The 90-Day AI Test That Costs Nothing to Run

Most founders who try AI do it the same way they try a new productivity app. They open it, poke around for a week, feel vaguely optimistic, and then forget about it when something urgent comes up. Six months later, they're not sure whether it helped.

That's not a technology problem. It's a measurement problem.

The barrier isn't the price tag

OECD surveys and national small business studies show that AI adoption among microbusinesses accelerated sharply after late 2022, driven mostly by generative AI embedded in tools firms already paid for, not by new purchases. The adoption curve is real. The value detection is not. Firms report using AI, but use stays shallow and fragmented because no one defined what "working" would look like before they started.

The GitHub Copilot research makes this concrete. Studies on developer teams show sizable efficiency gains when AI gets integrated into workflows with measurable outputs. The same research also shows that perceived productivity gains don't reliably appear in standard activity metrics. A developer who feels faster isn't always faster by the measures their team tracks. A founder who feels like AI is helping their content workflow isn't necessarily producing better content, more of it, or faster. The feeling and the data diverge because the measurement design was never set up to catch the signal.

This is the part most zero-budget AI guides skip. They tell you to start small. They don't tell you that starting small without a pre-defined output target produces 90 days of noise.

What the steelman looks like

The strongest objection to self-directed AI pilots is not that they're too simple. It's that founders lack the measurement skills to run them honestly.

The barriers research across multiple studies identifies skills gaps as a persistent obstacle in firms with fewer than ten employees. Designing a valid before/after comparison is itself a skill. A founder who tracks "number of blog posts published" will miss efficiency gains that show up in time-per-draft. A founder who tracks time-per-draft will miss quality improvements that only appear in reader engagement. The wrong metric produces a clean number that means nothing. Ninety days later, the founder concludes AI didn't help, when the AI worked fine and the measurement failed.

A consultant, at minimum, brings an external measurement framework. That's the legitimate part of the consultant pitch. The problem is that the framework a consultant brings wasn't built for a two-person content operation running on tools the founder already owns. OECD surveys and national small business studies consistently show that small-firm AI adoption runs on low-cost generative AI embedded in existing software, not on consultant-led implementations. The measurement frameworks consultants sell were calibrated for enterprise rollouts. They describe what to track in a 200-person marketing department, not in a founder's Tuesday morning writing session. [Inference: the mismatch between enterprise measurement frameworks and microbusiness workflows is not directly stated in the research, but follows from the documented divergence between how small firms actually adopt AI and how consultant-led implementations are typically structured.]

A vendor demo has the opposite problem. It shows best-case performance in a controlled environment against the vendor's chosen output metric. It produces no data about what happens when you, specifically, use this tool on your specific content workflow with your specific team skills. The 90-day pilot produces exactly that data. Nothing else does.

The design that makes the data usable

Single function. Existing tools. Pre-defined output targets set before week one.

The systematic review evidence and OECD survey pattern both point to the same conclusion: single-function, existing-tool experimentation aligns with how small firms already approach AI and offers the clearest path to detecting value. The GitHub Copilot research adds the critical condition: gains become detectable only when AI is integrated into concrete workflows with measurable outputs. Broad rollouts in sub-10-employee firms run into skills gaps and unclear business cases that a narrow pilot sidesteps entirely.

Pick one function. Content creation is a reasonable starting point because outputs are countable and quality is assessable without specialist knowledge. Pick two or three output targets before you start: time from brief to first draft, word count requiring revision, and number of revision cycles before publish. Write them down. These are your baseline for weeks one and two, before AI touches anything.

Weeks three through twelve, you run the same workflow with AI assistance embedded at one specific step — first draft generation, headline options, or research summarization, not all three at once. You track the same metrics. At week twelve, you compare the two sets of numbers.

This is not a controlled experiment. You are one person, there is no control group, and your skills will have changed over 90 days regardless of AI involvement. The data you produce is not publishable research. What it is, is specific to your workflow in a way no consultant pitch or vendor demo will ever be. It tells you whether this tool, used this way, on this task, changed these outputs in your operation. That is the only question worth answering before you spend money scaling.

What the 12 weeks actually look like

Weeks one and two: baseline only. No AI. Track your current output metrics obsessively. This is the step most people skip and the reason most pilots produce nothing useful.

Weeks three through six: introduce AI at one step. One step. Not the whole workflow. Track the same metrics.

Weeks seven and eight: compare the numbers. If time-per-draft dropped and revision cycles held steady, you have a signal. If time-per-draft dropped but revision cycles doubled, you traded one cost for another. If nothing moved, either the tool isn't helping or your metric isn't catching it — and you need to decide which before continuing.

Weeks nine through twelve: adjust the integration point based on what weeks seven and eight showed, then track again.

At week twelve, you have four data points: baseline output quality, baseline time cost, AI-assisted output quality, AI-assisted time cost. If two of those moved in the direction you wanted, the pilot paid off. If none moved, you spent 90 days learning that this tool doesn't fit this workflow, which is worth knowing before you pay for an upgrade or hire someone to implement it at scale.

The research is clear that scaling before clear benefit appears is where small firms go wrong. Ninety days of self-tracked data from a single function won't tell you everything. It will tell you the one thing a vendor demo never will: whether the tool works in your actual operation, on your actual tasks, with your actual team.