You already have an AI agent: start at the boundary of tedium

Say "AI agent" to a small-business owner and they picture something expensive and slightly science-fiction: an autonomous system running the company while everyone sleeps. So they file the whole idea under "not for us yet" and move on. That instinct skips the one place an agent actually earns its keep in a small firm, which is not the dramatic work at all. It is the dull, fiddly, repetitive work that sits in the gap between "this needs doing" and "not right now."

Azeem Azhar, who writes one of the most-read newsletters on where this technology is heading, coined a phrase for that gap. He calls it the boundary of tedium, and reframing the first agent around it is the most useful thing a small firm can do with the idea. "Every knowledge worker has a boundary of tedium," he writes: work that is "too boring or too fiddly for you to do yourself, but too complex or too specific to easily hand off to someone else." That is exactly where your first agent belongs.

Where the boundary sits

Every owner knows the feeling. There is work that is plainly worth doing, and yet it never gets done well or on time, because it is too dull to want to do and too particular to delegate cleanly. Tidying the CRM after a busy week. Pulling together the five things you need before every client call. Chasing the missing piece of information from three different people. Turning a pile of receipts into something your accountant can use. None of it is hard. All of it is friction, and friction is what quietly drags a small business down.

That friction is the boundary of tedium, and until recently it had nowhere to go. Too small to hire for, too specific to outsource, too boring to prioritise. An agent changes the economics of exactly that band of work, because software that can take a few steps on its own can absorb the fiddly middle while you keep the judgement.

What an agent actually is

Strip away the mystique. Dharmesh Shah, a software founder who builds these tools, describes the useful unit as a minimum viable agent: a small piece of software that, in his phrase, "goes through multiple steps to accomplish a goal." That is the whole idea. Not one answer to one prompt, but a short sequence run on your behalf: fetch this, check that, draft the next thing, flag anything odd. The first agent in a small firm is not a digital employee. It is a sequence that clears one repetitive job end to end.

Holding that plain definition matters, because it lowers the bar from "build a robot" to "automate one annoying routine," which is something a small team can actually do.

List your firm's tedium

Before you build anything, find the work. For one week, keep a running note of every task that makes you sigh: the small, repeated, fiddly jobs that you do badly or late because they bore you. Do not filter yet. Just catch them.

Then look for the ones with three features. They happen often, so clearing them is worth something. They follow roughly the same steps each time, so a sequence can handle them. And a mistake is cheap and easy to spot, so you are not handing a high-stakes job to early software. Those three features, frequent, repeatable, low-stakes, are what make a task a good first agent.

Your first agent is not a digital employee. It is one boring routine, finished without you.

Pick one and build the minimum version

Resist the urge to pick the biggest, most impressive candidate. Pick the most boring one that scores well on all three features. Meeting prep is a common winner: before each call, gather the recent emails, the last notes, the open actions and the relevant numbers into a short brief. Dull, frequent, low-stakes, and a real time sink. Ideal.

Then build the smallest version that works, not the complete one you can imagine. Shah's whole argument for the minimum viable agent is that you start with the simplest sequence that does the job, get it running on a real task, and only add to it once it has proved itself. A first agent that reliably preps your meetings is worth ten you sketched and never finished. Use the tools you already pay for where you can. The goal is one routine cleared, not a platform.

Other strong first candidates follow the same shape. The intake routine: when an enquiry arrives, pull the relevant details into one place, draft a first reply for you to approve, and add the lead to wherever you track them. The follow-up routine: after a call, turn your rough notes into a tidy summary and a short list of who owes what by when. The reconciliation routine: each week, gather the receipts and flag the ones that do not match. None of these are glamorous. All of them are the small recurring jobs that decide whether a Friday afternoon is spent on the business or on its admin. Clear one properly and the case for the next makes itself.

The honest limits

Keep a human owner and a check on every agent, especially early. An agent that takes steps on its own can take wrong steps on its own, confidently, and a small firm has no team standing by to catch it. So the routine should always surface what it did for a quick human glance, and there should be one named person who owns it when it misbehaves. Autonomy without an owner is how a helpful sequence quietly does the wrong thing for a fortnight before anyone notices.

Match the autonomy to the stakes, too. Letting an agent draft your meeting brief is fine if it occasionally misses something, because you read the brief anyway. Letting an agent send things to clients, move money, or change records without a human glance is a different risk, and the boundary of tedium is the wrong place to start taking it. Begin where a mistake costs you a minute, not a relationship.

What to do this week

Spend the week catching your tedium, then pick the single most boring task that happens often, follows the same steps, and is cheap to get wrong. Build the smallest sequence that clears it, run it for a fortnight with you checking the output, and only then decide whether to widen it. One annoying routine, finished without you, is a better introduction to agents than any demo, and it tends to make the next one obvious.

Working out which routines are worth automating, and which should stay in human hands, is a core part of the AI Advisory for Teams work: finding the genuine time sinks and turning them into safe, owned routines rather than scattered experiments. If that is where you want to start, book a business call.

Sources and further reading

• Azeem Azhar, "The boundary of tedium has moved," Exponential View, 2025. Where the "boundary of tedium" comes from: work too boring to do yourself and too fiddly to hand off.
• Dharmesh Shah, "The Secret Behind How Agent.AI Got Started," Simple.ai, 2024. Shah builds agent tools; this is his "minimum viable agent", software that takes a few steps to reach a goal.