The overage bill is the good news

More than once this year, I've watched a client get exactly what it wanted from an AI rollout: people started using it. Then the overage bill arrived.

The executives look at the new expense and ask the obvious question: what are we getting for this?

I tell them the truth: the overage is good news. It means adoption worked. It does not mean the investment has paid off.

Those conversations are a preview of 2027.

2025 was "what is AI?" Workshops, demos, and a lot of tire-kicking. I ran more training sessions that year than I can count.

2026 has been "get AI deployed." Pick an LLM. Get security comfortable. Buy the licenses. Train employees. Start the first projects.

2027 will be "what is AI doing for me?"

That question is going to arrive with an invoice stapled to it.

Companies are spending real money on AI now. Mid-market firms are into five and six figures a year. Large enterprises are spending seven. At some point, a CFO is going to look at that line item and ask the question CFOs are paid to ask: what did we get for this?

When a CEO asks me how their AI rollout has affected revenue or costs, I've started answering with a question: what financial return has your company gotten from PowerPoint?

Nobody can point to a number. Nobody would take PowerPoint away either.

That's where general-purpose AI for knowledge workers is heading. It's becoming table stakes, the way Excel is table stakes. It shows up as an expense, you can't cleanly attribute revenue to it, and removing it would make the company worse.

The productivity gains underneath are real. They're also hard to book.

Say John and Sally in finance become 25% more productive because they've automated the grind out of their weeks. Next summer, when the workload grows, the department absorbs it instead of hiring another person. That's a real six-figure benefit, fully loaded.

But can you prove AI was the reason? Maybe the workload changed. Maybe the manager reorganized the team. Maybe John and Sally simply got better at their jobs.

There's no invoice for the person you didn't hire. The value shows up as an absence, and absences don't make it into board decks.

There's another catch. When John gets two hours of his day back, those hours don't automatically become value. Left alone, they become more email. A saved hour is capacity. The company only benefits when that capacity gets put to better use.

This is why companies need two scoreboards for AI: capacity value and project value.

Trying to force both into one ROI calculation is going to make a lot of good AI work look bad.

Capacity value is what happens when hundreds or thousands of employees get through their existing work faster.

Someone in HR drafts a policy in twenty minutes instead of two hours. A manager works through a spreadsheet without waiting for an analyst. A salesperson researches an account before a meeting instead of going in cold.

Multiply those gains across a company and the capacity is enormous. The financial return appears slowly, though, and usually gets credited to something else.

Instead of asking employees how many hours they saved, start tracking what the company expected to need before AI changed the work:

Headcount plans. What roles did each function expect to add over the next 12 to 18 months, and at what cost?

Team scope. What is each department responsible for today? If the same team handles materially more work six months from now, that's a capacity gain worth investigating.

Operating metrics. Use numbers the department already cares about: invoices handled, requests resolved, reports completed, customers supported, or deals processed.

None of these will give you a perfect AI ROI number. They will give you better evidence than an adoption dashboard and a survey asking whether employees saved time.

Project value is different. It comes from defined work with a business outcome attached.

Every mid-market company has a gnarly process that eats half the week for three people. Automate it and you can name the hours, error rate, cycle time, and cost before and after. Those are numbers you can defend in a budget meeting.

AI also lets companies build things they couldn't have built before.

When I analyze a client's AI usage, one or two of the heaviest users often have jobs that have nothing to do with software. Some are generating hundreds of thousands of lines of code a month with Codex or Claude Code.

Whether all that code becomes valuable IP, a process advantage, or a maintenance problem is a fair question. It's probably some of each. But that capability did not exist inside most mid-market companies two years ago. Point it at the right product or workflow and the revenue or cost impact can be measured.

These projects bring their own homework. What a nontechnical employee builds often lives on a laptop, with only that employee knowing how it works. At some point, somebody has to move it to managed infrastructure, assign an owner, and pay to maintain it.

A project that saves $200,000 a year and costs $80,000 to operate is a good project. It isn't a $200,000 return.

Every AI project should have four things before the work starts: a baseline, a business owner, an estimated operating cost, and a date when someone will check the result.

If there's going to be a trough of disillusionment with AI, I think it shows up in 2027.

The technology will keep getting better. The problem is that the bills are arriving faster than the proof. Finance can read the cost to the penny while the value is spread across hundreds of people and thousands of small tasks.

The precise number usually wins the argument.

The companies that look smart next year will be the ones that started keeping score this year. They'll be able to say, "Our teams absorbed this much additional work without these planned hires, and these four projects produced this much revenue or reduced these costs."

Everyone else will have an adoption dashboard and a very large invoice.

Homework: Write down the headcount you currently expect to add over the next 18 months, by function and cost. Then list every active AI project and the business number it's supposed to change. If a project has no number, add one now. Twelve months from now, those two lists may be the best answer you have when the CFO asks, "What did we get for this?"

Two things I'd genuinely like back from you. First, hit Reply and tell me which part of this matches what you're seeing. Second, if there's someone in your world who should be on this list, forward it to them.

- Robbie