What does AI cost a mid-sized company, and how do you measure the return?

The three costs of AI for a company of 20 to 500 people, which one is smallest, how to keep each down, and the before-and-after method we use to measure the return.

A mid-sized company pays for AI in three places. There are seats on an AI plan, usage charges for anything you automate, and the people who set it up and keep it running. Usage is usually the smallest line. To measure the return, time each task before you automate it, time it again after, and compare the hours you got back with what you paid that month.

Start with the outcome

Taylor tells owners to start from the outcome they want, and he counts three. An owner wants more money, more time, or a better product. Every dollar of AI spend should trace back to one of those. If nobody at your company is checking that it does, some of that spend is wasted.

The three costs

CostWhat it isHow it growsHow to keep it down
SeatsA team plan for Claude, ChatGPT, or CopilotPer person, per monthPut everyone on the lowest team tier and upgrade the heavy users
UsageWhat an automation pays each time it calls a modelPer taskSend routine work to smaller models
PeopleTraining, setup, builds, and upkeepPer month or per projectPay for results and keep the builds in your own accounts

Seats

Start everyone on the lowest team tier. Move people up once you can see who the heavy users are, and expect your AI champions to be among them. Have finance manage the seats the way it manages any other subscription. One CPG client of ours now does exactly that.

The alternative costs more than it looks. One company we spoke with had a dozen people on individual Claude subscriptions, each with its own history, and no shared workspace between them.

Usage

Usage is what you pay when an automation calls a model, and it's billed per task. Reading a photo costs a few cents. A short text answer costs a fraction of a cent. A home services client worked out what it would cost to have AI read job photos on up to 30 quotes a day in busy season, and by their own estimate it came to a small monthly amount.

The charges go to your own accounts. Everything we build ends up in accounts the client owns, so the bill from the AI provider comes to you and you can see it.

Token costs get out of hand when every job runs on the biggest model. Taylor's rule is to plan with expensive models and execute with cheaper ones. People tend to do the reverse. A data layer that pulls from your systems on a schedule can run on small or open-source models, because the work is routine. We wrote a separate guide on matching models to jobs.

People

Someone has to train staff, connect the systems, build the automations, and fix them when a format changes. You can assign it to your own team, hire for it, or bring in an outside team. Owners sometimes ask why they'd pay outsiders when a senior analyst already knows the company. Our answer is that your analyst can't flag automations they don't know are possible. Their knowledge of the business still makes every build faster.

We don't bill by the hour. Taylor explains it with a report. If AI lets us finish a five-hour report in one hour, there's no reason to bill you for five. Billing by the hour would mean charging you more for working slowly.

Whoever you hire, ask how they stop the scope from growing. We hold a weekly check-in and a monthly scope review with every client. Ask as well who owns the result. Our clients own the code, prompts, and data, with no licence fee back to us.

Your budget sets the pace. A prospect asked Taylor whether they could start at half. We told them they could, and that interviewing their staff would then take two months where it would otherwise take one. With the tools changing this quickly, going slowly has a cost too.

How to measure the return

  1. Record a baseline before you build. We interview each person and write down how long their recurring tasks take today.
  2. Decide where the time goes. It might go to selling, to serving clients, or to work you used to turn down.
  3. Time the same tasks after the automation is live.
  4. Multiply the hours saved each week by what that time costs you, and add any revenue the extra capacity brought in.
  5. Compare that with what you paid that month, and repeat every month.

Before the interviews are done, any return figure is a guess. Once they are, we know where the hours sit, and we set the benchmark with the client and report against it monthly. If the value delivered doesn't beat the monthly cost, the work should stop or change. We apply the same test to our own fee.

A worked example

A skeptical owner brought Taylor a messy weekly file that took her team five hours to pull into shape. They ran it through Claude on the call and it was done in five minutes. At five hours a week, that single task was costing her team 260 hours a year.

It was a live demo, and an automation in daily use needs more care than that. It still shows how to find your first number. Pick a task with a known weekly cost and time it both ways.

Be careful with the numbers, including ours

Usage and return are different things. After a training week at an airport hotel, staff estimated their AI use went from half an hour to five hours per employee per week. That number measures adoption. What the five hours produced is a separate question.

Early figures are usually self-reported. One of our CPG clients reports more than five hours saved per person per week. That came from two of the first three people who answered a survey, and our client story says so.

We'd be wary of any return percentage quoted before someone has looked at your business. We don't publish one for that reason.

Your next step

Pick three recurring tasks this week and time them. Write down who does each one and what an hour of their time costs. That gives you a baseline to hold any AI spend against, including a proposal from us.

If you want help finding the three, book a free audit and we'll go through your team's week with you.

More guides.

Find out what we'd build for you.

The free audit is a 30-minute call. We ask how your team works today and tell you where we'd start.

Book a free audit ↗︎