"How do I know if AI is actually saving my business money?" The honest answer is that you know it the way you know anything else about your costs: from your own records, set against what you pay. Not from how the team feels, and not from what the vendor says. This article is for the owner of a company with ten to two hundred people who is paying for AI tools and has a nagging sense that nobody has added it up. It looks at the money side: cost per task, the checking time nobody counts, seats that sit unused, and a short review each quarter that ends in keep, change or cut. If you are still working out what these tools cost to begin with, how much does AI cost for a small business comes first.

I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them. I am a lawyer by training. Now I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies through 013 Labs. This is not financial, tax or accounting advice. How a saving or a cost shows up in your accounts is a question for your own accountant.

The short answer

You are saving money if the cost of a task, counted in full, is lower than it was before the tool, and the tool's own cost is smaller than that difference. Everything below is a way of filling in those two halves honestly, with numbers your company already holds.

Start with the cost of one task

Pick one task that happens often and that someone does the same way each time: answering a standard customer question, drafting a quote, summarizing a call, sorting incoming invoices. One task, not the whole company.

For that task, find the numbers that already exist in your records.

Time per task, before and after. Your timesheets, ticket system, job log or the manager's own notes will say how long the task took before the tool arrived. Compare it with the same measure now. If nobody recorded the before number, say so plainly and start recording today. A guess made afterwards is not a measure.

Turnaround. How long from the request arriving to the work going out? A quote that used to take two days and now takes one has a value, but only if it changed something you can see, such as more quotes sent or fewer lost.

Errors and corrections. How many pieces of work came back wrong, were corrected by a customer, or had to be redone? Your credit notes, complaints, returns and rework logs already hold this. A tool that is faster but produces more corrections may not be saving anything.

Volume. Did the same people handle more, or did the same work simply take less of their day? Both can be real. Only one of them lowers cost unless the freed time goes somewhere useful.

Multiply the time by what that time costs you, and you have a cost per task. Your accountant or bookkeeper can tell you what rate is fair to use. I will not tell you one.

Count the checking time

This is the cost most owners miss. When a tool produces a draft, a summary or a number, a person looks it over before it goes out. That looking takes time, and it belongs in the cost of the task.

Ask the people doing the work, or watch a few cases: how long does it take to check, fix and approve what the tool gives them? If the answer is close to the time the task used to take, the saving is small, and some tasks are slower than before once the checking is counted.

Two things follow. First, checking is not waste. For anything that reaches a customer, a lender or your books, someone should look at it before it counts, and the person should be named. Second, the checking cost should fall as people learn what the tool does well and badly. If it does not, that is useful to know. How to train employees to use AI is where that improvement usually comes from.

Count the seats and add-ons

Now the other half: what you pay. List every AI subscription, seat and add-on, from the invoices and card statements and not from memory. Include the ones a department bought on its own card.

For each one, write down who has it, who actually used it in the last month, and what task it is meant to help with. Most tools can show last-login or usage information in their admin page.

A seat nobody uses is a pure cost with no saving to set against it. So is a second tool that does what the first already does, and an add-on bought for a trial that was never switched off. These are usually the quickest savings you will find, and they need no judgment about the technology at all. Check the renewal dates too. A tool that renews automatically each year, with nobody reviewing it, will keep costing you whether or not it earns its place.

Put the two halves side by side

For each tool, one line on a single page:

You do not need a clever spreadsheet. A page of a dozen lines is enough. If you cannot fill in a line because the records are missing, that is the finding: nobody can say whether that tool is saving anything.

The quarterly review: keep, change or cut

Put it in the calendar, once a quarter, for thirty minutes with whoever owns each tool.

Keep the tool if the cost per task is lower with checking counted, the errors are not rising, and the seats are in use. Write down why, in one line.

Change it if there is a real gain but something is off: too many seats, the wrong people using it, checking that takes too long, or a task that suits it better than the one it is on. Name the change and a date to look again.

Cut it if nobody can show a gain, if usage has died away, or if it duplicates something else. Cancel before the renewal date, not after. Cutting a tool that did not earn its place is a good result of the review, not a failure of it.

The review also gives you a place to ask whether the first tasks you chose were the right ones. What to automate first in a small business covers how to pick tasks where a saving is easiest to see.

What not to count

How people feel. "It feels faster" and "the team loves it" are worth hearing, and they are a reason to look, not a result. Feelings can run ahead of the numbers in either direction.

The vendor's claims. A vendor's figure for savings comes from other companies' work, measured their way. It is not your task, your people or your records. Treat it as a reason to run your own check, never as the answer.

Time "freed up" with no destination. If a person saves an hour and the hour disappears into the day, no cost has fallen. Count a saving only where you can say what happened to the time: more work done, less overtime, fewer temporary hours bought.

One good week. A task that went well once is a story. A few months of the same measure is a record.

Savings you would have had anyway. If a task got faster because of a new process or a new hire at the same time, say so, and do not credit all of it to the tool.

What to tell a lender, a board or a buyer

Each of them is asking a version of the same question: is this real, and does someone own it?

A lender wants to know that AI spend is controlled and that nothing about the tools exposes the company. A page showing what you pay, what each tool is for, who checks the output and what you cut last quarter answers that without any claim of big savings. What to tell your bank about AI sets out what a lender is really asking.

A board wants the same page plus a view: which tools stay, which change and which go, and who is accountable. A board is more reassured by a company that cut something than by one that has kept everything.

A buyer will look for evidence that any gain is repeatable and does not depend on one person's private setup. Your quarterly pages, kept for a few quarters, are that evidence. Does AI make my business worth more when I sell goes into what buyers do and do not credit.

In all three cases, say only what your records show. Do not claim a saving you cannot trace. How any of it should appear in your accounts or any formal statement is for your own accountant, and what to ask your accountant about AI gives you questions to bring.

Picture a wholesaler of seventy people whose owner asks for the invoices behind every AI tool in the company. The list is longer than expected, because two departments each bought a similar tool on their own cards. The owner asks each tool's user how long the task took before, finds the order desk kept a log and the marketing team did not, and asks the desk how long checking takes. At the first quarterly review, one tool is kept with a note, one is changed from fifteen seats to five, one is cut before its renewal, and marketing is asked to start a simple log. This is a hypothesis, a composite and not a client.

Mistakes that repeat

Having no before number. Counting the vendor's saving instead of your own. Ignoring the time people spend checking. Paying for seats nobody uses. Letting tools renew with no review. Counting freed time that went nowhere. Telling a lender or buyer more than the records show.

Frequently asked questions

How do I know if AI is saving my business money?

Compare the full cost of one recurring task before and after, with checking time included, and set the difference against what you pay for the tool, every seat and add-on counted. Use your own records, not feelings or vendor claims.

What records do I need?

Time per task, turnaround, error and correction counts, volume, and the invoices for each tool. Most companies already hold most of these in timesheets, ticket logs, credit notes and card statements.

What if I never wrote down the before number?

Say so, and start recording now. A guess made afterwards is not a measure. Even two or three months of records going forward is better than nothing.

What is the hidden cost of checking AI output?

The time a person spends looking over, fixing and approving what a tool produces. It belongs in the cost of the task. If it is close to the original task time, the saving is small.

How often should I review my AI tools?

Once a quarter is enough for most companies, and thirty minutes per tool owner. The outcome for each tool is keep, change or cut.

Should I cut a tool nobody can show a gain from?

Usually yes, before the renewal date. If someone can name a task and a measure it should improve, give it a set time to show it first.

Is this financial, tax or accounting advice?

No. It is a method for checking your own records. How savings and costs are treated in your accounts is for your own accountant.

Not sure where you stand? Take the 3-minute AI readiness scorecard.