What to automate, and what to keep human

Two hosts talk through the one distinction that decides whether an AI project inside a small service business works: repetitive work versus consequential work. They walk two of the four workflows that leak money in an owner-led business — follow-up and approvals — in depth, rather than skimming all four, and land on why the scoping test matters more than the technology choice.

A: Here's the scene. It's Tuesday, 11am. The quote is written. It doesn't go out until 4. Nobody did anything wrong — the person who sends it was on site, then in a meeting, then had ninety unread emails. That's the whole story, and it's the one we keep coming back to in this season.

B: And the instinct, the very first instinct, is to fix that with AI. Let the system send it.

A: Right, and that's exactly the wrong first move, and we'll get to why. But let's start with why this keeps happening at all, because it's not really a technology gap.

B: It's a handoff gap. Most of what we're going to talk about this season isn't "the business doesn't have AI." It's "the business has a step where work sits, waiting on one person, and nobody ever defined what happens next."

A: Which is a very unglamorous diagnosis. Nobody wants to hear "your problem is a missing handoff." They want to hear "add an AI agent."

B: And most of the advice aimed at small businesses is written for companies that don't look like this one. It assumes a data team, a backlog of clean records, someone whose whole job is owning the rollout. In a twelve-person service business, the person evaluating AI is also the person quoting the job, chasing the invoice, and covering for whoever's on leave.

A: That's not a smaller version of the enterprise problem. It's a different problem.

B: So what do people actually try, when they hit this wall?

A: Usually the demo-shaped fix. Somebody sees an agent reply to a customer on its own, and the time saving looks enormous. And it is real — that's the thing, the time saving is genuinely real.

B: But then there's the day you spend apologizing for what it said.

A: Exactly. And that's the distinction the whole season sits on. There's work that's repetitive — reading, sorting, summarizing, retyping the same info into a second system — where nobody's judgment is actually being used. If a machine does that badly, you notice immediately and nothing is lost.

B: And then there's consequential work. Sending the actual reply to the customer. Issuing the quote. Approving a discount. Telling someone their application didn't go through.

A: Judgment is the entire point of that second category. If a machine gets it wrong, you find out from the customer.

B: So — automate the first kind freely. Keep a person in the second kind, always.

A: Which sounds obvious when you say it plainly. It stops sounding obvious the moment a slick demo blurs the line.

B: Let's ground this in one of the four leaks specifically, because "keep a human in consequential work" is easy to agree with and hard to actually apply. Follow-up is a good one. Ask any owner how many open enquiries the business has right now.

A: And if the answer takes more than ten seconds —

B: — the problem isn't reply speed. It's that no single place knows what exists. Enquiries come in on WhatsApp, email, a web form, the phone, sometimes Instagram. Every channel has an owner. The set of channels doesn't.

A: And the cost is invisible, which is what makes it expensive. A slow reply gets you a complaint — you hear about it. A missed enquiry produces nothing. No signal at all. Just someone who went elsewhere.

B: The instinct there is also the demo-shaped fix — put a bot on WhatsApp so nobody waits.

A: Which adds a fifth channel. It doesn't fix the four you already have.

B: The actual fix is almost insultingly boring. One queue, every channel lands in it, tagged with its source, one named owner for the set — not per channel, for the whole set.

A: And only after that is in place does AI enter — classifying, drafting, summarizing. Never sending.

B: Let's do the second one — approvals, because it's where this distinction gets tested hardest. The quote sitting from 11 to 4.

A: The tempting fix is: remove the person, let the system send it. And the arithmetic looks unanswerable — you remove the bottleneck, response time goes from hours to seconds.

B: What the arithmetic doesn't show you is the distribution of outcomes. Most automated replies are fine. A small fraction are wrong in expensive ways — a price you can't actually honor, a date you'll miss.

A: And you don't find that out from a log. You find out from the customer, late, and by then it's been repeating the mistake for two weeks.

B: So the fix isn't removing the approval. It's speeding it up — put the draft, the request, and the history on one screen so the decision takes seconds instead of minutes. Widen who's allowed to approve, because very often one person is the bottleneck only because nobody ever named a second approver.

A: Which is a staffing decision dressed up as a technology problem.

B: A lot of this season is that sentence, honestly.

A: So — where does a human absolutely have to stay?

B: Anywhere the action is irreversible, more than anywhere it feels important. A quote that goes out, a refund, a date you commit to. That's the actual test — not "is this significant," but "can I take it back."

A: Which is genuinely useful, because "important" is a feeling and "reversible" is a fact you can check.

B: So — one thing, for someone listening at their desk right now, to actually check this week.

A: Time it. Don't estimate it — actually time how long a status update or a quote takes to answer these three: how many open items exist right now, which is the oldest untouched one, and who owns it.

B: If you can't answer all three in under a minute, that's the workflow to fix first. Before any AI.

A: The full written version of this is on the site — atinnovators.in, under insights. That's it for this one.