One Giant AI Assistant or Twenty Narrow Ones? Give Every Agent One Job

Owners who start with AI agents often picture one assistant that does everything. It reads the inbox, finds clients, writes the proposals and posts the articles. In practice that agent ends up with every permission, every context and no clear way to check what it did. The alternative is a team of narrow agents, each with one job and a clear handover. This is a blueprint of what such a team could look like in a small firm. It is an illustrative pattern, not a description of any particular company.

Why a small firm would want digital employees at all

A one-person firm has the same jobs as a ten-person firm. Someone has to find work, research it, write to people, publish, and keep track of it all. When that someone is the owner, the work that pays gets squeezed by the work that doesn't. The point of agents is not to remove the owner. It is to take the routine steps off the owner's desk so their attention stays on judgment: which client to pursue, what to say, what to sign.

That only works if the work is mapped first. An agent can only follow a process someone has written down. If you want to see what a map looks like on paper, the business process mapping examples show the method.

The blueprint, desk by desk

Here is how a team of about twenty agents could divide a small firm's sales and publishing work. Most firms need far fewer; the count matters less than the rule that each desk has one job.

• Lead finders, one per market. Say six agents, each searching a single region or industry. Splitting the search keeps each one narrow and stops them duplicating work.
• A lead scorer. Every new lead gets a score, for example 0 to 10, once, when it arrives. Only leads above a threshold move on. Everything else waits in the database.
• A researcher and drafter. For each shortlisted lead, it reads up on the firm and writes a draft message. The draft is saved unsent.
• A pipeline grader. It reports what is moving and what is stuck, so nobody has to read every row.
• A small content desk. Scouts look for stories worth covering, one agent drafts, another edits. A person approves what goes live.
• A router. Work comes in through one agent that passes it to the right desk, so the owner gives instructions in one place.

Tools now exist that are built for this kind of team. Three of them are compared in Dots, Grok Bot or Muse.

Where the memory lives

Every agent team needs one shared place where the work lives: the list of leads, their scores and where each one stands. A notes tool is fine while the list is small. As it grows, agents spend more and more effort reading and re-ranking pages just to work out what to do next. Moving the list into a simple database, and scoring each item once, keeps that effort flat. The cost side of that choice, with a worked example, is in an AI agent that re-reads its list every 15 minutes.

The rule that matters most: agents draft, a person sends

The riskiest desk in the blueprint is the one closest to customers. An agent can find, score, research and draft at a pace no person can match. Letting it send messages to prospects or clients on its own is a different decision. Whatever it writes goes out in your firm's name, and you own it. A tribunal made that point when it held Air Canada responsible for its chatbot's answer.

Outreach is also about relationships. A message that a named person has read, corrected and chosen to send carries that person's judgment. The safer design is simple: the draft lands in a person's drafts folder, and that person reads it, fixes it and sends it. The agents do the heavy lifting; the person keeps the relationship.

What this means if you run a small firm

You probably don't need twenty agents. Most small firms need two or three, pointed at the loop that eats the most hours. There is a shortlist in AI agents for small business: the three to set up first. The rules carry over at any size:

• Map the work before you build anything.
• Give each agent one job. Narrow agents are easier to check and cheaper to run.
• Score at the source. Decide what is worth attention once, early, so nothing downstream wastes effort.
• Keep a person on anything that reaches a client. Agents draft. People decide and send.
• Price the finished job, not the tokens. See price 1 finished task, not 1 million tokens.

Not sure whether your task needs an agent at all? Take the test on AI agents for business. Some jobs need plain automation, and some should be left alone.

Frequently asked questions

How many AI agents does a small firm need?

Usually two or three to start, each pointed at one repeatable loop of work, such as turning inbox requests into records or scoring new enquiries. Larger teams make sense only once each job is mapped and checked.

Is it better to use one AI assistant or several narrow agents?

Narrow agents with one job each are easier to check, need fewer permissions and are often cheaper to run. One assistant doing everything ends up with every permission and every context.

Should AI agents send emails to clients on their own?

For most small firms, no. Let agents research and draft, and have a person read, edit and send anything a client or prospect will see. You are responsible for what goes out in your firm's name.

Want help designing yours?

Tell Ada about one loop of work that still runs through you. She'll help you decide which desk to build first and where a person must stay in the loop, and put it in your free report. Talk to Ada, or see how BuildBrain approaches AI agent setup for small firms.