Many AI projects cannot show a return, not because they failed, but because nobody decided in advance what return they were for. The model works, the demonstration impresses, and six months later no one can say whether the business is better off. Measuring return starts before the project does.
This piece sets out how we approach it, with real figures from systems we have shipped.
Pick one number the project has to move
Before any build begins, name the metric the project exists to improve, and record where it stands today. That baseline is the most important number in the whole project. Without it, every later claim of improvement is a guess.
One number is better than several. A project that promises to improve everything can always claim success somewhere. A project committed to one number is held honestly to account.
Know which kind of value you are after
AI projects tend to create value in a small number of distinct ways, and each is measured differently.
- Time returned. Work that used to be manual no longer is. Measure hours saved, and what those hours are now spent on.
- Revenue created. More pipeline, better conversion, new products. Measure revenue directly attributable to the system.
- Cost removed. A cheaper way to do something that already had to be done. Measure the cost of the old approach against the new.
- Quality improved. Fewer errors, better answers, faster decisions. Measure error rates or the downstream cost of mistakes.
Real examples from our work
These are the figures from projects we have shipped, each measured against a baseline.
Time returned
- Our AI Sales Development Representative took most of the manual prospecting off the team.
- Megan gives each healthcare provider hours back and handles a good share of the routine calls.
- Predictive Maintenance shortened the planning cycle.
Revenue created
- Multimodal Synthetic Surveys opened up a line of work the team could not take on before.
- The AI SDR put more worthwhile meetings in the calendar.
- Virtual try-on got more shoppers clicking through.
Cost removed
- The synthetic survey work replaced panels costing ten times as much.
- Our Dashboard & Billing System brought inventory costs down.
- Predictive maintenance cut down unplanned stoppages.
Quality improved
- Nexo made answers far more precise and cut the time spent hunting for them.
- Last-Mile Delivery Optimization got more deliveries there on time and shortened the routes.
Count the full cost
A return only means something against the full cost, and the full cost is larger than the build. Include:
- Development, including discovery and evaluation, not only engineering.
- Model and API usage, which scales with volume and can grow quickly.
- Infrastructure to run and store everything.
- Integration with existing systems, which is often underestimated.
- Ongoing monitoring and maintenance, which continues for the life of the system.
Projects that look inexpensive often become expensive in the last two.
Measure in production, not in testing
Test results show what a system can do. Production results show what it actually does. Measure the target metric where the business experiences it, and keep measuring after launch, because performance changes as data and conditions change.
Account for time to value
A system that returns a great deal eventually may be worth less than one that returns a moderate amount quickly. Consider how long each option takes to deliver its first measurable result. Starting narrow, proving value and expanding usually beats a large programme that delivers everything at once, or nothing.
The honest answer is sometimes “don’t”
Sometimes the calculation says AI is not the right tool for the problem: the volume is too low, the data is too poor, or a simpler solution would do. Finding that out before building is itself a return, and it is a conversation we would always rather have at the start.