Section 11

Reference use case 1: Professional services engagement launch

This reference use case applies the control-plane architecture to one of the most frequent, compliance-sensitive processes in a global audit, tax and advisory firm: accepting and launching a new client engagement. The architecture and discussion are vendor-neutral except for the transactional systems the process runs on: a CRM, SAP S/4HANA for projects and billing, Workday for people and skills, and Snowflake for analytics. The final subsection shows one way to realize the design with MuleSoft Agent Fabric and Informatica.

Overview

Large professional services firms launch thousands of engagements a year. Between a verbal win and the first chargeable hour sits a chain of manual, risk-sensitive steps: resolving exactly which legal entity the client is, checking conflicts of interest and auditor independence, drafting the engagement letter, building the budget, staffing the team, and creating the project, work breakdown structure (WBS) and billing plan in the ERP system. Each step lives in a different system and is owned by a different function, so the chain commonly takes days to weeks. The delays have direct costs. People wait on the bench, and work that starts before charge codes exist goes unbilled. A missed affiliate relationship can also become an independence breach with regulatory consequences when the client, or its parent, is an audit client.⁸⁹

In the target state, an Engagement Launch Orchestrator coordinates six specialist agents across these systems. People keep every judgment that carries professional or regulatory accountability: independence decisions, engagement terms and partner approval. Agents do the assembly, checking, drafting and system updates, 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.
Out of scope: delivery of the engagement itself and invoicing after launch.

Engagement launch stages with agents, human decisions, control plane, MCP tools and systems — *Figure 8. 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.*

Business case and expected outcomes

The value comes from compressing cycle time, removing manual assembly work and preventing compliance failures. The figures below are illustrative for a firm launching 10,000 engagements a year, to be replaced with the firm’s own baseline during discovery.

Value lever	How the solution delivers it	Illustrative annual value
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 ≈ $10.8M
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 ≈ $10.0M
Utilization	Teams are confirmed within days instead of weeks, reducing bench time	0.5 point × 5,000 client-facing staff × 1,800 hours × $200 ≈ $9.0M
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)

On these assumptions the program releases or protects roughly $30 million a year before platform, model and change costs, which should be estimated alongside the baseline. The business should also expect these outcomes:

Faster time to revenue, with verbal win to ready-to-charge in about two business days.
Lower cost to launch, with Risk, resourcing and finance working exceptions rather than re-keying data.
No known independence exceptions at engagement start, with a complete evidence trail for inspections.
A better client experience (engagement letter the same day) and a better staff experience (clear assignments, fewer chasers).

Key performance indicators

Baselines below are illustrative and should be measured during discovery.

KPI	Definition	Baseline (illustrative)	Target (12 months)
Win-to-ready-to-charge	Business days from verbal win to active charge codes	10 days	2 days
Manual effort per launch	Hours across all roles to launch one engagement	14 hours	5 hours
Independence cleared before start	Share of engagements cleared before any work begins	92%	100%
Post-start independence exceptions	Exceptions found after work begins, per 1,000 engagements	4	Below 0.5
First-time-right setup	SAP projects created without later correction	80%	97%
Staffing within 48 hours	Roles confirmed within 48 hours of verbal win	55%	85%
Agent quality	Evaluation pass rate for letters and independence summaries	Not measured	95% or higher
Governed AI traffic	Share of model and tool calls through the AI gateway	Not measured	100%
AI cost per launch	Model and platform cost per engagement launched	Not measured	Within target

Solution architecture

The architecture follows the five-layer reference architecture in Section 6, expanded to show the orchestration model, the MCP and A2A paths, the data-governance layer and the security controls that apply at every layer (Figure 9).

Layered vendor-neutral architecture for the engagement launch use case — *Figure 9. Use case 1 vendor-neutral solution architecture. Only the systems of record are named products.*

Clients and channels

Engagement managers use a chat copilot inside the firm’s collaboration tool, and partners approve from their phones with step-up authentication. Independence reviewers stay in their case workspace, while clients sign through an e-signature portal. Every request enters through the orchestrator and the control plane, so no user talks to a model directly.

Agents and autonomy tiers

Agent	Responsibility	MCP tools	Tier	Human role
Launch Orchestrator	Plans and tracks the launch Delegates tasks over A2A Enforces the regulated sequence	Registry lookup Task state	Tier 2	Sees plan and status
Entity Resolution	Matches the client to its MDM golden record Returns hierarchy and identifiers	resolve_entity get_hierarchy	Tier 2, read-only	None
Conflicts & Independence	Runs conflict and independence searches Assembles findings and precedent	search_restricted check_services get_relationships	Tier 1	Risk decides
Scoping & Letter	Drafts scope and engagement letter from approved templates and clauses	search_knowledge get_template create_document	Tier 1	Manager edits Partner approves
Pricing & Budget	Builds the budget from rate cards and historical actuals Checks margin	get_rate_card query_actuals	Tier 2	Manager accepts
Staffing	Proposes a team by skills, availability, location and independence clearance	find_available_staff get_clearance_flag	Tier 2	Manager confirms
Engagement Setup	Creates project, WBS, billing plan and charge codes	create_project create_wbs create_billing_plan	Tier 3, policy-gated	None (gated by policy)

Orchestration model selected

The firm uses the hybrid/federated pattern, in which the control plane centrally governs identity, policy, cost and audit and each function owns its agents. Risk & Quality owns the Conflicts & Independence agent and its rules, Finance owns Engagement Setup, and Talent owns Staffing. Within a launch, the orchestrator acts as supervisor, decomposing the launch into tasks and delegating them to specialist agents over A2A,²⁷ and agents reach tools and data only through MCP servers.²⁶ The regulated sequence is deterministic. No letter is sent before independence is cleared, and nothing is written to SAP before the partner and the client have approved. Model reasoning works inside the steps (resolving, summarizing, drafting, matching) and can never skip a gate.

Models and routing

Frontier model (primary): hosted under enterprise terms with zero data retention, it drafts engagement letters and summarizes complex findings. A second frontier provider is configured for failover.
Open-weight small language model: self-hosted in the firm’s virtual private cloud for entity extraction, classification and personal-data detection, so client-confidential text never leaves the firm for routine tasks.
Embedding model: on a private endpoint to index credentials, prior proposals and templates for retrieval.
Routing policy in the AI gateway: client-confidential payloads go only to private models or approved regions, and drafting goes to the frontier model with personal data masked. Each launch has its own token budget.

Control plane and policy

The control plane provides these services:

A registry of agents, MCP servers and APIs, each with an owner and an autonomy tier.
An AI gateway that mediates all LLM, MCP and A2A traffic.
A policy decision point that evaluates policy-as-code.
Identity services that issue per-agent workload identities and delegated, on-behalf-of tokens.
Evaluation, cost and observability services, and a kill switch with circuit breakers.

Representative policies for this use case:

Engagement setup in SAP requires independence status CLEARED, partner approval and a signed letter. The partner’s approval is bound to the exact setup request and expires at the end of the business day.
Engagement letters use approved templates, and any clause deviation routes to Risk & Quality.
Staffing proposals include only people with independence clearance for the client and the required credentials.
Client-confidential data is processed only by private models in approved regions.

MCP servers, APIs and systems of record

Each system of record is wrapped by governed APIs and exposed to agents as MCP tools with explicit read or write scopes: the CRM (opportunity and account), the independence and conflicts system (searches and a clearance flag), document management and e-signature, Workday (skills, certifications, availability), SAP S/4HANA (rate cards, projects, WBS, billing plans) and Snowflake (historical engagement economics through governed semantic views, read-only).

Master data and semantics

Independence and conflict checks depend on the entity data behind them, because a missed subsidiary means a missed conflict. Master data management therefore sits on the critical path:

Client and legal-entity master: one golden record per legal entity with the full corporate hierarchy (ultimate parent, affiliates, subsidiaries), identifiers such as the LEI and tax IDs, and links to the restricted-entity list.
Worker master: Workday remains the system of record, and MDM cross-references worker identifiers across the CRM, SAP and the independence system.
Engagement master: one engagement ID shared by the CRM opportunity, the independence case, the engagement letter and the SAP project.
Reference data: service catalog, service lines, industries and rate cards, with effective dates.

A semantic layer gives agents and analytics the same meaning for the same words: a business glossary (client, affiliate, audit client, restricted entity, engagement, realization, utilization), shared metric definitions, and a data contract for each MCP tool (schema, meaning and quality thresholds). A catalog records lineage from source to agent output, so a reviewer can see which records drove an independence finding. Data-quality rules watch hierarchy freshness, duplicate entities and missing identifiers, and access policies enforce classification, masking and information barriers between engagements.

Security at every level

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

Layer	Controls
Users	Single sign-on with phishing-resistant MFA Step-up authentication for partner approval Role- and attribute-based access
Agents	A unique workload identity per agent instance⁹¹ Least-privilege scopes Autonomy tiers Kill switch No agent approves its own output
A2A	Registered, signed agent cards Mutual TLS plus OAuth tokens on every call²⁷ Delegation chains carried through on-behalf-of token exchange⁹² with narrowed scopes Only registered agents may call each other
MCP	MCP servers sit behind the gateway as OAuth-protected resources Audience-bound tokens with no pass-through Per-tool scopes Input and output schema validation²⁶ As firm practice, allow-lists per agent, reviewed and pinned tool definitions to prevent tool poisoning, and separate read and write tools
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
Models	Access only through the gateway Enterprise terms with zero data retention and no training on firm data Prompt-injection and jailbreak screening of retrieved content Personal data masked before prompts Output validation⁹³
Data at rest	AES-256 encryption with customer-managed keys in a KMS or HSM Field-level encryption or tokenization for personal data such as personal financial interests in independence declarations Snowflake row-access and masking policies Vector indexes partitioned by client and engagement to preserve information barriers
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
Audit	Tamper-evident logs linking user, agent, tool call and record Retention aligned to professional-standards documentation requirements

Governance and observability

Risk & Quality owns the independence agent’s rules and approves changes to them, while an AI governance council approves autonomy tiers and new agents. Before promotion, every agent passes evaluation suites for letter clause accuracy, independence-finding recall and entity-resolution precision. Every hop emits OpenTelemetry traces using the generative-AI semantic conventions,⁹⁴ tied to one engagement correlation ID, so operations can see cycle time by stage, policy blocks, token cost per launch and evaluation drift on one dashboard.

Maya’s day as an engagement manager

Maya Chen is an engagement manager in the advisory practice. The timeline follows an illustrative launch for a new client, Project Atlas (Figure 10).

Timeline of eight moments in the engagement manager’s day — *Figure 10. A day in the life of an engagement manager launching Project Atlas (illustrative).*

At 08:05 the CRM opportunity for Project Atlas moves to verbal win, and within a minute Maya’s chat copilot shows a launch plan targeting ready-to-charge in two days. By 08:12 the entity-resolution agent has matched the client to its golden record and found that it is a subsidiary of a group whose parent is an audit client, so an enhanced independence review starts automatically. At 09:40 a reviewer in Risk & Quality approves the engagement with one condition: no work on financial-reporting systems.

The scoping agent drafts the engagement letter with that carve-out by 10:30, and Maya rewrites two paragraphs. The pricing agent’s budget arrives at 11:20 with a 31% margin against a 30% target. At 13:05 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. The partner approves the letter and budget on mobile at 14:30 with step-up authentication, and the letter goes to the client for signature. At 16:10 the client signs, the setup agent creates the SAP project, WBS and billing plan, and the team receives charge codes. The whole launch, which used to take about two weeks, finished the same day.

How the components work together

Figure 11 traces the same day through the architecture, and the table explains what each component did at each moment and which controls applied.

Sequence diagram of agents, control plane, models, MCP servers and systems — *Figure 11. Component interaction for the Project Atlas launch. Numbered steps follow the day in the life.*

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 MDM MCP tools (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

Realizing the use case with MuleSoft and Informatica

Every component above can be assembled from multiple vendors. MuleSoft Agent Fabric and Informatica cover the integration, control-plane and data-governance layers as one stack, with Informatica as the data foundation, while the firm keeps its choice of collaboration tools, models, agent frameworks and systems of record (Figure 12).

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 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 Mobile approval for high-risk steps²¹ A custom Omni Gateway policy 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 Attribute AI spend to each launch Letters and findings are tested before release with the firm’s evaluation tool²¹
MCP servers and APIs	MCP Bridge 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 APIs⁷⁷
Semantics, catalog, lineage	Data Governance and Catalog, with the CDGC Metadata Search MCP server (Informatica)	Shared glossary and lineage for agents and people Agents find data assets by searching the catalog⁷⁷
Data quality	Data quality, with the Address Verification MCP server (Informatica)	Validate entity data and addresses at the point of entry Route exceptions to data stewards⁷⁷
Deployment	CloudHub or customer-managed Runtime Fabric (MuleSoft)	Keep gateways and MCP servers inside the firm’s VPC where confidentiality requires²¹

Reference architecture with MuleSoft and Informatica components highlighted — *Figure 12. Use case 1 realized with MuleSoft Agent Fabric and Informatica. Colors show which vendor fills each role.*

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. The firm’s own agents, models and user tools plug into both.⁷⁷

Delivering the launch in phases

The use case is delivered in the four phases of Section 8, so value arrives at every milestone rather than at the end. Each step widens the scope or raises the autonomy tier only after the previous one has proved its value and its controls.

Milestone	What goes live	KPIs at the milestone	Value (illustrative)
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. 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)
Month 8 (phase 2)	Half of all launches. Scoping and pricing agents draft letters and budgets from approved clauses and rate cards. 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, utilization about $4.5M)
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
Month 24 (phase 4): all goals met	Run rate and KPI targets sustained. Adjacent processes, such as engagement changes and closure, added on the same controls. Models swapped without application changes.	All KPIs held at target	About $33M cumulative by month 24

Values are illustrative and before platform, model and change costs. Cumulative value counts each run rate from the milestone at which it was reached. 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 S/4HANA under bound approvals.