We turn down AI projects fairly often. Not because the client is wrong about wanting one, but because the thing they want can’t work yet, and building it anyway would burn their budget and their patience.
This piece is the checklist we run through before saying yes. It’s deliberately unflattering. If you fail three of the five, the correct move is to fix the plumbing and come back in a quarter.
1. Your data exists somewhere a computer can read
This is the one that kills most projects. An AI assistant that answers questions about your business needs your business to be written down somewhere other than someone’s head, and somewhere other than a shared drive of scanned PDFs where the scans are photographs of printouts.
“Do we have data” is the wrong question, because everyone has data. What matters is whether the thing you want the model to know is in a system with an API, a database, or at minimum a folder of text-bearing documents.
A good sign: your job history is in a CRM and you can export it. A bad sign: your pricing logic is a spreadsheet that only Dave understands, and Dave is retiring in August. In the second case, the AI project is actually a documentation project wearing a costume.
2. Someone can say what “correct” means
You can’t evaluate a model output without a definition of a good output. It sounds obvious, and it gets skipped constantly.
Before we build a quote-drafting assistant, we ask for twenty real quotes and a person willing to say which ones were good and why. If nobody can articulate the difference between a good quote and a bad one, we have no way to test the thing we build, no way to know when it regresses, and no way to defend it when it produces something odd.
It’s also the cheapest possible filter. Collecting twenty examples takes an afternoon. Discovering after eight weeks of development that the team disagrees about what a good output looks like takes eight weeks.
3. The task has a tolerable failure mode
Models are wrong sometimes. Not often, but not never, and the wrongness is confident. So the real question is: what happens when this is wrong?
Drafting an internal summary that a human reads before acting is fine, because being wrong costs someone ten seconds of confusion. Auto-sending a price to a customer with no review is not fine, because being wrong costs you a contract you have to honor or a reputation hit you have to eat.
The useful frame isn’t accuracy percentage. It’s blast radius. We’re happy to ship something that’s right 85 percent of the time if a human sees every output before it leaves the building, and thoroughly unhappy to ship something right 99 percent of the time if the 1 percent silently wires money.
Most good small-business AI lives in the draft-and-review slot. We don’t apologize for that. It’s where the value is, and pretending otherwise is how people end up with expensive incidents.
4. There is a person who owns it after we leave
Every deployed system decays. Prompts drift out of date because your pricing changed. An API deprecates. A model version gets retired and the replacement behaves slightly differently. Somebody has to notice.
At a 15-person company, that person is real and named, and they need maybe two hours a month. If nobody can be named, the system will quietly rot, and in nine months someone will say “we tried AI, it didn’t work.” They’ll be right, and the reason will have nothing to do with AI.
We’d rather scope smaller and hand over something one person can hold than build something impressive that needs a team you don’t have.
5. You have a problem expensive enough to be worth it
This one gets skipped because it feels rude to ask. But a custom AI build, a real one with evaluation, integration, and support, starts in the low five figures. If the problem it solves costs you $4,000 a year, don’t build it. Buy something off the shelf or leave it alone.
The problems worth a custom build tend to look like one of these:
- A task consuming more than 10 hours a week across the company, every week, forever.
- A bottleneck where the delay itself costs revenue: quotes going out in three days instead of three hours.
- Institutional knowledge trapped in one or two people who are leaving, retiring, or already overloaded.
- A compliance or documentation burden where errors carry real penalties.
- Volume you can’t hire your way out of at a price that works.
Notice that none of those are “we should be using AI.” If the driver is that a competitor mentioned it on LinkedIn, the honest answer is to wait.
Owners hate hearing this, and they’re usually relieved a month later.
What to do if you failed some of these
Failing is normal. Most companies we assess fail two or three the first time. Here is the order we would fix them in.
- Fix data location first. Get the thing you care about into one system with an API. This is usually a CRM consolidation or a document migration, and it has standalone value even if the AI project never happens.
- Then write down the rules. The pricing logic, the qualification criteria, the escalation thresholds. Documentation is a prerequisite, not a deliverable.
- Then automate the deterministic parts. Most of what people want from AI is actually a routing or data-movement problem. Do that with plain automation, no model involved. It’s cheaper, faster, and it never hallucinates.
- Then, and only then, look at what’s left. Whatever survives that filter genuinely needs judgment, language, or pattern recognition. That’s the AI project, and it’ll be smaller and better-defined than what you started with.
Companies that do steps one through three often find the remaining AI scope is a quarter of what they imagined. Take the win. It means you spent money on the part that needed it.
The version of this that is actually optimistic
None of the above says AI is useless for small business. It says the useful applications are narrower and less glamorous than the pitch decks suggest, and that the preconditions are mostly boring operational hygiene.
The companies getting real value right now aren’t doing anything exotic. They have their data in one place, they picked one painful task, they put a human in the review loop, and they named an owner. All of it is available to a 20-person plumbing company today, and none of it requires anyone to have an opinion about which model is winning.
If you want to know where you actually stand, run the five questions honestly with your leadership team and write down the failures. That document is worth more than any vendor demo you’ll sit through this year.