What 'agentic AI' actually means for a ten-person company

"Agentic AI" is everywhere at the moment, and almost all of it is written for companies with a thousand staff and a technology budget to match. For an owner of a ten-person firm, the phrase mostly produces a vague sense of being behind, without much idea of what it would actually mean on a Tuesday. So here is the plain version, stripped of the conference language: an agent is software that can take several steps on its own toward a goal you set, rather than waiting for you to direct each step.

That is the whole of it. An ordinary assistant answers when you ask. An agent, given a goal, will go and do a short sequence of things to reach it: look something up, use that to do the next thing, check a condition, produce a result. Azeem Azhar, who writes about where the technology is heading, describes agents working in the background as "an extension of your will," carrying out your intent without you steering every move. For a small firm the useful question is not whether this is impressive. It is which jobs it can genuinely take on, and which it cannot yet be trusted with.

A plain definition of an agent

Hold three features in mind and you will not be fooled by the marketing. An agent has a goal, given by you. It takes multiple steps, in sequence, deciding what to do next based on what just happened. And it acts, rather than only advising: it actually does the steps, not merely suggests them.

That is genuinely different from the chatbot most people picture. A chatbot is a conversation. An agent is a short errand run on your behalf. The difference matters because acting on its own is exactly what makes an agent useful and exactly what makes it riskier, and a small firm needs to hold both of those at once.

Assistant versus agent for a small firm

The distinction worth getting right is between an assistant that helps you do a task and an agent that does the task. We have written separately about sorting assistant, automation and agent work, and the short version is that they suit different jobs. An assistant is the safe, everyday workhorse: you stay in control of every step, and a mistake is caught the moment you read the output. An agent hands over the steps as well as the work, which buys you more time and asks for more trust.

For most small firms, assistants will do the bulk of the useful work for a while yet. Agents earn their place on the narrow set of jobs that are repetitive, multi-step, and forgiving enough that a wrong move costs little. The mistake is reaching for an agent because it sounds advanced when an assistant would have done the job with less risk. Match the tool to the task, not to the trend.

Two realistic small-business agent jobs

Skip the science fiction. Two kinds of job are genuinely within reach and genuinely useful.

The preparation errand. Before a meeting or a call, an agent gathers what you need: recent emails, last notes, open actions, the relevant figures, and assembles them into a short brief. Multi-step, repetitive, and low-stakes, because you read the brief before you walk in. If it misses something, you notice and add it.

The tidy-up errand. After an enquiry comes in or a job finishes, an agent does the routine follow-through: log the details where they belong, draft the standard reply for your approval, flag anything that does not look right. Again, several steps, done often, with a human glance before anything reaches a customer.

Notice what these have in common. They are dull, frequent, and built so a mistake is cheap and visible. That is the shape of a good first agent job, and it is the opposite of the dramatic, autonomous use cases the hype tends to lead with.

The right first agent does a boring errand you can check, not a dramatic job you have to trust blind.

The review-and-proceed pattern

The pattern that makes agents safe in a small firm is simple, and Nate B. Jones describes a clean version of it. The most reliable agent work, in his framing, is the kind where AI drafts, a human reviews, and the AI then proceeds. The agent does the legwork up to a decision point, a person makes or confirms the call, and the agent carries on from there. You keep judgement at the moments that matter and hand over the routine around them.

This is the design that lets a small firm use agents without losing control. You are not choosing between full autonomy and doing everything yourself. You are inserting a human checkpoint at the point of consequence, and letting the agent run freely on either side of it. Get the checkpoint in the right place and an agent becomes a trustworthy worker rather than an unsupervised risk.

The honest limits

Autonomy needs an audit trail. An agent that takes steps on its own must leave a record of what it did, so that when something goes wrong, and eventually it will, you can see what happened and fix it. In a small firm with no one watching the system full time, the ability to look back and understand an agent's actions is not a luxury. It is the thing that lets you sleep while it runs.

And keep the early agents where mistakes are cheap. The places agents fail are the places they act confidently on a wrong assumption, and the cost of that depends entirely on what you let them touch. An agent preparing your meeting brief can be wrong harmlessly. An agent sending invoices, moving money or messaging customers without a checkpoint is a different proposition, and not where a small firm should be learning. Start where being wrong costs a minute, and earn your way toward more.

It is also worth being sober about the hype, because a lot of what is sold as agentic is not yet reliable enough to run unwatched on real work. The demos are real and the trajectory is genuine, but the gap between a polished demonstration and a routine you would trust with a customer is wide, and it is your firm that pays when it is crossed too early. The sensible posture is neither dismissal nor breathless adoption. It is to use agents now on the narrow band of jobs that are ready, watch them closely, and widen the band as they earn it.

What to do next

Pick one repetitive, multi-step job that is cheap to get wrong, the meeting prep, the enquiry tidy-up, and sketch it as an errand: the goal, the steps, and the one point where a human should review before it proceeds. Build the smallest version, keep the audit trail on, and run it with you checking the output for a fortnight. You will learn more about what agentic AI means for your firm from one real errand than from a month of reading about it.

Working out which jobs are genuinely ready for an agent, and where to put the human checkpoint, is a core part of the AI Advisory for Teams work. If "agentic AI" has been a fog of jargon, book a business call and we will make it concrete.

Sources and further reading

• Nate B. Jones, "Executive Briefing: Your 2025 AI Agent Playbook in 10 Minutes," Nate's Newsletter, 2025. A clean version of the review-and-proceed pattern: AI drafts, a human checks, the AI carries on.
• Azeem Azhar, "AI in 2030," Exponential View, 10 May 2025. His description of agents working in the background as an extension of your will.