Measuring the return on an AI project matters as much as the project itself, yet it gets far less attention. Pilots start with enthusiasm, and a few months later nobody can say clearly whether the project actually delivered anything. That uncertainty hurts both ways: a project that works doesn't get enough support, and one that doesn't keeps quietly generating costs. In this post we propose a simple, practical framework for measuring AI projects.
You can't talk about returns without a baseline
The most common mistake is not recording the situation before the project starts. A feeling that "this used to take forever" is not enough for comparison. Before the project begins, collect at least a few weeks of the following data:
- Monthly volume of the work (how many emails, quotes, documents)
- Average time per item (even a rough estimate)
- Waiting time (for example, from when a request arrives to the first reply)
- Error or correction rate
- Who does the work, and in which role
Most of this data already sits in your mailbox, CRM or ERP. Collecting it doesn't take a big effort; what matters is doing it before the project. We explain how to choose your first project in our post on where to start with AI.
Which metrics should you look at?
Calculating the return only in terms of "hours saved" falls short. We recommend looking at four groups:
| Group | Example metric | Where it is measured |
|---|---|---|
| Time | Time per item, first response time | System logs, timestamps |
| Quality | Error rate, need for corrections, customer complaints | Approval records, support tickets |
| Capacity | Volume handled by the same team | Monthly volume comparison |
| Revenue impact | Quote conversion rate, number of missed requests | CRM, sales reports |
For each project, choose one primary metric from these groups and let the rest play a supporting role. In an email classification project, for example, the primary metric could be "first response time to quote requests." Several primary metrics make the decision harder when results are mixed.
Where does the saved time go?
An often overlooked point: when AI frees up a few hours in an employee's week, those hours don't automatically turn into value. If it isn't clear where that time goes, the return stays on paper. So when you launch the project, also answer "what will the freed-up time be used for?": getting back to more customers, faster quotes, finishing postponed work and so on. Your measurement should look at that goal too.
Account for the hidden costs
The other half of the return calculation is cost. Items commonly left out of AI projects:
- Review time. Reading and approving the output takes time too. In the first months, it may take longer than expected.
- Usage fees. AI services are often billed by usage; as volume grows, so does the cost.
- Maintenance. Updating instructions, keeping the knowledge source current and monitoring integrations require ongoing effort.
- Training and adjustment. The team may need a few weeks to get used to the new way of working, and productivity may dip during that period.
- Cost of errors. A wrong draft that slips through can lead to an outcome that is hard to reverse.
Estimating these items upfront puts the project on a realistic footing.
Decision time: continue, fix or stop
At the end of the pilot period, you can use a simple set of questions to choose one of three options:
- Is there a meaningful improvement in the primary metric?
- Has anything gotten worse in the quality metrics?
- Is the gain still visible once hidden costs are included?
- Does the team want to keep using the new way of working?
If the answers to the first three are positive, continue and expand the scope carefully. If there is improvement but quality is a problem, fix the instructions and sources and run another short pilot. If there is no improvement, stopping is also a successful outcome; you have learned what doesn't work.
Tie measurement to your reporting
Collecting metrics by hand gets abandoned after a few months. If possible, connect the measurement to your existing reporting: response times from the mailbox, quote conversions from the CRM, and revenue and collection impact from the ERP should arrive automatically. On the dashboard and ERP side, we turn such metrics into a regular flow in our ERP reporting work.
How we do it at Globya
When we propose an AI project, we write the measurement plan along with it: baseline data, primary metric, pilot period and decision date. As part of how we work, we share progress through concrete outputs. As your digital IT consultant, our goal is not to put AI everywhere, but to make it permanent where it truly delivers value, and to say so plainly when it doesn't.
Frequently asked questions
How long does it take to see the return on an AI project?
It depends on the work and its scope. For small, repetitive tasks, even a pilot of a few weeks shows the direction; for broader projects, you may need to wait a few months because of the team's adjustment period.
Do we have to calculate the return in monetary terms?
It isn't required, but it helps. Metrics such as time saved, fewer missed requests or a lower error rate can be roughly converted into a monetary value; what matters is writing down the assumptions clearly.
We started without a baseline. What can we do?
It is often possible to reconstruct the past period from system records; email timestamps and CRM records are good sources. If not, you can work without AI for a short period and collect comparison data.
The Globya assistant is online 24/7; it answers right away and passes your question to the team if needed.