Section 2

Framing the use case and the value hypothesis

Framing decides everything that follows. The goal of the first workshop is one use case with a named sponsor and business owner, a clear trigger and end state, a scope small enough to start in 90 days, and a value hypothesis the business owner believes enough to measure.

Choosing the use case

Score each candidate from 1 to 5 against the criteria below, with evidence. As a rule of thumb, a strong first use case scores 28 or more out of 35, and nothing below 3 on a safe first step or reachable systems. Fix those first or choose another candidate.

Criterion	What good looks like
Value at stake	Volume times the time, money or risk per item is large enough to matter to the sponsor.
Visible pain	The people doing the work, and their leaders, already feel the delay, cost or risk.
Reachable systems	The systems of record have APIs or connectors, or can get them within the first phase.
Data readiness	The master data the agents need exists and can be made trustworthy.
Safe first step	The first phase can run read-only or propose-only, with people keeping the decisions.
Measurable in 90 days	A baseline exists or can be taken now, and the first slice can show movement within a quarter.
Sponsor and owner	An executive sponsor and a business owner are named and available.

Describing the use case at a glance

Capture the use case in the client’s words, on one page (template T1):

Problem today. The chain of manual, risk-sensitive steps and what the delays and errors cost.
Target state. What the agents will do, what the people will keep and which systems are involved.
Trigger and end state. The event that starts the process and what is true when it is done.
Users and systems of record. Who takes part, and the systems the process runs on.
Out of scope. What the first phase excludes. Saying it early prevents scope creep later.
Stages. Five to seven stages from trigger to end state, each with its human step, its work mode (human-led, AI-assisted or AI-led), its agent and autonomy tier, and its system of record.
Goals. Business goals the sponsor will judge success by, and control-plane goals: governed traffic, cost per outcome, evaluation before promotion and oversight people can exercise.

Deciding which work stays human

Decide which work should stay human before deciding how to automate the rest. BCG suggests asking two questions of every step: could AI do it as well as or better than people, and should it, given regulatory risk, customer trust, ethics and safety?⁸⁵ The answers sort each stage into one of three work modes (T1):

Work mode	What it means	Typical steps	Tiers and controls
Human-led	A person does the work and decides AI may inform	Steps that carry professional, regulatory or safety accountability, or where trust depends on a person	Tier 1 at most: agents assemble and check, and a person decides
AI-assisted	AI drafts, proposes or checks A person reviews and accepts	Judgment work with clear criteria and a reviewer who can check the result	Tiers 1 and 2: drafts and proposals, with bound approvals for anything that writes
AI-led	Agents do the work within bounds People supervise and handle exceptions	High-volume, rule-based work whose results can be checked and reversed	Tiers 2 and 3: limits, monitoring, a kill switch and a takeover runbook

Ask the should-AI questions of every stage an agent could do:

Regulatory risk: does a law, regulator or professional standard expect a person to decide?
Customer trust: would the customer or client expect a person here?
Ethics: could the decision treat people unfairly, and could anyone explain it?
Safety: could an error harm people, assets or the environment, and can it be reversed?

Record the reason for every human-led step, since it is what risk teams and boards ask about first. Record the split before any decision on roles, too: Deloitte’s 2026 Human Capital Trends found that 57% of leaders who cut staff because of AI have come to question that decision.⁸⁷

Writing the value hypothesis

Write the business case as a small set of value levers, each with a formula the business owner and finance can check (template T2). Use the client’s volumes and rates, keep the assumptions conservative, and show the arithmetic.

Value lever	What it measures	Formula
Capacity released	Hours of manual work removed, valued at loaded cost	Volume × hours saved × loaded cost per hour
Revenue protected or accelerated	Revenue that is no longer lost, written off or delayed	Volume affected × revenue per item recovered or pulled forward
Utilization and throughput	More output from the same people or assets	Points of utilization × people × hours × rate
Cost avoided	Penalties, fails, outages, rework or external spend avoided	Events avoided × cost per event
Risk avoided	Regulatory, compliance, safety or reputational exposure reduced	Usually not quantified Describe and evidence it

Show one formula per lever, with its inputs and their source.
Separate run-rate value (a year at full scale) from one-time value. The plan phases the run rate.
Quote values before platform, model and change costs until the project team has estimated them. Then show net value and payback.
Describe value that is hard to quantify, such as risk avoided, in words and evidence, without inventing a number.
Set the case against published payback benchmarks, and judge generative and agentic slices on separate horizons (Section 6).
Label every figure illustrative until the baseline is measured and finance agrees the method.