The first question every team asks about AI is "what's the most impressive thing we could do with this?"
It's the wrong question. The impressive thing is usually the risky thing, and the risky thing never ships. The better question is duller: what work happens every single day, follows a shape, and costs us money when it's late?
Answer that honestly and you land on four workflows. Almost every business has them. They're not glamorous. They're where the money is.
1. Intake
Something arrives — a WhatsApp message, a form, an email, a missed call. Right now a human reads it, works out what it is, and puts it somewhere. That's intake, and it's pure overhead. Nobody was ever promoted for correctly filing an enquiry.
An agent is good at this because the job is classification plus extraction, not judgment. What kind of request is this? Who's it from? What are they asking for? Is anything missing? Those are answerable from the message itself, and when they aren't, the correct behaviour is to ask one clarifying question — which is also something an agent does well.
Start here because intake failures are cheap. If it mis-files an enquiry, someone re-files it. Nobody gets hurt, and you get a live stream of real inputs to measure against.
2. Qualification
Once you know what arrived, you need to know whether it's worth your time. Is this a real buyer or someone comparing prices at midnight? Do they have a timeline? Are they in your service area? Have you spoken to them before?
Qualification is where most businesses quietly leak revenue, because it's done inconsistently — thoroughly on a slow Tuesday, not at all on a busy Friday. An agent applies the same criteria at 2am on a public holiday as it does at 10am on a Monday. That consistency is worth more than any individual judgment call it makes.
The key design decision is that qualification should score and route, not reject. An agent that says "this looks low-intent, here's why" and moves the lead down the queue is useful. An agent that deletes the lead is a liability.
3. Scheduling
Booking a time is a solved problem that remains, somehow, unsolved. It's a negotiation with three constraints — your calendar, their availability, and travel or prep time — conducted over a channel that wasn't designed for it.
This is a genuinely good fit for an agent, because the constraints are all machine-readable. Your calendar is an API. Their availability arrives as free text ("anytime Thursday after lunch, or Friday morning") that a model reads without breaking a sweat. The output is a calendar event and a confirmation message.
It's also the first place where a customer feels the difference. A reply at 11pm that offers three real slots and holds one is not automation they tolerate. It's service they notice.
4. Follow-up
The most valuable of the four, and the one humans are worst at. Not because we're lazy — because follow-up requires remembering something at a moment when nothing is prompting you to remember it.
An agent has no such problem. It knows the quote went out eleven days ago, that nobody replied, that the last message ended on a question, and that this client historically responds to a short nudge rather than a long one. It drafts accordingly.
Follow-up is where the compounding happens. Every business has a pile of conversations that died from silence rather than rejection. Working that pile consistently is often the single highest-return thing an agent does in its first month.
Why these four and not something cleverer
Three reasons.
They're high-frequency. You get enough runs per week to actually measure whether the thing works. A workflow that fires twice a month teaches you nothing for a year.
They're recoverable. Every failure mode has a cheap fix — re-file it, re-score it, re-send it. Compare that to an agent that issues refunds or signs off on compliance. Those come later, once you've earned the trust, and they come with an approval step.
They're where the hours actually go. Ask anyone in your team to log a week honestly and this quartet will eat a third of it. Not the hard third. The tiring third.
The trap: automating the task instead of the loop
The most common mistake is to automate one step and leave the human holding the rest. An AI that drafts a follow-up email but doesn't know whether it was sent, doesn't see the reply, and can't tell you what happened next hasn't saved you anything. You've swapped writing for supervising.
The unit that matters is the loop: notice → decide → act → record → watch for the response. An agent that owns the whole loop for a narrow workflow beats one that owns one step of a broad one. Every time.
That's the shape of the crew we run at Microcrew — conversations in on WhatsApp and Telegram, bookings out the other side, with the four loops above closed end to end.
Where to start on Monday
Pick one. Follow-up, most likely, because it's the one nobody defends.
Write down what happens today, step by step, including the bit where someone forgets. Count how often the loop runs in a week. Decide what a bad outcome costs — genuinely, in dollars or reputation. If the answer is "a mildly awkward email," you've found a safe first workflow.
Then decide how much of it the agent gets to do without asking. That's a separate question, and an important one — we'll get to it on Friday.