Three Numbers That Defend Your AI Spending

Most founders who adopt AI tools feel the difference before they measure it. Tickets move faster. Blog posts take less time. The inbox clears. Then a partner asks what the AI is actually costing, and the answer is a shrug.

The shrug is not dishonesty. It is a measurement problem.

The counterargument deserves to be heard first

A California Management Review meta-analysis reviewed experiments and adoption studies on AI and productivity, controlled for publication bias, and found no robust relationship between AI adoption and aggregate firm-level productivity. That finding is worth sitting with before you build any measurement system around AI, because it tells you something specific: task-level gains do not automatically add up to a better business.

The MIT Sloan research on GPT-based tools found roughly 40% performance gains when workers stayed inside the tool's capability range, paired with significant drops when they pushed it into tasks the tool handles poorly. A single before-and-after number averages those two outcomes together and tells you nothing about which situation you are in.

So the honest version of what three simple indicators give you is not proof that AI improved your business. It is evidence of what changed at the task level, which is a narrower claim and a more defensible one.

What the research actually confirms at task level

Brynjolfsson, Li, and Raymond studied more than five thousand customer support agents using a generative AI assistant and measured a 14-15% increase in issues resolved per hour. The Federal Reserve Bank of St. Louis surveyed workers and found average time savings of 2.2 hours per week for a 40-hour worker. A systematic analysis of AI chatbot deployments found response time reductions approaching 99.6% alongside customer satisfaction score increases of roughly 18 percentage points.

These are not economy-wide claims. They are counts of observable events at the task level, and they map directly onto three things you track without a spreadsheet degree.

The two-week method

Pick a two-week window before you add the AI tool. Count three things each week: hours spent on content creation and customer replies, number of content pieces finished, and average time from customer message to your first reply.

Run the AI tool for the next two weeks. Count the same three things.

Track both the tasks where the AI worked and the ones where it did not. If your chatbot resolved 80% of queries in seconds and 20% required you to step in and rewrite the response from scratch, record both numbers. The MIT Sloan capability-frontier problem is real, and a measurement that only captures the 80% is not a measurement, it is a highlight reel.

One honest limitation worth naming: the research on whether small business owners track AI failures as diligently as AI wins comes from enterprise-scale studies, not from firms of five or ten people. The discipline required to record the bad outcomes alongside the good ones is harder when you are also running payroll. [Inference: small business owners tracking this method without a formal logging habit are more likely to undercount failure cases than enterprise workers in monitored environments.]

What you bring to a stakeholder conversation

A lender or business partner asking about AI spending is not asking whether AI lifted the national productivity index. They are asking whether the money you spent changed anything they can see.

If your content output went from two pieces per week to five, and your first-reply time dropped from four hours to under one, those are direct answers to a direct question. Brynjolfsson's team measured issues resolved per hour. You measure the same thing at smaller scale. The logic is identical.

The California Management Review finding does not make these numbers meaningless. It makes them honest about their scope. You are not claiming AI transformed your firm. You are showing a stakeholder what changed in the specific tasks you used it for, over a specific two weeks, with the failure cases included.

That is a claim the numbers support. It is also a claim a stakeholder can verify.