AI front offices for small businesses

Three years into large language models, the distribution of who has actually benefited is lopsided. Large companies can put money and headcount against the problem and absorb a few failed pilots on the way. Consumers have more assistants and subscriptions than they can evaluate. Small and mid-sized businesses, which are most businesses, have been served the least.

The technology is not what is missing. The distance between a capable model and something a business owner will put in front of their customers is long, and most of it is not model work.

An owner has to be understood well enough that the system asks the questions their business actually asks. They have to trust it enough to stop listening in. And it has to hold up on the phone line where real revenue arrives, in a business with no engineer to catch it when it fails. Each of those is substantial, and together they are why plenty of capable teams are entering this space and plenty of them stall in it.

That difficulty is the opportunity, and it is why these owners need someone doing this now rather than waiting for the software industry to reach them. Piston is my attempt at it.

What it does

Piston answers a business’s calls at any hour. It works out what the caller needs, sends a genuine emergency straight to a person, books the rest into a live calendar, and texts the confirmation. The owner opens one screen and sees the call, what the caller wanted, and what was done about it.

What was hard

Earning the right to be left alone. Success is not the system working, it is the owner no longer checking. Getting there takes a period where they do listen back, find it handled the call the way they would have, and gradually stop. None of that phase is technical and none of it can be skipped.

Never claiming something that did not happen. A voice agent that tells a caller they are booked for Tuesday when the calendar write failed is worse than one that never booked at all, because now nobody is coming and the customer is expecting someone. The agent can only say a booking is made once the calendar confirms it, and hands off to a person when that fails. It is allowed to be less impressive. It is not allowed to be wrong.

Knowing when to get out of the way. The most valuable call an HVAC business takes is no heat in January, and it must not queue behind somebody pricing a replacement unit. Triage happens before anything else is asked, not after a scripted greeting and not after the caller has described the problem twice.

Fitting a business without rebuilding for it. A locksmith and a law firm want different questions, different urgency rules and a different definition of a good call. They run the same product with different configuration, which is what makes it possible for one person to support five industries at once.

Where it runs

Five business types are built on it: HVAC, locksmith, spa and salon, restaurant, and healthcare-law intake. One HVAC company runs it in production today.

Those figures are the owner’s own attribution, from his books rather than a measurement I made.

That deployment stops short of booking on purpose. It takes the job and promises a callback within the hour, because dispatch there depends on which technician is closest to which job, and a slot booked by software at nine at night can be the wrong answer by seven the next morning. Booking runs live for the businesses where a slot is simply a slot.

What I would fix next

The intake and booking paths grew separately and share less than they should. Every new industry still costs a conversation with the owner about what a good call means to them. That conversation is worth having and I do not want to remove it, but the part of it that repeats between businesses should not be work.