Plenty of companies have already approved their first AI project. Some added an assistant to answer customer questions, others use AI to read supplier invoices. A few months later the owner asks whether the project was worth the money, and nobody can answer with confidence. The cause is almost always the same. Nobody recorded how things worked before the AI went in.
Measuring the ROI of an AI project is not complicated, as long as you prepare for it from the start. This article covers what to measure, how to run a fair pilot, and the mistakes that make ROI figures look better or worse than reality.
Start with a baseline
A baseline is a picture of your process today, before any AI. Without it, every claimed saving is a guess. Take one process you want AI to help with and record the following over a few weeks of normal work:
- How much work comes in, such as invoices, tickets, or customer messages per day.
- How long one item takes from the moment someone picks it up until it is done.
- How often mistakes or revisions happen.
- How long customers or other teams wait for the result.
- Who does the work and how many staff hours it uses.
The record does not have to be perfect. A simple spreadsheet that an admin fills in each day is far more useful than anyone's memory. A clinic that wants AI to answer questions about doctors' schedules, for example, only needs to log the number of chats per day and how long patients wait for a first reply.
How to measure the ROI of an AI project
An honest ROI looks at several sides at once. Looking only at time saved is often misleading.
Time saved
Compare the time per item before and after AI. Include the time spent checking the AI's output, because review time is easy to forget. Then find out what the saved time is actually used for. If an admin can handle more orders without overtime, that is a real benefit. If the time simply goes unused, the saving has not yet turned into value for the business.
Quality of the output
AI that is fast but often wrong creates more work. Measure the error rate the same way you did for the baseline, such as the number of invoices entered incorrectly or customer replies that needed correcting. Build a set of real examples with known correct answers and test the AI against them regularly.
Adoption
An AI feature that nobody uses produces nothing. Track how many staff actually use it each week and on how much of their work. Low adoption is usually a clue. The workflow may be awkward, people may not trust the results, or the team has not been trained.
Cost
The cost of an AI project is more than the API bill. Include development or subscription costs, monthly API usage, the team's time for review and maintenance, and the cost of fixing mistakes. API costs also grow with volume, so work out the cost per item, not only the monthly total.
Once all four sides are recorded, the ROI formula itself is simple. Take the value of the benefits, subtract the total cost, and divide by the total cost. The hard part is putting a rupiah value on benefits such as faster customer replies. Agree on how to value them with the business owner before the pilot starts, so the method is not argued over once results come in.
Setting up a fair pilot
A pilot is a limited trial before AI is rolled out across the company. A fair pilot gives you an answer you can trust, whether the result is good or bad. You can use these steps:
- Choose one process with enough volume and results that are easy to measure.
- Set metrics and targets before the pilot begins, and do not change them halfway through.
- Use real data and real users, including difficult cases and messy documents.
- If possible, keep a comparison group working the old way.
- Run it long enough to cover one full work cycle, including busy days.
- Record every cost, including the hours your team spends preparing and checking.
- Finish with a clear decision to continue, change course, or stop.
Stopping a pilot that did not work is also a valuable result. You get an answer at a small cost, before any larger investment.
Mistakes that distort ROI figures
A few mistakes show up again and again. The first is calculating savings from a demo. Demos use tidy examples, while daily work is full of blurry documents and unexpected questions. The second is comparing periods from different seasons. A distributor comparing the weeks before Lebaran with an ordinary month will get numbers that mean nothing.
The third mistake is assuming saved hours automatically become money. As long as headcount and workload stay the same, the real value lies in what the team can do with the extra time. Finally, many teams ignore long-term costs. Models get updated, prompts need adjusting, and quality has to be monitored after launch. All of that belongs in the cost.
AI projects that are measurable from day one
Through our AI Integration service, Averoa builds AI features into the apps and processes you already run, with evaluation on real data and model cost control from the start. We can help you set up the baseline and the pilot before a project is expanded.
