"Only if necessary" is a good AI plan for a small business

On September 14, the Small Business Expo research team published a survey asking owners whether they would replace an employee with AI. The biggest answer, at 45.2%, was “only if necessary.” Another 28.9% said never. About one in four have done it already or would.

The same survey explains the caution. 58.5% of owners say customers are spending less than last year. 57.9% use personal debt to keep the business running, at least now and then. The survey’s own read is that necessity, not excitement, is what drives adoption.

That middle group gets described as hesitant. We think it is the most sensible group in the room. An owner borrowing against their own credit card can’t afford a bet they can’t check. “Only if necessary” is a person asking for proof first, and that’s the right instinct. It just needs a plan attached.

Start with the work nobody is doing

“Replace an employee” is the hardest possible first move. You lose a person who knows your customers, and you find out whether the AI is any good after the fact.

There is almost always an easier place to start. The chat message at 9pm. The emails that pile up on Saturday. The fifth “where is my order” of the morning while the one person on support is trying to sort out an actual problem. That work is either slow or not getting done, and handing it to an agent doesn’t cost anybody a job.

Time profiled Hospitable, a short-term rental software company, in May. Its AI agents handle about 70% of support questions. The company didn’t lay off its support staff. It hired more slowly instead, and the CEO said they would have needed to triple the support team without the agents. That’s a growing business keeping its people and adding capacity, which is what most owners actually want.

Big companies are landing in a similar place. A Gartner survey of 321 service leaders found 85% are adding new responsibilities to their human agents, and 75% are moving agents into new roles.

What “proof first” should mean

If you’re in the “only if necessary” group, here is a short test to hold any AI support tool to before a customer ever talks to it.

1. See it answer your real messages first, in private. Not a demo script. Your actual inbox, with replies that go nowhere until you’ve read them.
2. Turn it on one subject at a time. Order status and store hours are not the same risk as refunds. You should be able to let the agent answer one and keep the other for yourself.
3. Set limits it can’t argue with. “Never give a refund over $200” should be a rule in the software, checked every time, not a sentence the AI was asked to remember.
4. Count what your person did with the time. If the agent takes the repetitive questions and your employee spends the hours on callbacks, reviews, and the customers most likely to leave, you learned something useful and you kept the person who knows the business.

Run that for a month and “necessary” stops being a guess. You’ll have seen hundreds of drafts on your own customers, and you’ll know which topics you trust and which ones you don’t yet.

How we built for this owner

This is the owner we had in mind when we designed Celeste.

Celeste learns from your own support history, then practices on your real incoming messages in private. You read the drafts. Nothing goes to a customer until you decide it should. Then you turn it on one topic at a time, and you can pull a topic back if you change your mind. The limits you set are checked by our code, so the agent proposes an action and the software decides whether it’s allowed.

It also fits a tight budget. You pay per conversation Celeste resolves, with no per-seat fees, and every workspace starts with a free month on your own data and no card to begin. The details are on the pricing page.

And for the owners who answered “never,” or who worry about what happens to support workers in general, we made a commitment of our own. Under The Dividend Standard, IMCeleste sets aside, every year, the greater of 5% of its qualifying revenue or 70% of its adjusted profit for the workers whose jobs AI automates.

If you’re waiting until AI is necessary, you can still get ready now. Watch it practice on your own messages, and when the day comes you’ll already know what it can handle.