What an AI assistant for small business should actually do

An AI assistant for a small business earns its keep two ways. It learns the specific way your business runs, down to which parts vendor quietly ships you the reman instead of the new unit. And it watches the paperwork coming at you, then says something when a decision is sitting there unmade. Everything else is a chatbot wearing your logo.

Most of these products fail the first part and overdo the second. The advice would fit any business in your zip code, and the alerts arrive so fast you stop reading them by week three.

Rather than argue that in the abstract, I'm going to run every question against one shop. Kestrel Auto: three bays in Fort Collins, four techs, an owner named Dana who does the books Sunday nights with the TV on.

Generic advice is the default failure mode

Ask a general-purpose AI what to do about slow Tuesdays and you get a listicle. Run a promotion. Post more. Email your list. Start a loyalty program. Dana has known Tuesday afternoons are dead for six years. What she wants settled is narrower: fill that window with $89 oil changes that tie up a lift, or hold it open for the two Front Range Plumbing vans so Thursday morning stays clear for diagnostics?

To answer that you'd have to know bay hours are the constraint and tech hours aren't. You'd have to know the $145 diagnostic fee is the most profitable hour in the building. You'd have to know Front Range is net 30 on paper and day 44 in practice, which Dana has never once made an issue of, because eleven vans that keep coming back beat a stranger who pays on time once.

None of that is in a model's training data. It gets there because someone typed it in, or because the software watched long enough to work it out.

So skip the demo question. Bring the third-weirdest problem you deal with on a regular basis. If the answer would work just as well for the dental office down the street, you have a search engine with better manners.

Read access first, and probably read access only

Almost every one of these tools opens by asking for your email. Gmail usually goes first, since that's where the vendor invoices and the insurance adjuster's third follow-up land.

A tool that can read that inbox and a tool that can send from it are two different products, whatever the pricing page says. The worst a reader can do is misunderstand a thread and hand you advice you ignore. Write access has a longer tail. Four hundred customers emailed about a promotion you never approved. The message you needed for a chargeback dispute, archived. One clumsy auto-reply to the Front Range fleet manager and a $34,000-a-year account starts taking calls from the shop on Riverside.

Ask flatly: can this send, delete, archive, move money, or change a calendar? If the rep has to go check, assume yes. A tool that wants send permissions "so we're ready for what's coming" is asking you to pre-approve a product that doesn't exist yet.

Ask to see the memory page

The pitch for all of these is that they learn your business. Fine. Ask to see what it learned.

Two weeks in, the memory should read like sentences you could hand a new hire. "Front Range Plumbing runs 11 vans, invoices net 30, pays around day 44." "Warranty claims on reman alternators go back to the supplier, not the manufacturer." If what you get instead is a chart of embeddings, or the phrase "learned implicitly," then you can't audit it and you certainly can't correct it.

This is the part I'd lean on hardest in a demo, because it's cheap to fake and expensive to build. Memory also rots. Kestrel dropped a parts vendor in March after three straight backorders, and an assistant still holding the old fact will keep quoting jobs against a supplier Dana stopped calling four months ago. Fixing that should take ten seconds on a page. If it takes a support ticket and a four-day wait, you don't have a memory page, you have a screenshot of one.

Deleting works the same way. A tech leaves badly. A dispute drifts toward small claims. You should be able to pull those facts out on your own say-so, that afternoon.

The good ones send you less

Every monitoring product dies the same death. Week one you read every alert. Week nine they route to a folder you never open, and the one that mattered is sitting in there with the rest.

The fix is structural. Standing work should arrive in a few known shapes, and you should know which shape you're getting.

A daily briefing reports no matter what, quiet days included, because the value is in the predictability. Dana reads hers at 7:10 in the morning while the bay doors go up, and "nothing needs you today" counts as a result. A watch job is a different animal. It has a threshold, and silence means nothing crossed it. That silence is only worth anything if you trust where the bar sits, so ask how the bar gets set and whether you can move it yourself.

Then there's the periodic sweep for things you could be doing better. Most tools lose the plot here. Generating suggestions is cheap; ranking them is not, and a sweep that comes back with fourteen ideas has quietly handed the sorting job to you. A hard cap of three decisions a week forces the software to make a choice it would rather avoid.

So ask a vendor what its product does on a day when nothing is wrong. If the answer is a digest of insights, you'll have it muted by month two.

What screen access buys, and what it costs

A lot of what runs a business never touches an inbox. It happens while Dana is deep in a supplier's warranty portal at 4:40 on a Friday, deciding whether to eat $310 on a rebuilt alternator. An assistant that can see on-screen context catches decisions that email never hears about.

It's also the part that should make you nervous, and I don't think vendors are honest enough about that. Your screen has your payroll login on it. Somebody's FMLA paperwork. A customer's card number, read out loud over the counter while you typed.

What makes it workable is the order of operations. Personal details get stripped before a note is written down, not after, and the off switch sits one click away rather than three menus deep, so you can reach it before you open payroll. Ask when the stripping happens. "Before storage" is an answer. "After storage" means it was stored.

Questions worth asking before you pay

• Can it write, send, or spend anything? If so, what can I turn off?
• Show me the memory page. The page itself, not a description of one.
• What does it do on a day when nothing is wrong?
• How do I delete one specific thing it learned, and how long does that take?
• What would it know about my shop that a competitor's assistant wouldn't?

A vendor who gets vague on the second one is telling you something. And if the honest answer to the last one is nothing, you're paying for advice every owner on the block already gets free.

Scaffle is an AI advisor for businesses. It learns how yours runs and keeps watch through the tools you connect. Connectors are read-only, Gmail first. What it stores are notes in plain language on a memory page you can read, correct, or wipe. Standing jobs come in the shapes described above, and the recommendation sweep is capped at three decisions. A desktop companion can pick up on-screen context, and it strips personal details before anything is kept. We're early, and the checklist above is the one I'd want an owner to hold us to. scaffle.ai