"AI saves money" is a sentence nobody pays for. To defend a budget in front of an owner — or in front of yourself — you need numbers that survive the question "where does that come from?" Here is how to get them, and which of them you should show nobody.
The base formula: ROI = (Benefit − Cost) ÷ Cost × 100%
The formula is trivial. All the difficulty lies in what you are entitled to put in the numerator.
What counts as cost
All three categories have to be counted, otherwise you are not calculating ROI but flattering yourself.
- One-off: the build, data preparation, training the team.
- Monthly: model API usage, infrastructure, support.
- Indirect: your own people's time — preparing the base, attending calls, updating data every week.
The third category is the one most often forgotten, and it is not zero. If someone spends an hour a week keeping the knowledge base current, that is a week and a half of work per year.
A first-year example: a pilot $1,500, plus post-launch refinement, plus API costs in the range of $10–30 a month, plus roughly 50 hours of internal time. Exact figures depend on the process; the budget breakdown is in how much AI automation costs.
What counts as benefit
1. Time saved
Hours saved × cost per hour × 12 months
An illustrative model. The agent closes 70% of routine enquiries. A support operator on $800 a month used to spend all of their time on them. On paper the saving is $800 × 0.7 × 12 = $6,720 a year.
The caveat without which that number is untrue: you only realise this money if the freed-up time turned into something. If the operator stayed on the payroll and simply became less busy, you saved hours but not dollars. Real savings appear when you avoid hiring a second operator as volume grows, or when the person moves to work that generates revenue.
2. Revenue growth
Additional deals × average order value
The effect here is usually larger than in cost savings — and simultaneously harder to prove. First response time affects conversion; that is a durable pattern, not a marketing claim. But by how much in your business, you will only learn by measuring.
How to measure it honestly: record your enquiry-to-meeting conversion for two or three months before launch. Compare with the same figure two months after. Multiply the difference by your average order value — but only the part that cannot be explained by seasonality or a new ad campaign.
3. Fewer costly errors
Cost of the average error × errors prevented
This works where errors are visible and expensive: manual entry into an accounting system, document reconciliation, price calculation. The condition for using it is that you knew the cost of an error before automating. If you did not, "prevented" cannot be counted, and it is more honest to leave this item out.
4. Faster processes
Not everything converts to money directly, but it can be measured in speed and is still a result:
- enquiry handling time: hours → seconds;
- preparing a recurring report: hours → minutes;
- screening a hundred CVs: days → minutes.
Do not put these into the ROI figure — show them separately. Mixing measurable money and measurable time into a single percentage produces a number that collapses at the first question.
Five metrics to monitor monthly
| Metric | What it shows | Reference point |
|---|---|---|
| Automation rate | Enquiries closed without a human | 70–85% on a mature process |
| Average first response time | Before and after deployment | An order of magnitude or better |
| Volume handled | How much work actually flows through the agent | Growing month over month |
| Cost per enquiry | API spend ÷ number of enquiries | Cents, not dollars |
| Answer quality | Post-conversation rating or a manual sample | Not declining month over month |
Reference points are exactly that, not promises. An automation rate of 70–85% is achievable on a process with a solid knowledge base and after several weeks of refinement; in the first month it is almost always lower.
A monthly report template
One page, eight lines, ten minutes to fill in:
- enquiries handled: [number]
- of those, automatically: [number] ([%])
- escalations to a human: [number] ([%])
- average first response time: [seconds]
- hours saved: [number]
- what those hours turned into: [text]
- API spend: [$]
- what we fixed this month: [text]
The sixth line is the most important and the most uncomfortable. It is what separates a report from a presentation.
Four traps
No baseline. The most common and the most painful. If you never measured how long the process took before automation, the effect cannot be calculated and you are left arguing about impressions. Measuring takes a week and has to happen before the work starts.
Attributing everything to the agent. In the same quarter you also launched ads, hired a manager and raised prices. Revenue growth is not entirely the agent's doing. Where the influences cannot be separated, it is more honest to rely on metrics with a direct link: response time, automation rate.
Hours that went nowhere. Described above and worth repeating: saved time becomes money only through a decision — not hiring, reassigning, or taking on more clients with the same headcount.
Measuring too early. In month one the agent performs worse than it will in month three: the base grows, the instructions get sharper. Calculating ROI from the first month understates the result. Calculating it from your one good month overstates it. Take three consecutive months.
Frequently asked questions
How long until it pays for itself?
It depends on what the process cost before automation. Divide the build cost by the monthly saving and you get the payback period in months. If it comes to more than a year, the process was badly chosen — look for one that hurts more.
What if the ROI is negative?
That is a legitimate pilot outcome, not a disaster. There are usually three causes: a process with too little volume, a weak knowledge base, or people continuing to work the old way around the agent. All three are diagnosable in a week and cost less than an annual contract signed blind.
How do we count a process that is not about money directly?
Through the cost of the time of the people involved. An internal knowledge base generates no revenue but frees up managers' hours — and a manager's hour has a price.
Do we need a dedicated dashboard?
Not for the first process. Eight rows in a spreadsheet are enough. A dashboard starts to make sense when you have three or four processes and want to compare them.
Who should own this report?
Whoever owns the process, not the contractor. The contractor has an interest in attractive numbers — that is not malice but a conflict of interest, and it is better removed structurally.
Where to start
Before launching any agent, measure two things: how many hours a week the process consumes, and what your current conversion or response time is. It takes a week and determines whether you will be able to prove anything at all three months from now.
Not sure what to measure in your case? We work through it in a free process assessment, along with an estimate of what your process costs today.