Many organisations can tell you how much they spend on AI, how many licences they have bought and how many employees have tried a copilot. Far fewer can show what improved as a result.
That gap matters. Usage is evidence that a tool is being used. It is not evidence that the organisation is more productive, that customers are better served or that the investment is earning its keep.
A credible AI business case starts with a problem, not a model.
Start with the cost of the problem
Choose a workflow where the current friction is visible: slow case handling, repeated document review, long software review cycles, expensive support escalations or time-consuming internal search.
Before selecting a tool, document the baseline. How many cases arrive each week? How long does handling take? What proportion require rework? How often are mistakes made? What does a delay cost? If nobody can answer these questions, the first task is measurement—not buying more AI.
Build a one-page business case
Every AI initiative should have five things:
- An accountable owner. One person responsible for the outcome, not just the rollout.
- A baseline. A reliable picture of performance before the change.
- A target outcome. A measurable improvement linked to the business problem.
- Guardrails. Quality, safety, privacy, compliance and service levels that must not deteriorate.
- A decision date. When the team will decide to scale, adjust or stop.
For example, a support team might test whether AI-assisted drafting reduces median handling time by 15% while keeping quality scores stable and increasing neither complaint rates nor escalations. The 15% is a proposed experiment target, not a forecast. The baseline and the actual result determine whether the intervention worked.
Measure the full cost, not just the licence
The total cost of an AI workflow can include model usage, licences, integration, data preparation, security review, evaluation, human verification, exception handling, training and ongoing monitoring. Hidden review work can erase the apparent saving from faster generation.
Calculate value conservatively. If a task saves five minutes but requires four minutes of extra checking, the net time saving is one minute—not five. If the workflow is faster but produces more costly errors, the financial case may be negative.
Distinguish capacity released from cash saved. Giving a team back ten hours a week is valuable, but it is not automatically a ten-hour reduction in payroll. State how the capacity will be used: shorter customer wait times, more cases handled, reduced backlog or work that would otherwise have required additional hiring.
Use a balanced scorecard
For each initiative, select one primary outcome measure and a small number of guardrails. Examples include:
- Productivity: cycle time, handling time or throughput per person.
- Quality: defect rate, correction rate, audit score or customer satisfaction.
- Reliability: failure rate, escalation rate, rollback rate or service availability.
- Experience: time to resolution, abandonment or employee effort.
- Economics: cost per completed task and total cost to operate.
Do not choose ten metrics simply because they are available. Pick the few that determine whether the intervention is worth keeping. Define each measure before the pilot so the team cannot move the goalposts after seeing the results.
A 90-day decision cycle
Days 1–14: establish the baseline. Select one workflow, define task boundaries and capture volume, time, quality and cost.
Days 15–70: run a controlled pilot. Introduce the AI-assisted process to a suitable group or task type. Keep safety controls in place and record exceptions, rework and human checking.
Days 71–90: make the decision. Compare the results with the baseline, account for seasonality and task mix where possible, and review whether guardrails held. Decide whether to scale, revise the workflow or stop.
A pilot that fails to meet its target is not necessarily wasted effort. It becomes waste when nobody learns from it, nobody owns the decision, and the initiative continues indefinitely because it is fashionable.
Prove, scale, stop
Prove that the workflow improves a meaningful outcome. Scale only when the result is repeatable and the controls are sustainable. Stop when evidence does not justify the cost, risk or complexity.
This discipline is more useful than an AI adoption target on its own. Adoption can be a leading indicator; business value is the result to optimise.
The question for leadership teams: which AI initiative would you fund again today if you had to justify it using measured outcomes rather than enthusiasm and usage numbers?