Twelve months after an AI initiative launches, someone asks a simple question: what did we get for it? Too often the answer is a list of activity. Licences bought, pilots launched, people trained. The value may be real, but nobody set up the measures to show it.
This two-part series sets out a practical way to measure AI value. Part 1 covers what needs to be in place before launch. Part 2 covers how to review, calculate and decide once the work is running.
Three reasons value goes unmeasured
No baseline
Performance before launch was never recorded, so any improvement afterwards is an estimate that is easy to dispute.
Activity, not outcomes
Logins, prompts and pilots show use, not value. They are useful signals, but they are not results.
No owner
Nobody is accountable for the number, so it is not reviewed and nothing changes when it slips.
None of these is a technology problem. Each is a management decision that was not made before the money was spent.
Measure the whole chain
AI value builds in four layers. Each depends on the one before it, so measure all four and read them from left to right.
Adoption
Is it being used as intended?Active users as a share of the target group. Share of eligible work that goes through the tool.
Performance
Is the work better?Cycle time, error and rework rate, first-contact resolution, throughput per person.
Outcome
Did the business result move?Cost per transaction, backlog, customer satisfaction, revenue per account.
Value
What is it worth, net of cost?Annual net benefit, payback period, return on investment.
The measurement chain. Example measures for each layer.
Reading the chain in order turns a disappointing result into a diagnosis:
Set the baseline before launch
Five decisions, made before go-live, separate an initiative that can show its value from one that can only describe it.
- 1Name the outcome and its ownerState the business result the initiative exists to move, and name the person accountable for it. If no one will own the number, the initiative is not ready.
- 2Choose a few measuresPick one or two measures for each layer of the chain. A short list gets reviewed; a long one gets ignored.
- 3Record todayMeasure current performance over a representative period, such as the last full quarter, using the same definitions you will use after launch.
- 4Agree the decision ruleDecide in advance what result means continue, adjust or stop, and by when. This keeps the later review about evidence, not advocacy.
- 5Fix the data sourceConfirm where each number comes from, who reports it and how often, before the first review is due.
You can often recover one. Use system data from the period before launch, or compare against a team or region that has not adopted the tool yet. Record the method so the comparison can be repeated.
Choosing measures by function
The right KPIs depend on the workflow. The examples below are starting points to adapt, not targets. Choose measures you can track the same way before and after launch.
| Function | Adoption signal | Performance KPI | Business outcome |
|---|---|---|---|
| Customer service | Share of inquiries handled with AI assistance | Average handling time; first-contact resolution | Cost per contact; customer satisfaction |
| Finance | Share of invoices processed without manual keying | Processing time per invoice; exception rate | Cost per invoice; days to close the month |
| Operations and supply chain | Share of planners using forecasts in decisions | Forecast accuracy; late schedule changes | Inventory carrying cost; on-time delivery |
| Human resources | Share of policy questions answered through self-service | Time to answer; time to fill roles | HR cost per employee; new-hire retention at 90 days |
| Sales and marketing | Share of proposals drafted with AI support | Proposal turnaround time; content production time | Win rate; cost per qualified lead |
| IT service desk | Share of tickets triaged automatically | Mean time to resolve; reopen rate | Cost per ticket; user satisfaction |
| Public sector and programs | Share of applications pre-screened with AI support | Processing time per application; error rate | Backlog size; service standard met |
| Legal and compliance | Share of standard contracts reviewed with AI support | Review time per contract; issues found at final review | External legal spend; contract cycle time |
With the baseline set and the measures chosen, the initiative is ready to launch in a way that can show what it returns. The next step is to review it with discipline.