There's a question that separates people who use AI from people who rely on it, and it isn't technical.
Would you let it send that, right now, to your best client, under your name, without reading it first?
Almost everyone hesitates. That hesitation is worth taking seriously, because it's the actual adoption barrier — not accuracy, not price, not integration. Everything else is downstream of it.
What you're really worried about
When people say "I don't trust it," they mean one of four distinct things, and it's worth knowing which.
"It'll be wrong about a fact." Quotes a price that changed last month. Confirms availability that isn't there. This is the most-cited fear and, oddly, the easiest to engineer around — facts come from your systems, not the model's memory, and a claim without a source can be blocked before it leaves.
"It'll sound like a robot." The client reads it and knows. This one's underrated, because the damage is subtle and permanent: you don't lose the client, you become slightly less human to them. Worth noting the bar here is your actual messages — the rushed one-liner you sent from a car park, not the polished thing you'd write with an hour spare.
"It'll say something it shouldn't." Agree to a discount. Commit to a deadline. Make a representation you'd be held to. This is the serious one, because it's the one with a legal shape.
"I won't know what it said." The quiet fear. You walk into a meeting and the client references a conversation you weren't part of, that happened under your name. Not being wrong — being uninformed about yourself.
They need different answers. Most AI pitches address the first and ignore the fourth, which is why the pitch doesn't land.
The threshold isn't accuracy
The intuitive model is that you'll trust it once it's right often enough. Pick a number — 95%? 99%? — and let it go when you hit it.
That's not how it works, and the reason is that the errors aren't evenly distributed in cost. Ninety-nine harmless replies and one message promising a delivery date you can't hit is not 99% good. It's one incident, and the ninety-nine don't offset it.
So the real threshold isn't how often is it right. It's what is the worst thing it can do, and would I survive it.
Which is a much more answerable question, and it has an engineering answer: constrain what the agent is able to do, so that the worst case is something you'd survive. An agent with no authority to offer discounts cannot offer a discount — not because it was told not to, but because that isn't among the actions available to it. An agent that can only propose times your calendar actually has cannot double-book you.
Trust follows capability limits, not accuracy percentages. That's why "how accurate is it" is the wrong first question and "what exactly can it do" is the right one.
The answer to the fourth fear
The "I won't know what it said" worry deserves its own treatment, because it's the one that persists after the others are solved.
The fix isn't a report. Nobody reads a weekly summary of their own conversations. The fix is that the agent works in your channel — the same WhatsApp thread, the same inbox, the same client record — so its messages appear in the history you already look at, in the place you'd naturally look. You're not being briefed on a parallel conversation. There is only one conversation, and some of it was drafted for you.
That single design decision resolves more anxiety than any amount of accuracy improvement.
A practical way through
You don't have to answer "would you let it email your client" as a yes or no. Nobody should.
Answer it per message type instead:
- Availability and scheduling. Low stakes, high frequency, verifiable against your calendar. Most people are comfortable here within a fortnight.
- Answers to routine questions. Fine when every fact is sourced from your own systems and shown as such.
- Follow-ups and nudges. Comfortable quickly, because the failure mode is a slightly redundant message.
- Anything with a number, a date, or a promise in it. Draft only. Possibly forever, and that's a legitimate permanent answer, not a failure to progress.
- Complaints, disputes, bad news. Human writes it. The agent's job is to make sure a human knows within minutes.
Notice that the first three cover most of the volume and none of the risk. That's the whole trade — you're not handing over your client relationships, you're handing over the logistics that surround them.
Where the line actually sits
Here's the reframe that tends to end the conversation.
You already delegate this. If you have staff, someone junior has emailed your client under a shared inbox, using a template you approved, following rules you set. You didn't read every message before it went. You set boundaries, checked the important ones, and stayed reachable when it mattered.
You trusted a process, not a person's infallibility.
An agent is the same arrangement with better instrumentation. You can see every message. You can cap what it's allowed to do. You can stop it in one action. You can read the full reasoning behind any decision it made — which, in fairness, you cannot do with a person.
The question was never whether it's perfect. Nothing in your business is perfect. The question is whether the boundaries are real and whether you'd know quickly if something went wrong.
If both are yes, the hesitation has done its job. You can let it send.
Wednesday: how we write an eval set for a workflow nobody documented.