Product lists age in weeks; this guide is a decision process instead. It exists because the most common AI purchase in small business goes like this: a tool looks impressive in a demo, gets bought on an annual plan, gets used enthusiastically for two weeks, and is quietly abandoned by month two — while the subscription runs on. The failure is almost never the model. It’s the fit between the tool and a workflow nobody examined first.
The process below has four steps: name the workflow, disqualify fast, pilot small, and decide on numbers.
Step one: start from a workflow, not a tool
An AI tool only pays for itself inside a workflow — a recurring chunk of work with a start, an end, and a cost. “We should be using AI” is not a workflow. “Every inbound quote request takes 40 minutes to turn into a written quote, and we do twelve a week” is.
Make the shortlist of candidate workflows yourself, before looking at any product:
- Recurring, not occasional — weekly or daily, or the savings can’t compound.
- Text- or document-shaped — drafting, summarizing, classifying, extracting, replying. That’s where current models are strong.
- Tolerant of review — a human can check the output faster than they could produce it. Drafting a quote qualifies; sending unreviewed legal commitments does not.
- Genuinely costly — count the hours; multiply by what an hour costs you. If the workflow costs less than the tool, stop there.
Rank by cost, pick the top one, and only then go shopping. Shopping first reverses the logic: the demo defines the problem, and demos are built around the workflows the vendor looks best at.
Step two: the three disqualifying questions
Most vendors can be eliminated in ten minutes with three questions. Ask them before any trial, in writing where the answers matter.
1. What happens to our data?
You’re going to paste customer details, financials, or contracts into this thing. So: is your data used to train anyone’s models? Can you delete it? Where is it stored, and does that satisfy whatever rules you operate under? A vendor that answers vaguely — or buries “we may use your content to improve our services” in the terms — is disqualified for any workflow touching customer or financial data. Reputable vendors answer this question crisply because they’re asked daily.
2. What happens when it’s wrong?
Every AI tool produces wrong output sometimes. The honest vendors design for it: confidence flags, source links, human-review queues, easy correction. The dangerous ones design around it — demos only, “99% accuracy” with no definition of the measurement, no visible path from wrong answer to fixed answer. Ask for the failure story: “show me what the product does when it makes a mistake.” A vendor with no answer has never thought about it, and you’ll be the one finding out.
3. What does leaving look like?
Can you export your data — prompts, outputs, configurations, history — in a usable format? Month-to-month pricing, or annual lock-in? If the tool disappeared on a Friday, what would Monday look like? This question matters double in AI, where products are young, pivots are constant, and acquisitions kill tools monthly. Prefer tools that wrap portable assets (your documents, your templates, standard formats) over tools that accumulate unexportable state.
A useful pattern emerges after a few of these conversations: the questions don’t just filter products, they reveal vendor maturity. Teams that answer all three without flinching tend to build trustworthy products. Teams that handle them badly tend to handle everything badly.
Step three: pilot small, with a number
Never start with an annual plan, a team rollout, or a migration. Start with one workflow, one or two people, one month, and a measurement designed before the pilot starts:
| decide before the pilot | example |
|---|---|
| The baseline | ”A quote currently takes 40 minutes” |
| The metric | ”Minutes per quote, and error rate caught in review” |
| The bar | ”Under 15 minutes with no rise in errors, or we stop” |
| The reviewer | ”Sara checks every output in week one, spot-checks after” |
| The end date | ”Decision on the 30th, in the calendar now” |
The end date is the step everyone skips. Without it, mediocre tools survive on inertia — nobody cancels, because nobody decided to decide. With it, the pilot produces the only thing a pilot is for: a clear keep-or-kill call.
During the pilot, log the failures, not just the wins. Five minutes saved drafting doesn’t help if review now takes ten because trust is low. The honest metric is end-to-end time including the checking — which is also why workflows “tolerant of review” were a step-one filter.
Step four: decide on numbers, then expand slowly
If the pilot clears the bar, expand one workflow at a time, re-running the same cheap measurement loop. If it misses, kill it without ceremony — a failed pilot that cost one month and one seat is the process working, not failing. Either way, write down what happened; after three or four pilots the notes become your own product list, calibrated to your business instead of to a vendor’s demo reel.
Two structural cautions for the expansion phase:
- Mind tool sprawl. Five AI subscriptions doing five small jobs usually lose to one capable general assistant doing all five at 80% quality. Sprawl costs money, logins, and training time. (The next guide, on the solo operator’s stack, takes this head-on.)
- Re-evaluate yearly, not monthly. The underlying models improve fast, and a category that failed a pilot in January may genuinely work by the following year. But chasing every release mid-year burns time for marginal gains. A standing yearly review captures most of the upside at a fraction of the churn.
The free-tier-first corollary
Before piloting any paid tool, spend thirty minutes establishing the baseline it has to beat: the same workflow done with the general assistant you (probably) already pay for, plus a saved prompt. A surprising share of the “AI for X” category is a prompt template with a logo — and if the assistant-plus-prompt gets you 90% of the demo, the specialized tool’s real pitch is that last 10%, priced accordingly. Sometimes that 10% is worth it (real integrations, real volume handling); often it’s a subscription for something you could have saved as text. You only find out by testing the cheap version first, and the test doubles as pilot preparation: it forces you to articulate the workflow precisely enough to prompt it.
What to ignore
You can safely ignore: anything sold primarily through urgency (“your competitors are already using AI”), anything that can’t name its failure modes, anything demanding annual payment before a pilot, and any tool whose pitch is a capability (“powered by GPT-5!”) rather than a workflow outcome. The model inside matters far less than the fit around it — every vendor has access to roughly the same models, and the differentiator is everything else.
None of this is a guarantee of results — no process is. What the process guarantees is cheaper, faster, honest failures: you find out for one seat-month what most businesses find out for twelve. In a category moving this fast, that’s the durable advantage.