Section 8

Use case 1 with every template filled in

Download the templates as a Word document (.docx, 1.5 MB): Part 2 has these templates filled in for use case 1, and Part 1 has them blank. Appendix G describes the document.

Client	A global audit, tax and advisory firm (illustrative)
Use case	Engagement launch: from verbal win to ready-to-charge
Assessment	Five weeks, eight workshops (Appendix B) and a readout to the managing partner for operations and the leadership team (illustrative)
Volumes and rates	10,000 engagement launches a year, $120 an hour loaded cost, $200 an hour average rate and 5,000 client-facing staff working 1,800 hours a year (illustrative)

T1 filled: use case at a glance

Use case	Engagement launch: from verbal win to ready-to-charge
Problem today	Between a verbal win and the first chargeable hour sits a chain of manual, risk-sensitive steps: resolving the legal entity, checking conflicts and auditor independence, drafting the letter, pricing, staffing, and creating the SAP project and billing plan. Each step lives in a different system and is owned by a different function, so the chain takes days to weeks. People wait on the bench, early work goes unbilled, and a missed affiliate can become an independence breach.⁸⁹
Target state	An Engagement Launch Orchestrator coordinates six specialist agents across the CRM, the independence system, document management, Workday, SAP S/4HANA and Snowflake. People keep every judgment that carries professional or regulatory accountability. Agents assemble, check, draft and update, and the control plane makes every step governed, observable and auditable.
Trigger	An opportunity reaches verbal win in the CRM
End state	Independence cleared and documented, engagement letter signed, team soft-booked, and the SAP project, WBS, billing plan and charge codes active
Users	Engagement managers and partners, Risk & Quality (independence reviewers), resource management and finance
Systems of record	CRM Independence and conflicts system Document management and e-signature Workday HCM SAP S/4HANA Snowflake
Out of scope	Delivery of the engagement itself and invoicing after launch

Stages

Stage	Human step	Work mode	Agent and tier	System of record
Verbal win	Manager sees the launch plan	AI-led: planning and tracking	Launch Orchestrator (Tier 2)	CRM
Accept client	None	AI-led, read-only: matching to the golden record	Entity Resolution (Tier 2, read-only)	Client master (MDM)
Risk clearance	Risk decides	Human-led: independence is a regulated professional judgment	Conflicts & Independence (Tier 1)	Independence system
Scope and price	Manager edits the letter	AI-assisted: agents draft, and the manager edits and accepts	Scoping & Letter and Pricing & Budget (Tiers 1–2)	DMS SAP S/4HANA Snowflake
Staff team	Manager confirms	AI-assisted: the agent proposes and the manager confirms	Staffing (Tier 2)	Workday HCM
Approve and set up	Partner approves Client signs	Human-led approval, then AI-led setup under the bound approval	Engagement Setup (Tier 3, gated)	SAP S/4HANA

Goals

Goal	Type	Measured by
Faster time to revenue: verbal win to ready-to-charge in about two business days	Business	Win-to-ready-to-charge
Lower cost to launch: Risk, resourcing and finance work exceptions instead of re-keying data	Business	Manual effort per launch
No known independence exceptions at engagement start, with a complete evidence trail for inspections	Business	Independence cleared before start Post-start exceptions
A better client experience (engagement letter the same day) and staff experience (clear assignments, fewer chasers)	Business	Staffing within 48 hours
Every model and tool call governed by one control plane, with AI cost per launch within its target	Control plane	Governed AI traffic AI cost per launch Agent quality

Selection scores

Criterion	Score (1–5)	Evidence
Value at stake	5	10,000 launches a year, with about $30M a year at stake (T2)
Visible pain	5	Partners and managers chase status, and early hours go unbilled
Reachable systems	4	SAP, Workday, Snowflake and the CRM have APIs, but the independence system needs a search API
Data readiness	3	Client hierarchy spread across three systems, with stewardship needed first
Safe first step	5	Entity resolution and independence can start read-only, with Risk deciding
Measurable in 90 days	4	Cycle time and effort can be measured in one service line
Sponsor and owner	4	Managing partner for operations sponsors Head of practice operations owns
Total (out of 35)	30	Strong first use case

*Figure 9. Use case 1 at a glance: the launch stages, the agent and human step in each, and the governed path to each system of record (handbook Figure 8).*

Why it is filled this way

The use case is named after its business outcome, from verbal win to ready-to-charge, so the sponsor recognizes the value in the title.

The out-of-scope row excludes delivering the engagement and invoicing afterwards, which keeps the first phase to the launch alone.

Every stage shows who keeps the decision. The use case is safe to start because independence, engagement terms and partner approval stay with people while agents assemble, check, draft and update.

Each stage has a work mode and, where it is human-led, the reason. Risk clearance stays human-led because independence is a regulated professional judgment: the should-AI answer is no, whatever a model could do.

Data readiness scores 3 out of 5 because the client hierarchy is spread across three systems. That gap puts the golden record into phase 1 without stopping the use case.

T2 filled: business case and value hypothesis

Value lever	How the solution delivers it	Formula and inputs	Annual value (illustrative)	Confidence
Capacity released	Manual assembly, checking and data entry falls from about 14 to 5 hours per launch across managers, Risk, resourcing and finance	10,000 launches × 9 hours × $120 loaded cost	About $10.8M a year	Medium
Revenue protected	Charge codes exist when work starts, so early hours are billed instead of written off	2,500 affected launches × 20 hours × $200 average rate	About $10.0M a year	Medium
Utilization	Teams are confirmed within days instead of weeks, reducing bench time	0.5 point × 5,000 client-facing staff × 1,800 hours × $200	About $9.0M a year	Low
Risk avoided	Every independence decision is made by a person before work starts, with evidence and lineage attached	Avoided remediation, regulatory and reputational cost	Not quantified	Not applicable

Assumption	Value	Source	Agreed by
Engagement launches a year	10,000	CRM (illustrative)	Business owner
Loaded cost per hour	$120	Finance (illustrative)	Finance partner
Average rate per hour	$200	Finance (illustrative)	Finance partner
Launches starting before charge codes exist	About 2,500 a year, about 20 hours each	SAP and time records (illustrative)	Finance partner
Client-facing staff and hours	5,000 staff 1,800 hours a year	Workday (illustrative)	Talent

Run rate at full scale	About $30 million a year released or protected, before platform, model and change costs
Costs	To be estimated with the project team: platform, implementation, model usage and change
Expected business outcomes	Verbal win to ready-to-charge in about two business days Lower cost to launch No known independence exceptions at engagement start, with a complete evidence trail The engagement letter the same day
Payback benchmark	Typical AI payback is two to four years.⁸² This plan releases value from month 3 and reaches the full run rate at month 12, so the payback month depends on the costs the project team will estimate.
Horizons	The read-only first slice is judged in months. The gated SAP setup, which acts in a system of record, is judged on the 12-month horizon.

Why it is filled this way

Each lever has a formula that finance can check line by line. The inputs are labeled illustrative until the firm’s own baseline replaces them.

Utilization carries low confidence because it depends on how quickly freed time is re-staffed, which the firm will measure in phase 2.

Risk avoided is described without a price. The firm may have its own estimate of remediation and regulatory cost, and until that estimate is available, evidence and lineage are the value.

The benchmark row sets the case beside what finance will compare it with: the two to four years a typical AI use case takes to pay back.

T3 filled: asset inventory

The inventory below is illustrative. For each tool, the table shows whether it exists today or must be built, and the phase column shows when the plan needs it.

*Figure 10. Use case 1 assets by layer, from the asset inventory (client deck, Part 1).*

T3a. Applications and systems of record

System	Owner	Role in the use case	Interfaces	Agent access	Phase
CRM	Sales operations	Opportunity and account The verbal-win event starts the launch	Event Opportunity and account API	Read	1
Independence and conflicts system	Risk & Quality	Conflict and independence searches, restricted-entity list, clearance flag	Search API Clearance-flag API	Read (a person decides)	1
Master data hub	Data governance	Client, worker and engagement golden records with hierarchies	MDM API MCP tools (resolve_entity, get_hierarchy)	Read	1
Document management and e-signature	Risk & Quality Practice operations	Templates, clause library, precedent Engagement letters Client signature	Search and document APIs Signature event	Read Create drafts	2
Workday HCM	Talent	Skills, certifications, availability	Workday API	Read	2
SAP S/4HANA	Finance	Rate cards Project, WBS, billing plan and charge codes	SAP APIs	Read (phase 2) Write, policy-gated (phase 3)	2–3
Snowflake	Finance and data platform	Historical engagement economics through governed semantic views	Semantic views	Read	2
Collaboration tool (chat copilot)	IT	The engagement manager’s entry point, in Slack or Teams	Chat app	Channel only	1

T3b. Data assets and master data

Data asset	System of record	Golden record today?	Quality and risk notes	Sensitivity	Owner or steward
Client and legal-entity master	Master data hub, fed by the CRM, SAP and the independence system	To extend	A missed subsidiary is a missed conflict: hierarchy freshness, duplicate entities, missing LEI or tax IDs	Confidential	Risk & Quality data steward
Restricted-entity list	Independence system	Yes	Must link to the client master	Restricted	Risk & Quality
Worker master	Workday (system of record), cross-referenced in MDM	Partly	Worker IDs differ across the CRM, SAP and the independence system	Confidential (personal data)	Talent
Engagement master	Master data hub	New	One engagement ID for the opportunity, the independence case, the letter and the SAP project	Confidential	Finance
Reference data: service catalog, rate cards	SAP S/4HANA Service catalog	Partly	Effective dates must be respected	Internal	Finance
Independence declarations	Independence system	Yes	Personal financial interests: field-level encryption or tokenization Others see a yes or no clearance only	Restricted	Risk & Quality
Templates, clauses and precedent	Document management and knowledge base	Not applicable	Approved templates only Information barriers between engagements	Confidential	Risk & Quality
Historical engagement economics	Snowflake semantic views	Not applicable	Shared metric definitions (realization, utilization)	Confidential	Finance

T3c. Integrations, APIs, events and MCP tools

Tool or interface	System	Operation	Scope	Exists or to build	Used by
Verbal-win event	CRM	Start the launch	Event	Exists	Launch Orchestrator
resolve_entity get_hierarchy	Master data hub	Match the client and return hierarchy and identifiers	Read	New MCP tools	Entity Resolution
search_restricted check_services get_relationships	Independence system	Conflict and independence searches	Read	To build	Conflicts & Independence
search_knowledge get_template	Document management and knowledge base	Precedent and approved templates	Read	To build	Scoping & Letter
create_document	Document management	Create the draft letter	Write (drafts only)	To build	Scoping & Letter
get_rate_card	SAP S/4HANA	Rate cards	Read	To build	Pricing & Budget
query_actuals	Snowflake	Historical actuals through a semantic view	Read	To build	Pricing & Budget
find_available_staff	Workday HCM	Skills, availability and location	Read	To build	Staffing
get_clearance_flag	Independence system	Yes or no clearance for a person and client	Read (yes or no only)	To build	Staffing
create_project create_wbs create_billing_plan	SAP S/4HANA	Project, WBS, billing plan and charge codes	Write, policy-gated	To build (phase 3)	Engagement Setup
Signature event	E-signature	Client signed	Event	Exists	Launch Orchestrator

T3d. Agents

Agent	Responsibility	Tier	Human role	Owner	Phase
Launch Orchestrator	Plans and tracks the launch, delegates tasks over A2A and enforces the regulated sequence	Tier 2	Sees plan and status	Practice operations (proposed)	1
Entity Resolution	Matches the client to its MDM golden record and returns hierarchy and identifiers	Tier 2, read-only	None	Data governance (proposed)	1
Conflicts & Independence	Runs conflict and independence searches and assembles findings and precedent	Tier 1	Risk decides	Risk & Quality	1
Scoping & Letter	Drafts scope and engagement letter from approved templates and clauses	Tier 1	Manager edits Partner approves	Risk & Quality (templates)	2
Pricing & Budget	Builds the budget from rate cards and historical actuals and checks margin	Tier 2	Manager accepts	Finance (proposed)	2
Staffing	Proposes a team by skills, availability, location and independence clearance	Tier 2	Manager confirms	Talent	2
Engagement Setup	Creates project, WBS, billing plan and charge codes	Tier 3, policy-gated	None (gated by policy)	Finance	3

Agent	Sponsor	Lifecycle stage	Next recertification	Retirement criteria
Launch Orchestrator	Head of practice operations	Register and onboard (phase 1)	Every six months	The launch process changes or a successor replaces it
Entity Resolution	Head of data governance	Register and onboard (phase 1)	Every six months	Precision below threshold for a quarter, or matching moves into MDM
Conflicts & Independence	Risk & Quality leader	Register and onboard (phase 1)	Quarterly (regulated work)	Independence rules change beyond its evaluation suite
Scoping & Letter	Risk & Quality leader (templates)	Propose (phase 2)	Quarterly	The clause library is replaced, or accuracy falls below threshold
Pricing & Budget	Finance partner	Propose (phase 2)	Every six months	The rate-card system is replaced
Staffing	Talent leader	Propose (phase 2)	Every six months	Workday matching covers the need
Engagement Setup	Finance partner	Propose (phase 3)	Quarterly (writes to SAP)	SAP setup is automated another way, or a failed binding check goes unexplained

T3e. Models

Model	Provider and hosting	Tasks	Data rules	Fallback
Frontier model (primary)	Provider under enterprise terms with zero data retention	Drafts engagement letters and summarizes complex findings	Personal data masked Approved regions No training on firm data	Second frontier provider
Frontier model (failover)	Second provider, same terms	Failover for drafting and summaries	As above	Not applicable
Open-weight small language model	Self-hosted in the firm’s VPC	Entity extraction, classification and personal-data detection	Client-confidential text never leaves the firm for routine tasks	Not applicable
Embedding model	Private endpoint	Indexes credentials, prior proposals and templates for retrieval	Indexes partitioned by client and engagement (information barriers)	Not applicable

Routing policy. Client-confidential payloads go only to private models or approved regions. Drafting goes to the frontier model with personal data masked, and token budgets are set per launch.

T3f. Identities, users and channels

Actor	Channel	Authentication	Authority and approvals
Engagement managers	Chat copilot	Single sign-on with phishing-resistant MFA	Edit letters, accept budgets, confirm teams
Partners	Mobile approval	Step-up authentication	Approve letter and budget Approve the SAP setup request (bound to that request, expires at the end of the business day)
Risk & Quality reviewers	Case workspace	Single sign-on with MFA	Decide independence (Tier 1)
Resource management and finance	Existing tools	Single sign-on with MFA	Work exceptions
Client signatories	E-signature portal	Portal authentication	Sign the engagement letter
Agents (seven)	None (machine actors)	A unique workload identity per agent instance⁹¹	Least-privilege scopes On-behalf-of tokens with narrowed scopes⁹² No agent approves its own output

T3g. Policies and controls

Policy or control	Type	Enforced today	Target enforcement	Owner
Engagement setup in SAP requires independence status CLEARED, the partner’s approval bound to the exact setup request (expiring at the end of the business day) and a signed letter	Approval and gate	Manual checks	Policy decision point A custom Omni Gateway policy, built with the Policy Development Kit, for approval binding	Finance Risk & Quality
Engagement letters use approved templates, and any clause deviation routes to Risk & Quality	Content	Manual review	Template and clause checks in the Scoping & Letter agent	Risk & Quality
Staffing proposals include only people with independence clearance for the client and the required credentials	Authorization	Manual	Clearance flag (yes or no) checked by the Staffing agent	Talent Risk & Quality
Client-confidential data is processed only by private models in approved regions	Data and routing	Not enforced	AI Gateway routing policy	CISO
Personal data is masked before prompts	Data	Not enforced	AI Gateway masking and PII detection	Privacy officer
Token budget per launch AI cost per launch within its target	Cost limit	Not measured	AI Gateway token-based rate limits Spend attribution in LLM Governance	Finance Platform owner
Information barriers between engagements	Data access	Partly (document permissions)	Retrieval filtering Partitioned indexes Snowflake row-access policies	Risk & Quality
Autonomy tiers and new agents approved by the AI governance council Evaluation before promotion	Governance	None	Agent Registry tiers Evaluation suites (any evaluation tool)	AI governance council
Auditor-independence rules⁸⁹	Regulatory	Manual evidence	Evidence and lineage attached to every decision	Risk & Quality

T3h. Infrastructure, runtime and observability

Component	Notes
Firm virtual private cloud	Gateways, MCP servers and the self-hosted model run inside the firm’s VPC where confidentiality requires
Private connectivity	Private links to SAP, Workday and Snowflake, with no public endpoints for systems of record
Keys and secrets	Customer-managed keys in a KMS or HSM AES-256 at rest
Transport	TLS 1.3 everywhere mTLS between services Hybrid post-quantum key exchange at gateways as vendors support it
Observability	OpenTelemetry traces for every hop, tied to one engagement correlation ID, feeding the firm’s SIEM and an operations dashboard⁹⁴

Why it is filled this way

The inventory counts six systems of record, seven agents, four models, fifteen MCP tools, four master data domains and nine policies. Those counts become the discovery slide, and every row has an owner.

The master data hub appears as a system even though the golden record is “to extend”: a missed subsidiary is a missed conflict, so the client master is on the critical path from phase 1.

Most policies are “manual” or “not enforced” today. That column makes the case for the control plane, which enforces the same rules at runtime in the gateway path.

Owners marked “proposed” are the lead architect’s suggestion for the firm to confirm. The handbook names Risk & Quality, Finance and Talent as owners of their agents.

Every agent also has a named sponsor and a recertification date. The independence and setup agents are recertified quarterly, because one supports regulated work and the other writes to SAP.

T4 filled: requirements register

ID	Type	Requirement	Priority	KPI	Phase	Owner
BR-01	Business	Cut verbal-win-to-ready-to-charge from 10 to 2 business days	Must	Win-to-ready-to-charge	3	Business owner
BR-02	Business	Cut manual effort per launch from 14 to 5 hours	Must	Manual effort per launch	1–3	Business owner
BR-03	Business	Every independence decision is made by a person before work starts, with evidence and lineage	Must	Independence cleared before start	1	Risk & Quality
FR-01	Functional	On verbal win, create a launch plan with a correlation ID and show it to the engagement manager within a minute	Must	Win-to-ready-to-charge	1	Practice operations
FR-02	Functional	Resolve the client to its golden record with the full hierarchy (ultimate parent, affiliates, subsidiaries) and identifiers	Must	Post-start exceptions	1	Data governance
FR-03	Functional	Run conflict and independence searches and assemble findings and precedent for a reviewer to decide (Tier 1)	Must	Independence cleared before start	1	Risk & Quality
FR-04	Functional	Draft the engagement letter from approved templates and clauses, and route any deviation to Risk & Quality	Must	Manual effort per launch	2	Risk & Quality
FR-05	Functional	Build the budget from rate cards and historical actuals and check the margin against target	Should	Manual effort per launch	2	Finance
FR-06	Functional	Propose a team by skills, availability, location and independence clearance	Must	Staffing within 48 hours	2	Talent
FR-07	Functional	Create the SAP project, WBS, billing plan and charge codes only when independence is cleared, the partner has approved that exact request and the letter is signed	Must	First-time-right setup	3	Finance
DR-01	Data	Maintain one golden record per legal entity with hierarchy, LEI and tax IDs, linked to the restricted-entity list, and monitor freshness and duplicates	Must	Post-start exceptions	1	Data governance
DR-02	Data	Share one engagement ID across the opportunity, the independence case, the letter and the SAP project	Must	First-time-right setup	1–3	Finance
DR-03	Data	Publish a glossary, shared metric definitions and a data contract for each MCP tool, and capture lineage from source to agent output	Should	Agent quality	2	Data governance
IR-01	Integration	Expose each system of record as governed APIs and MCP tools with explicit read or write scopes	Must	Governed AI traffic	1–3	Enterprise architect
NFR-01	Non-functional	Fail over the primary model to a second provider, and make every SAP write idempotent	Must	First-time-right setup	2–3	Platform owner
NFR-02	Non-functional	Set a token budget per launch, keep AI cost per launch within its target, and attribute every model and tool call to the engagement and its cost center within a day	Must	AI cost per launch	1–2	Finance
SR-01	Security	Give every agent instance its own workload identity with short-lived credentials and no static API keys Use on-behalf-of tokens with narrowed scopes No agent approves its own output	Must	Oversight coverage	1	CISO
SR-02	Security	Process client-confidential data only on private models in approved regions, and mask personal data before prompts	Must	Governed AI traffic	1	CISO
SR-03	Security	Bind the partner’s approval to the exact setup request and expire it at the end of the business day	Must	Oversight effectiveness	3	CISO Finance
CR-01	Compliance	Keep the evidence behind every independence decision for inspections, with retention aligned to professional-standards documentation requirements	Must	Independence cleared before start	1	Risk & Quality
OR-01	Operational	Trace every hop with OpenTelemetry under one engagement correlation ID, and show cycle time, policy blocks, token cost and evaluation drift on one dashboard	Must	Governed AI traffic	1	Platform owner
OR-02	Operational	Pass evaluation suites before promotion (letter-clause accuracy, independence-finding recall, entity-resolution precision), then score sampled production traces continuously	Must	Agent quality	2	AI governance council
OR-03	Operational	Before the setup agent reaches tier 3, write and drill a takeover runbook: open launches move to the finance setup queue with their trace, pending approvals and data touched	Must	Time to contain	3	Finance

Why it is filled this way

Every requirement names the KPI it moves and the phase in which it must be met, so phase 1 stays small: the gated SAP write (FR-07) and the bound approval (SR-03) are phase 3.

The data requirements (DR-01, DR-02) are Must and phase 1 because entity resolution and independence depend on them from the first slice.

The operational requirements treat evaluation as a lifecycle (OR-02) and make the takeover runbook a condition of tier 3 (OR-03). NFR-02 makes every model and tool call attributable to the engagement that caused it.

The register stops at 23 rows. It covers what shapes the architecture and the plan and leaves detailed user stories to delivery.

T5 filled: security and compliance requirements

Security follows zero-trust principles:⁹⁰ every actor is authenticated, every call is authorized for the specific task, and nothing is trusted because of where it sits on the network.

Layer	Requirement (controls)	Enforced by	Phase
Users	Single sign-on with phishing-resistant MFA Step-up authentication for partner approval Role- and attribute-based access	Firm identity provider Trusted Agent Identity mobile approval	1
Agents	A unique workload identity per agent instance⁹¹ Short-lived credentials and no static API keys⁴⁶ Least-privilege scopes Autonomy tiers Kill switch No agent approves its own output	Agent Registry (owner and tier) Trusted Agent Identity Agent Kill Switch	1
A2A	Registered, signed agent cards, verified by the registry (A2A version 1.0) Mutual TLS plus OAuth tokens on every call²⁷ Delegation carried through on-behalf-of token exchange with narrowed scopes⁹² Only registered agents may call each other	Agent Registry Agent Broker A2A support Omni Gateway policies	1–2
MCP	MCP servers behind the gateway as OAuth-protected resources, audience-bound tokens with no pass-through, per-tool scopes, input and output schema validation, and policy on MCP’s standard request headers, with every call joined to its trace²⁶ As firm practice: allow-lists per agent, only registered and version-pinned servers, and separate read and write tools	AI Gateway MCP policies MCP Bridge	1
APIs	OAuth 2.0 client credentials or signed JWTs mTLS to backends Rate limits, payload and injection threat protection Private connectivity to SAP, Workday and Snowflake	Omni Gateway policies⁷⁸	1
Models	Access only through the gateway, enterprise terms with zero data retention and no training on firm data, prompt-injection and jailbreak screening, personal data masked before prompts, and output validation⁹³	AI Gateway routing, masking and token-based rate limits	1
Data at rest	AES-256 with customer-managed keys in a KMS or HSM Field-level encryption or tokenization for personal financial interests Snowflake row-access and masking policies Vector indexes partitioned by client and engagement	Firm KMS or HSM Snowflake policies Informatica classification	1
Data in motion	TLS 1.3 everywhere mTLS between services Private links to SaaS Hybrid post-quantum key exchange at gateways as vendors support it	Gateways Private links	1
Audit	Tamper-evident logs linking user, agent, tool call and record Retention aligned to professional-standards documentation requirements	Agent Visualizer OpenTelemetry traces to the firm SIEM	1

Autonomy by agent

Agent	Tier	Human role	Approval bound to	Limits and kill switch
Launch Orchestrator	Tier 2	Sees plan and status	Not applicable	Token budget per launch Kill switch
Entity Resolution	Tier 2, read-only	None	Not applicable	Read-only scopes
Conflicts & Independence	Tier 1	Risk decides	The reviewer’s decision, signed and logged	Read-only No agent decides
Scoping & Letter	Tier 1	Manager edits Partner approves	The partner’s approval of the letter	Approved templates only Drafts only
Pricing & Budget	Tier 2	Manager accepts	Not applicable	Read-only
Staffing	Tier 2	Manager confirms	Not applicable	Clearance returned as yes or no only
Engagement Setup	Tier 3, policy-gated	None (gated by policy)	The partner’s approval of that exact setup request, expiring at the end of the business day	Idempotent writes Kill switch drilled before go-live

Promotion and demotion

Agent	Current tier	Evidence to promote	Demotion triggers
Engagement Setup	Tier 3, gated, from month 12	At month 8 it proposes setups that match the partners’ approvals for a full quarter Red-team findings closed Impact assessment updated Finance and Risk & Quality agree	An approval fails its binding check Independence status changes after approval Stale hierarchy data Any security event
Scoping & Letter	Tier 1, drafts only	Stays at tier 1: the partner approves every letter. Standard letters could move to tier 2 only after clause accuracy holds for two quarters and Risk & Quality agrees	Clause accuracy below threshold A template change not yet evaluated
Conflicts & Independence	Tier 1 by design	Not promoted: independence stays a human-led decision	Finding recall below threshold The restricted-entity list out of date
Staffing	Tier 2, proposes	Stays at tier 2: managers confirm every team	The clearance-flag check fails Stale Workday data

Human takeover

Agent	Who takes over	Information handed over	Degraded mode and service level	Drills
Engagement Setup	Finance project-setup team	The launch trace, the partner’s bound approval, the setup request and any SAP objects already created	Manual SAP setup as today, within one business day	Before tier 3 goes live, then quarterly
Launch Orchestrator	Practice operations	The launch plan and status of every launch in flight	Managers follow the launch checklist by hand	Before phase 2, then twice a year
Staffing	Resource management	Proposed teams and their clearance flags (yes or no only)	Manual staffing as today	Twice a year

Threat coverage

OWASP agentic risk	What it looks like here	Controls in this design
ASI01 Agent Goal Hijack	Hidden instructions in client documents or precedent redirect a letter or a finding	Input and tool-result screening at AI Gateway Template and clause checks The partner approves the exact letter
ASI02 Tool Misuse	The setup agent creates the wrong SAP objects	Separate read and write tools Schema validation Idempotent writes Per-launch limits
ASI03 Identity & Privilege Abuse	An agent acts beyond the authority of the person it acts for	One identity per agent On-behalf-of tokens with narrowed scopes The approval bound to the exact setup request
ASI04 Agentic Supply Chain Vulnerabilities	A poisoned or changed MCP server enters the launch path	Only registered, version-pinned MCP servers Reviewed tool definitions Gap: the platform team owns what enters the registry
ASI05 Unexpected Code Execution	A drafting request triggers code or commands	No code or shell tools for any launch agent
ASI06 Memory & Context Poisoning	Wrong entity data leads to a missed conflict	Context only from the golden record, the catalog and lineage Quality rules on the client hierarchy
ASI07 Insecure Inter-Agent Communication	A spoofed agent asks the orchestrator for independence findings	Signed agent cards mTLS plus OAuth on every A2A call Only registered agents may call each other
ASI08 Cascading Failures	A bad hierarchy update spreads to every launch that day	Propose-only on stale master data Per-launch limits The kill switch
ASI09 Human-Agent Trust Exploitation	A polished summary leads a reviewer or partner to approve without reading	Findings shown with evidence and lineage Approver training Verification latency watched Gap: Risk & Quality owns what reviewers are shown
ASI10 Rogue Agents	An unregistered agent calls SAP or the independence system	Agent Scanners and the registry Decoy tools Gateway-only access to systems of record

Compliance and evidence

Obligation	Evidence required	Where it is produced	Owner
Auditor-independence rules⁸⁹	Independence cleared before work starts, with the findings and the reviewer’s decision	Independence system Decision log with lineage	Risk & Quality
Professional-standards documentation	Retained evidence for inspections	Tamper-evident audit logs Document management	Risk & Quality
Data protection	Personal data masked Personal financial interests protected	AI Gateway masking Tokenization in the independence system	Privacy officer
AI governance standards the firm reports against (ISO/IEC 42001; NIST AI RMF)^72,73	Inventory, risk tiers, impact assessments, evaluation and monitoring records	Agent Registry, evaluation records and audit logs (handbook Appendix C maps the controls)	AI governance council

Why it is filled this way

Only one agent writes to a system of record, and only under three conditions: independence cleared, the partner’s bound approval and a signed letter. Every other agent reads, drafts or proposes.

Under data minimization, the staffing agent sees a yes or no clearance flag and never the personal financial interests behind it, which keeps sensitive data where it belongs.

Most controls are phase 1 because security starts on day one, while autonomy is phased in later.

The setup agent earns tier 3 on evidence: a quarter of proposed setups that match the partners’ approvals. It drops back automatically if a binding check fails or independence status changes, and finance can take over its open launches within a business day.

The threat table shows coverage and the two gaps that need named owners: what enters the registry (the platform team) and what reviewers are shown (Risk & Quality).

T6 filled: KPIs and scorecard

KPI	Definition	Data source	Owner
Win-to-ready-to-charge	Business days from verbal win to active charge codes	CRM and SAP timestamps	Business owner
Manual effort per launch	Hours across all roles to launch one engagement	Time study Activity logs	Business owner
Independence cleared before start	Share of engagements cleared before any work begins	Independence system SAP	Risk & Quality
Post-start independence exceptions	Exceptions found after work begins, per 1,000 engagements	Independence system	Risk & Quality
First-time-right setup	SAP projects created without later correction	SAP change logs	Finance
Staffing within 48 hours	Roles confirmed within 48 hours of verbal win	Workday CRM	Talent
Agent quality	Evaluation pass rate for letters and independence summaries	Evaluation suites	AI governance council
Governed AI traffic	Share of model and tool calls through the AI gateway	AI Gateway	Platform owner
AI cost per launch	Model and platform cost per engagement launched	LLM Governance Finance	Finance
AI cost visibility	Share of AI spend for launches attributed to an engagement and its cost center within a day	AI Gateway LLM Governance	Finance

KPI	Baseline	Month 3	Month 8	Month 12	Month 24
Win-to-ready-to-charge	10 days	Measure in scope	Measure	2 days	2 days
Manual effort per launch	14 hours	About 10 hours in scope	About 8 hours	5 hours	5 hours
Independence cleared before start	92%	Measure in scope	Measure	100%	100%
Post-start independence exceptions	4	Measure in scope	Measure	Below 0.5	Below 0.5
First-time-right setup	80%	Manual, as today	Manual, as today	97%	97%
Staffing within 48 hours	55%	Measure	Toward 70%	85%	85%
Agent quality	Not measured	Baseline set	Track	95% or higher	95% or higher
Governed AI traffic	Not measured	100% in scope	100% in scope	100%	100%
AI cost per launch	Not measured	Baseline set	Within target	Within target	Within target
AI cost visibility	Not measured	100% in scope	100% in scope	100%	100%

Control-plane scorecard for the use case

Metric	What it shows for this use case	Leading indicator	Data source
Agent coverage	Every launch agent registered with an owner, a sponsor and a tier	Launch agents found by scanners but not yet registered	Agent Registry and Scanners
Governed traffic	Every model and tool call in scope through AI Gateway	New launch agents whose first production call went through the gateway	AI Gateway
Cost per outcome	AI cost per launch against its target	Cost per successful task: letters and budgets that passed their checks	LLM Governance
AI cost visibility	AI spend for launches attributed to an engagement within a day	Share of the previous day’s spend tagged to an engagement	AI Gateway LLM Governance
Value realized	Run rate of the live slice against the business case: about $0.5M, $8M, then $30M a year	Acceptance rate of letters and budgets without significant edits	Business-case tracking
Oversight effectiveness	SAP setups executed exactly as the partner approved, within the approval’s validity	Takeover rate: how often a manager or reviewer takes over a task an agent started	Gateway approval-binding events
Time to contain	Kill-switch and takeover drill passed before tier 3 goes live	Drill time: minutes to stop the setup agent and hand its open launches to finance	Agent Kill Switch Takeover drill records

Why it is filled this way

Baselines are illustrative. Three are “not measured” because the firm has no gateway today, and measuring them is itself a phase 1 outcome.

Month 3 targets apply to one service line in one region, and say so. All targets are met by month 12, and month 24 shows them sustained.

The scorecard ties the use case to the platform: agent coverage, governed traffic and cost per outcome are the evidence the control plane is working.

Every scorecard metric has a leading indicator that moves within weeks, and AI cost visibility is measured from phase 1, when every call in scope is tagged to its engagement.

T7 filled: north-star architecture (vendor-neutral)

Layer	This use case
Users and clients	Chat copilot (engagement managers) Mobile approvals with step-up MFA (partners) Risk workspace (independence reviewers) E-signature portal (client signatories)
Agents and tiers	Entity Resolution (Tier 2, read-only) Conflicts & Independence (Tier 1, Risk decides) Scoping & Letter (Tier 1, drafts only) Pricing & Budget (Tier 2, proposes) Staffing (Tier 2, proposes) Engagement Setup (Tier 3, gated writes)
Orchestration	Hybrid/federated Launch Orchestrator as supervisor Deterministic gates for independence, approvals and signature A2A delegation²⁷ CRM and e-signature event triggers
Control plane	Registry of agents, tools and APIs AI gateway for LLM, MCP and A2A traffic Policy decision point (policy as code) Identity and delegation Evaluation and cost Kill switch
Models	Frontier model for drafting (zero retention) Second frontier provider for failover Self-hosted open-weight SLM for extraction Embeddings on a private endpoint
MCP servers and APIs	CRM, independence system, knowledge and document management, Workday, SAP S/4HANA and Snowflake, each with explicit read or write scopes²⁶
Data governance	Golden records for client, worker and engagement Semantic layer and data contracts Catalog and lineage Data quality Access policies and information barriers
Systems of record	CRM Independence system Document management and e-signature Workday HCM SAP S/4HANA Snowflake
Across every layer	Security: MFA, per-agent identity, OAuth scopes per tool, mTLS, masking, encryption, tamper-evident audit. Governance: autonomy tiers, human approval gates, agent owners, OpenTelemetry traces, token cost per launch, evaluation scores

Design decisions

Decision	Choice and rationale
Orchestration pattern	Hybrid/federated: the control plane governs identity, policy, cost and audit centrally, while each function owns its agents (Risk & Quality, Finance, Talent). Within a launch, a supervisor pattern applies.
Model hosting and routing	A frontier model under enterprise terms with zero data retention for drafting, a second provider for failover, and a self-hosted open-weight model for extraction so client-confidential text stays in the firm.
Where gateways and MCP servers run	Inside the firm’s VPC where confidentiality requires (customer-managed Runtime Fabric in the realization).
Data residency	Client-confidential payloads only to private models or approved regions Private links to SaaS Customer-managed keys
Human approvals and tiers	Independence, engagement terms and partner approval stay human. The setup write is gated, and the partner’s approval is bound to the exact request and expires at the end of the business day.
Observability and audit	OpenTelemetry traces for every hop under one engagement correlation ID, feeding the firm’s SIEM and an operations dashboard.⁹⁴

Design review against the firm’s framework

Check	Result	Evidence
Every agent has an owner, a sponsor, a risk tier and an identity of its own	Yes	Seven launch agents, each with an owner, a sponsor, a tier and its own workload identity (T3d, T5)
Each agent’s tool access is least-privilege and allow-listed	Yes	Allow-lists per agent and separate read and write tools (T5)
Human oversight matches the risk and reversibility of each action	Yes	Tier 1 for independence and letters, and the SAP write gated by a bound approval (T5)
Agents’ data access is aligned with their users’ access	Yes	On-behalf-of tokens carry the manager’s or partner’s narrowed scopes (T5)
Only trusted, registered MCP and A2A servers are reachable	Partly	Registry and gateway from phase 1, with version pinning for MCP servers to be set up in phase 1
New policies start in audit mode, then enforce	Yes	Routing and masking policies start in audit mode in the phase 1 scope, then enforce
Agent reasoning, tool calls and hand-offs are traced end to end	Yes	One engagement correlation ID across every hop (OR-01)
Agent costs are tagged to cost centers, with alerts as spend nears budget	Yes	Token budget per launch and spend attributed to each engagement (NFR-02)
Adversarial testing runs before release and on a schedule	Partly	Evaluation suites before promotion (OR-02) The red-team schedule is to be agreed with the CISO
There is a tested way to disable a misbehaving agent and hand over its work	Yes	Kill switch and takeover runbook drilled before tier 3 (T5, OR-03)

*Figure 11. Use case 1 vendor-neutral solution architecture. Only the systems of record are named products (handbook Figure 9).*

Why it is filled this way

The regulated sequence is deterministic and is not left to model reasoning. There is no letter before independence is cleared, and nothing is written to SAP before the partner and the client have approved. Models reason inside steps and never skip a gate.

Data governance is a layer of its own because independence checks are only as good as the entity data behind them.

Only systems of record are named. Everything else stays vendor-neutral until T10.

The design review found two gaps to close in phase 1: version pinning for MCP servers and an agreed red-team schedule.

T8 filled: a day in the life

Persona	Maya Chen, engagement manager in the advisory practice, launching Project Atlas for a new client
The case	Project Atlas, a new client that turns out to be a subsidiary of a group whose parent is an audit client (illustrative)

Time	Moment	What happens	Who decides
08:05	Verbal win	The CRM opportunity for Project Atlas moves to verbal win. Within a minute Maya’s chat copilot shows a launch plan targeting ready-to-charge in two days.	No
08:12	Client resolved	The entity-resolution agent matches the client to its golden record and finds that it is a subsidiary of a group whose parent is an audit client. An enhanced independence review starts automatically.	No
09:40	Independence cleared	A reviewer in Risk & Quality approves the engagement with one condition: no work on financial-reporting systems.	Yes: Risk decides
10:30	Letter drafted	The scoping agent drafts the engagement letter with the carve-out, and Maya rewrites two paragraphs.	Yes: Maya edits
11:20	Budget ready	The pricing agent’s budget arrives with a 31% margin against a 30% target.	Yes: Maya accepts
13:05	Team proposed	The staffing agent proposes a team of five, having already replaced a consultant who is not independence-cleared for this client. Maya sees only that a substitution was required, not why.	Yes: Maya confirms
14:30	Partner approves	The partner approves the letter and budget on mobile with step-up authentication, and the letter goes to the client for signature.	Yes: partner approves
16:10	Ready to charge	Once the client signs, the setup agent creates the SAP project, WBS and billing plan, and the team receives charge codes. What used to take two weeks finished the same day.	Client signs Setup gated by policy

*Figure 12. A day in the life of an engagement manager launching Project Atlas (illustrative; handbook Figure 10).*

Why it is filled this way

The story includes one moment where a control changes the outcome, and it is the moment a CTO remembers: the entity-resolution agent finds the audit-client parent, which starts an enhanced review.

Maya sees only that a substitution was required, so the story shows data minimization without a lecture.

The day ends in business terms: what used to take two weeks finished the same day.

T9 filled: how the components work together

Time	What happens	Components and protocols	Controls applied
08:05	Verbal win starts the launch	CRM event triggers the Launch Orchestrator, which registers the plan with a correlation ID	Orchestrator workload identity Trace starts
08:12	Client resolved to its golden record	Orchestrator delegates to Entity Resolution over A2A Agent calls MCP tools on the master data hub (resolve_entity, get_hierarchy)	Read-only scopes Gateway authorization Lineage captured
09:40	Independence decision	Conflicts & Independence agent calls independence-system tools Private SLM summarizes Review task goes to Risk	Tier 1: a person decides Decision signed and logged
10:30	Letter drafted	Scoping agent retrieves precedent through the knowledge MCP, then drafts with the frontier model via the AI gateway	Retrieval filtered by information barriers Personal data masked Template and clause checks
11:20	Budget built	Pricing agent reads SAP rate cards and a Snowflake semantic view through MCP (read)	Row-access policies Metric definitions from the semantic layer
13:05	Team proposed	Staffing agent calls Workday MCP tools and the clearance-flag API	Data minimization: clearance returns yes or no only Manager confirms
14:30	Partner approval	Mobile approval with step-up MFA, recorded against the engagement ID	Non-repudiation Separation of duties
16:10	Engagement set up	Setup agent calls SAP MCP write tools after the policy decision point allows	Gate: cleared, approved and signed Idempotent writes Kill switch available

Delegation check

Hand-off	Unit of work and its check	Model	Permissions passed on	When it stops and asks
Orchestrator to Entity Resolution	Resolve one client to its golden record, checked against the hierarchy and identifiers returned	Self-hosted small model	Read-only MDM scopes, on behalf of the manager	An ambiguous match goes to a data steward, not to the next step
Orchestrator to Conflicts & Independence	Assemble the findings for one launch, checked by the Risk reviewer	Self-hosted small model for summaries	Read-only independence scopes with no decision rights	Always: a person decides
Orchestrator to Scoping & Letter	Draft one letter from approved templates, checked by template and clause rules	Frontier model through AI Gateway	Drafts only, in document management	Any clause deviation goes to Risk & Quality
Orchestrator to Engagement Setup	Create one SAP project, WBS and billing plan, checked against the approved request	None needed for the write	Write scopes for this launch only, after the gate allows	Any mismatch with the approval stops the write

*Figure 13. Component interaction for the Project Atlas launch. Numbered steps follow the day in the life (handbook Figure 11).*

Why it is filled this way

Each moment names the protocol: A2A between the orchestrator and specialist agents, MCP from agents to tools, APIs to systems of record.

The last row shows the policy decision point allowing the SAP write only when the three conditions are met. Every write in every use case needs a row like it.

Every hand-off narrows authority: the setup agent receives write scopes for one launch only, after the gate allows, and any mismatch with the approval stops the write.

T10 filled: realizing the use case with MuleSoft Agent Fabric and Informatica

Reference component	MuleSoft or Informatica capability	Role in this use case
Registry	Agent Registry and Agent Scanners (MuleSoft Agent Fabric)	Registers every launch agent, including agents other teams built on other platforms, with owner and autonomy tier²¹
Orchestration	Agent Broker with guided determinism (MuleSoft Agent Fabric)	Supervisor routing over A2A, with fixed gates for independence, approvals and signature²¹
AI gateway	AI Gateway for LLM and MCP traffic and Omni Gateway policies for APIs and A2A (MuleSoft)	One policy point for token limits, routing, masking, authentication and rate limits^21,78
Agent identity	Trusted Agent Identity (MuleSoft Agent Fabric)	Agents act on behalf of Maya or the partner, with mobile approval for high-risk steps²¹ A custom Omni Gateway policy, built with the Policy Development Kit, binds each approval to the executed call⁷⁸
Operate and control	Agent Kill Switch, Agent Visualizer, AI cost controls (MuleSoft Agent Fabric)	Stop a misbehaving agent, trace the launch, apply token-based rate limits and attribute AI spend to each launch²¹ Evaluation suites run with any evaluation tool
MCP servers and APIs	MCP Bridge, and System and Process APIs with MuleSoft connectors for SAP S/4HANA, Workday, Snowflake and the CRM (MuleSoft)	Existing APIs become governed MCP tools without rewriting integrations²¹
Master data	Multidomain MDM (Informatica)	Client, worker and engagement golden records with hierarchies, which agents reach through governed APIs exposed as MCP tools by MCP Bridge^21,77
Semantics, catalog, lineage	Data Governance and Catalog, with the CDGC Metadata Search MCP server (Informatica)	Shared glossary and lineage for agents and people, with agents searching catalog metadata through MCP⁷⁷
Data quality	Data quality rules, with the Address Verification MCP server (Informatica)	Validate entity data, such as addresses and identifiers, at the point of entry, and flag duplicates and stale hierarchies to data stewards⁷⁷
Deployment	CloudHub or customer-managed Runtime Fabric (MuleSoft)	Keep gateways and MCP servers inside the firm’s VPC where confidentiality requires²¹

The evaluation suites (OR-02) run with any evaluation tool, and their results are recorded in the registry so that promotion decisions rest on them.

What stays the firm’s choice

Component	Firm’s choice
Channels and collaboration tools	Chat copilot in Slack or Teams The Risk workspace The e-signature portal
Agent platforms and frameworks	Any platform, with agents registered and governed centrally
Models and providers	Frontier models from any provider through AI Gateway The self-hosted small model Embeddings
Systems of record	CRM, independence system, document management and e-signature, Workday, SAP S/4HANA, Snowflake
Identity provider and SIEM	The firm’s existing services, federated by the control plane

Coexisting with the firm’s other platforms

Collaboration platform	The chat copilot runs in Slack or Teams. The platform’s own admin controls stay in place, and the copilot’s calls into the launch agents pass through Agent Fabric.
Agents built elsewhere	Agents other teams built on other platforms are registered in Agent Registry. Their native planes stay authoritative for identity and runtime in their own estates.
Identity	The firm’s identity provider stays authoritative for people. Agent identities are federated, not duplicated.
Agent Fabric is authoritative for	The launch agents’ inventory, policy on every call into SAP, Workday, Snowflake, the CRM and the independence system, and AI cost and audit.

*Figure 14. Use case 1 realized with MuleSoft Agent Fabric and Informatica. Colors show which vendor fills each role (handbook Figure 12).*

Why it is filled this way

The mapping is one-to-one with the vendor-neutral design. MuleSoft supplies the control plane, the A2A and MCP paths and the governed APIs into SAP, Workday and Snowflake. Informatica supplies the golden records, semantics and quality that make independence checks trustworthy.

Approval binding is labeled custom: it is a custom Omni Gateway policy built with the Policy Development Kit, so the firm or a partner owns its build, testing and support.⁷⁸

Coexistence is settled in writing: other platforms keep their own controls and identities, and Agent Fabric is authoritative for the launch agents, every call into the firm’s systems, and cost and audit.

T11 filled: phased implementation with ROI and business value

Milestone	What goes live	KPIs at the milestone	Value (illustrative)	Formula	Foundations left behind
Month 3 (phase 1)	One service line in one region, about a tenth of launches. Entity resolution and independence agents work read-only at tiers 1–2, while people draft letters and set up SAP as today. Every model and tool call goes through the gateway from day one.	Manual effort per launch down from 14 to about 10 hours in scope Governed AI traffic 100% in scope	About $0.5M a year	1,000 launches × 4 hours × $120	Agents registered with owners and tiers Spend visible by use case The gateway live
Month 8 (phase 2)	Half of all launches. Scoping and pricing agents draft letters and budgets from approved clauses and rate cards, and the staffing agent proposes teams from Workday. The launch capabilities are published as governed MCP tools.	Manual effort about 8 hours in scope Staffing within 48 hours rising toward 70%	About $8M a year	Capacity about $3.6M (5,000 × 6 hours × $120) Utilization about $4.5M	Governed APIs and MCP tools reusable by the next use case Routing and token-based rate limits on
Month 12 (phase 3)	Every launch. The setup agent creates the SAP project, WBS and billing plan at tier 3 once the partner’s approval, bound to that exact request, and the client’s signature are in place.	All KPI targets met, including win-to-ready-to-charge in 2 days and 5 hours of effort per launch	About $30M a year	The full business case	Agent identities, bound approvals, drills and evaluation in every release
Month 24 (phase 4): all goals met	Run rate and KPI targets are sustained. Adjacent processes, such as engagement changes and closure, are added on the same controls, and models are swapped without application changes.	All KPIs held at target	About $33M cumulative by month 24	Each run rate counted from its milestone	Federated control plane Post-quantum plan on track

Platform phases and use case milestones phased together — *Figure 15. The platform and the use case, phased together (client deck, Part 1). Values are illustrative, before platform, model and change costs.*

Quick-win test for the month-3 slice

Test	Result	Evidence
Shows measurable value within 90 days in one scope (one service line, market, region or site).	Yes	One service line in one region, about 1,000 launches a year Effort down from 14 to about 10 hours in scope
Runs read-only or propose-only (tiers 1 and 2): no writes to systems of record yet.	Yes	Entity resolution and independence agents at tiers 1–2, while people still draft letters and set up SAP
Uses systems and data that are reachable now, or with one connector.	Yes, with one build	MDM and the independence search API (the search API is built in phase 1)
Saves time or money that the users and the sponsor can see.	Yes	About four hours saved per launch for managers and Risk reviewers
Leaves a reusable foundation behind: a registered agent, a governed MCP tool, a golden record or the gateway itself.	Yes	Registry, AI Gateway, the client golden record and the first MCP tools
Has a baseline taken before it starts, so the movement can be proved.	Yes	Cycle time and effort measured in the service line before go-live

Cumulative value and ROI

Period	Run rate in force	Value
Months 3–8	$0.5M a year for 5 months	About $0.2M
Months 8–12	$8M a year for 4 months	About $2.7M
Months 12–24	$30M a year for 12 months	About $30.0M
Total by month 24		About $33M

*Figure 16. Run-rate value at each milestone and how the cumulative value adds up (client deck, Part 1; illustrative).*

Net ROI and payback follow once the project team has estimated platform, model and change costs: ROI = (cumulative value − cumulative cost) ÷ cumulative cost, and payback is the month in which cumulative value first exceeds cumulative cost. Report both at every milestone review.

Payback benchmark	Typical AI payback is two to four years, against seven to 12 months for other technology investments, while 67% of CEOs expect AI returns within one to three years^82,83
This plan	Value from month 3 in one service line and the full run rate from month 12, with the payback month following once costs are estimated (illustrative)
Horizons	Read-only and drafting slices (months 3 and 8) judged in months The gated SAP setup, which acts in a system of record, judged on the 12-month horizon
Cost data	AI spend attributed to each engagement from phase 1 (NFR-02) Billing data in the FOCUS format where the providers support it

Why it is filled this way

The first slice passes every quick-win test, so value is visible within a quarter while the platform’s foundations go in.

Capacity and utilization gains start early, from assembling, checking and drafting. The revenue protected by timely charge codes arrives in phase 3, when agents can write to SAP under bound approvals, which is why month 12 carries most of the value.

Cumulative value counts each run rate only from its milestone. Costs are left to the project team’s estimate and are not guessed.

T12 filled: decisions, risks and next steps

Decision	What it covers	Owner	By when
Name an owner	One accountable executive and an AI governance council spanning architecture, security, data and the business, covering agents, models and AI spend, with a sponsor for every agent.	Managing partner, operations	Readout plus two weeks
Fund phase 1	A 90-day program in one service line in one region. Entity resolution and independence agents run in read-only mode, every model and tool call goes through AI Gateway, and KPIs are baselined with finance.	Managing partner, operations	Readout plus two weeks
Adopt the standards	API-led connectivity, MCP and A2A as enterprise policy, with every agent registered and routed through the control plane, whatever platform built it.	Chief architect	Readout plus two weeks
Agree how value is reported	Baseline the KPIs with finance now and report value at every milestone review	Finance partner	Before phase 1 starts

Risks

Risk	Impact	Mitigation	Owner
Hierarchy data quality	A missed subsidiary is a missed conflict	Data-quality rules for freshness, duplicates and missing identifiers Stewardship before phase 1 goes live	Data governance
MCP coverage for data services	The Informatica MCP servers may not cover every master data operation the agents need, such as resolve_entity and get_hierarchy	Map each data operation to an Informatica MCP server or an MDM API during planning, and expose the APIs as MCP tools through MCP Bridge	Lead architect
Evaluation tooling	Promotion depends on evaluation suites, and the firm has no evaluation tool today	Choose an evaluation tool and agree the thresholds before phase 2 Record the results with each promotion decision	Lead architect AI governance council
Connector coverage	The independence system and document management may need custom APIs	Confirm connector coverage for every system and size custom API work in phase 1	Lead architect
Custom policy ownership	Approval binding is a custom Omni Gateway policy, not a MuleSoft product	Agree who builds, tests and supports it before phase 3⁷⁸	Enterprise architect
Adoption	Risk & Quality must trust agent-assembled findings Partners must use mobile approvals	Start in assist mode, show evidence and lineage, and train reviewers and partners	Business owner
Takeover readiness	If the setup agent is stopped, launches in flight could stall at the last step	Write and drill the takeover runbook before tier 3 (OR-03), with finance setting up SAP by hand meanwhile	Finance

Assumptions and dependencies

Assumption. 10,000 launches a year, 5,000 client-facing staff and 1,800 hours per person a year (illustrative until measured).
Assumption. $120 an hour loaded cost and a $200 an hour average rate (illustrative until agreed with finance).
Assumption. About a quarter of launches (2,500 a year) start work before charge codes exist, at about 20 hours each.
Assumption. Values are before platform, model and change costs, which the project team estimates with the client.
Dependency. API access to SAP S/4HANA, Workday, Snowflake, the CRM and the independence system, with private connectivity.
Dependency. Single sign-on with phishing-resistant MFA and step-up authentication for partners.
Dependency. Model contracts under enterprise terms with zero data retention and approved regions.
Dependency. A runtime inside the firm’s VPC for gateways and MCP servers where confidentiality requires.
Dependency. Named data stewards for the client, worker and engagement masters.

The first 90 days

Weeks	Step	Activities
Weeks 1–2	Pick the scope and baseline	One service line in one region. Measure today’s cycle time, effort and exceptions with finance.
Weeks 3–6	Govern the traffic	Run Agent Scanners and register agents with owners and tiers. Route every model and tool call through AI Gateway, and stand up the client hierarchy.
Weeks 7–10	Start read-only	Entity resolution and independence agents at tiers 1–2 in assist mode, with evidence and lineage for reviewers.
Weeks 11–13	Prove and decide	Report the KPIs against the baseline, the run-rate value and the AI cost per launch, then decide on phase 2.

Why it is filled this way

The decisions are the handbook’s three for the first 90 days, plus agreeing how value is reported, which makes the business case provable from the start.

Every risk has an owner and a mitigation, including the implementation ones (MCP coverage for data services, evaluation tooling, connector coverage and custom policy ownership) and takeover readiness before tier 3.

The 90-day plan ends with a decision on phase 2 based on measured KPIs, run-rate value and AI cost per launch.

T13 filled: operating model and adoption plan

Roles and decision rights

Role	Named person or team	Owns	Decides
Accountable executive	Managing partner, operations	The launch program, its funding and its business case	Which phases proceed, and when
AI governance council	Chief architect (chair), CISO, Risk & Quality, data governance, finance and practice leaders	Policy, autonomy tiers, standards and no-go zones, such as “no agent decides independence”	Promotions to tier 3 Exceptions
Business owner	Head of practice operations	The launch KPIs, the value and the process	Which stages stay human-led, AI-assisted or AI-led Acceptance of results
Agent sponsors	One named person per agent (T3d)	Their agent from registration to retirement	Recertification and retirement The first call in an incident
Platform team	Integration and AI platform team (proposed)	Agent Registry, AI Gateway, identity, evaluation and observability	Shared policies and onboarding standards
Security and risk	CISO, with Risk & Quality	Threat model, red-teaming, incident response and drills	Blocking a release Demoting or stopping an agent
Finance	Finance partner	AI cost attribution and value tracking	How value is reported AI budget thresholds
Skills and change	Talent and learning	Training for partners, reviewers, managers and sponsors	When each group is ready to go live

RACI for the control plane

Activity	Accountable	Responsible	Consulted	Informed
Register a launch agent with its owner, sponsor, tier and system card	Agent sponsor	Platform team	CISO Risk & Quality	AI governance council
Promote the setup agent to tier 3	AI governance council	Finance partner (its sponsor)	CISO Risk & Quality Head of practice operations	Managing partner, operations
Recertify or retire a launch agent	Agent sponsor	Platform team	Head of practice operations	AI governance council
Change a shared policy, or approve a new MCP server or model provider	AI governance council	Platform team	CISO Chief architect	Agent sponsors
Stop an agent and run the takeover	CISO	Agent sponsor Finance setup team or practice operations	Platform team	Head of practice operations Managing partner, operations
Report value and AI cost per launch at each milestone	Head of practice operations	Finance partner	Platform team	Managing partner, operations

Training plan

Audience	What they need to learn	When
Partners	Approving letters, budgets and SAP setups on mobile: what an approval binds, when it expires and how to stop a launch	Phase 1 for letters and budgets Before tier 3 for setups
Risk & Quality reviewers	Reading agent-assembled findings critically: evidence, lineage and the limits of the agent	Before phase 1 goes live
Engagement managers	Working with the launch copilot: editing drafts, confirming teams and taking over a launch	Before their service line goes live
Agent sponsors	The agent lifecycle, recertification and the takeover runbook	Before they register their agent
Platform team	Operating Agent Registry, AI Gateway, identity, evaluation and cost controls	Phase 1

Adoption measures

Measure	Definition	Target
Acceptance rate	Letters and budgets accepted without significant edits	Tracked from month 8 Target agreed with Risk & Quality
Takeover rate	Launch tasks a manager or reviewer had to take over	Falling quarter on quarter
Verification latency	Time from an assembled independence finding to the reviewer’s decision	Very short times reviewed, as they can mean approval without review
Trained before go-live	Partners, reviewers and managers trained before their service line goes live	100%

Why it is filled this way

Every agent has a named sponsor, and the council’s first no-go zone is written down: no agent decides independence.

The RACI makes the one step that raises autonomy explicit: promoting the setup agent to tier 3 needs the council, with the CISO and Risk & Quality consulted.

Adoption is measured by acceptance and takeovers, not logins, and partners are trained before they approve their first SAP setup on mobile.