Section 6

Phasing for early value

CIOs and CTOs are right to worry about returns. In a survey of more than 4,400 chief executives published in January 2026, 56% said AI had delivered no significant financial benefit to date.⁷⁹ In a mid-2026 survey, 37% of respondents attributed any EBIT impact to AI,⁸⁰ and many agentic AI projects are expected to be canceled over cost, unclear value or weak risk controls.⁶ The response is to keep the ambition and improve the sequence: adopt the reference architecture in phases, and put a slice of a real use case into production in every phase, so value arrives from the first milestone while the platform matures.

The four platform phases

Put visibility and one policy point first, then reusable capabilities and routing, then trust, and leave federation for last. Published phasing guidance likewise starts with governance, a single policy point and baseline guardrails, and federates only at scale.²⁰

Phasing the use case

The same pattern works inside each use case, with milestones at months 3, 8, 12 and 24 (template T11):

• Month 3 (phase 1). Start read-only or propose-only, in one scope (a service line, market, region or site), with every model and tool call through the gateway from day one.
• Month 8 (phase 2). Widen the scope and add drafting and proposals. Publish the use case’s capabilities as governed APIs and MCP tools the next use case can reuse.
• Month 12 (phase 3). Let agents act in systems of record, but only under approvals bound to each action, limits that hold across sessions, and a kill switch people have drilled. This is where most of each business case sits.
• Month 24 (phase 4). All goals met: the run rate is sustained, adjacent processes are added on the same controls, and models can be swapped without application changes.

Setting expectations with payback benchmarks

Finance will test the milestone values against published payback data, so show the benchmark first, and how phasing beats it (T2 and T11):

Phasing is how a program beats the benchmark: read-only and propose-only slices release capacity within months and help fund the controls that the larger gains depend on. Judge the early slices on months, and the agentic ones that act in systems of record on a longer horizon.

Quick wins first

Early quick wins keep sponsorship. A good first slice passes every test below. If it does not, shrink the scope or lower the autonomy until it does.

• Shows measurable value within 90 days in one scope (one service line, market, region or site).
• Runs read-only or propose-only (tiers 1 and 2): no writes to systems of record yet.
• Uses systems and data that are reachable now, or with one connector.
• Saves time or money that the users and the sponsor can see.
• Leaves a reusable foundation behind: a registered agent, a governed MCP tool, a golden record or the gateway itself.
• Has a baseline taken before it starts, so the movement can be proved.

Typical quick wins come from the platform and from the use case. The platform delivers visibility on its own: every agent registered with an owner and a tier, duplicate and unused agents retired, and model spend visible by use case. The use case delivers capacity from assembling, checking and drafting, which agents can do safely in read-only or propose-only modes. The larger gains usually arrive in phase 3, once agents can act in systems of record under the approvals and limits that phase puts in place.

KPIs

Set three kinds of KPI, each with a baseline, a target for every milestone, a data source and an owner (template T6):

• Business outcomes. The measures the sponsor will judge success by: cycle time, effort, error or exception rates, revenue or cost effects.
• Quality and adoption. Evaluation pass rates, first-time-right rates, the share of work flowing through the new process.
• Control plane. The scorecard below, reported monthly to the executive committee and quarterly to the AI governance board.²⁹

Leading indicators

The scorecard’s metrics are outcomes a board can read, and they move slowly. Delivery teams also need early signals that move within weeks, defined so that every team computes them the same way. Evaluate the trajectory as well as the answer: whether the agent chose the right tools, in the right order, at an acceptable cost.⁶⁰ Measure value instead of adoption. In BCG’s 2026 survey of close to 12,000 employees and leaders, the share whose organizations had put agents into workflows rose from 13% to 30% in a year, yet half said their companies lack clear governance for teams of people and AI.⁸⁸ Use the library below in T6:

AI cost visibility

Cost is where the control plane pays for itself soonest, and where most clients are blind: in KPMG’s second-quarter 2026 pulse, only 26% of organizations had full, real-time visibility into what their AI systems cost to operate, and in the third quarter 43% had set usage or token budgets.⁸ Treat cost visibility as an active control: tag every model and tool call to an owner, a use case and a cost center, alert as spend nears its budget,¹⁴ and compare the cost of each completed outcome with the cost of the same work done by people, because the economics depend on the whole system of people and agents, not on the model bill alone.²² Ask the client for billing data in the FinOps Foundation’s FOCUS format where its providers support it, and use the FinOps unit measures, such as cost per inference and per token, anomaly detection and time to business value.²³

ROI and business value

Build the value side from the business case and the phasing, then add the cost side with the project team:

1. Run rate at each milestone. Volume in scope × improvement per item × value per unit, for each lever live at that milestone.
2. Cumulative value. Count each run rate only for the months it is in force: for milestones at months 3, 8, 12 and 24, that is five, four and twelve months.
3. Costs. Platform subscriptions, implementation services, model usage, infrastructure, and change and training, from the project team’s estimate and the client’s own costs.
4. ROI and payback. ROI = (cumulative value − cumulative cost) ÷ cumulative cost. Payback is the month in which cumulative value first exceeds cumulative cost.
5. Reporting. Agree the method with finance, report value realized at every steering review, and keep values illustrative until the baseline is measured.

For use case 1, the run rate is about $0.5 million a year at month 3, $8 million at month 8 and $30 million from month 12. Counted from each milestone, cumulative value by month 24 is about $0.2 million + $2.7 million + $30 million, or about $33 million, before platform, model and change costs (illustrative).

Business value beyond the numbers

• Risk avoided. Decisions made by the right person before work starts, with evidence and lineage.
• Compliance evidence. Audit trails that answer regulators and inspectors in hours instead of weeks.
• Experience. Faster answers for clients and fewer chasers for staff.
• Reuse. Every governed API, MCP tool and golden record makes the next use case faster and cheaper to deliver.

The operating model

Technology is the smaller part of the work: BCG estimates that becoming AI-first “involves 30% technology and 70% people and organization.”⁸⁵ In Microsoft’s 2026 Work Trend Index, only one in four people using AI at work said their leadership is clearly and consistently aligned on AI,⁸⁶ and in Google Cloud’s 2026 survey, 48% of the organizations whose AI returns are accelerating had “extremely clear” ownership and decision rights for AI and agentic initiatives.⁸⁴ Agree who owns what before the agents go live (T13):

Build skills in step with the plan: approvers trained on what they approve, sponsors on the agent lifecycle, and business teams on working alongside agents and taking over when one is stopped. Measure adoption by value and acceptance instead of counting users or prompts.