How to plan an AI workforce for a small team

An AI workforce is a group of AI agents assigned to carry out defined pieces of work for a business. For a small team, start with two agents, each responsible for a recurring task with an output someone can check. Give each a spending limit, decide which actions need approval, and review the arrangement after 30 days. There is no reason to build a large roster before you know which work is worth handing over.

The temptation is to start with job titles. Hire an AI researcher, an AI assistant, perhaps an AI operations manager. Soon there is an impressive-looking organisation chart and nobody has decided who checks the figures in Friday's report. We would start with the work already sitting on someone's desk. That tells you what the agents should do and what a satisfactory result looks like.

Make a list of work that keeps coming back

Take the past two working weeks and look at the tasks your team repeated. Check the support inbox, meeting calendar, finance folder and shared task list. Look for work that returns predictably, takes recognisable inputs and leaves behind something another person uses.

Do not worry about agent templates yet. Write the actual task, preferably using the team's own words. 'Prepare the Monday customer issue report' is useful. 'Improve customer experience' is too broad to assign or inspect.

A small operations team might produce this list:

• Sort support requests by topic and urgency, then prepare suggested replies.
• Collect competitor announcements every Friday and prepare a research brief.
• Match supplier invoices against purchase records and flag missing information.
• Prepare background notes before recurring client calls.
• Update customer records after support conversations.
• Negotiate unusual refund requests and decide who receives exceptions.

These tasks do not all belong in the same bucket. The first four produce material someone can inspect before using it. Updating records changes information people rely on. Refund negotiations require judgement about relationships and money. Both might eventually involve agents, but they make poor opening assignments for a team still learning how to supervise one.

Choose two tasks whose outputs can be checked

Suppose you choose support triage and the weekly research brief. They cover different kinds of work, yet both produce material that can be reviewed before anyone acts. A support lead can compare suggested categories with original messages. An operations lead can open the research sources and check whether the brief represents them accurately.

For each candidate, ask how a person would recognise an error. A task is easier to delegate when the answer does not depend on the agent's opinion of its own performance.

Pick only two. Starting with six makes a poor result harder to diagnose. Was the brief unclear? Was access missing? Did reviewers lack time? With two assignments, you can find the cause without changing the team's entire working routine.

There is another useful test. Can the reviewer inspect the output faster than producing it? If a research brief takes nearly as long to fact-check as it would to write, narrow its scope or specify better evidence before proceeding.

Give each agent one job and a budget

Write one brief for each selected task. Describe the input, required work, finished deliverable and circumstances in which the agent must stop. Name the person responsible for reviewing the result. The support lead owns support triage. The operations lead owns the research brief.

A support assignment might say: 'Each weekday, review new support messages, place them into our existing categories and prepare draft responses. Use approved support material. Flag requests about payments, account access or unclear policy. Do not send replies. Deliver a queue for the support lead to inspect.'

That brief has boundaries. It does not tell the agent to 'handle support' or give it permission to invent refund policies. The agent prepares the material, while a person decides what the business promises.

Set a maximum budget for each delegation. A research assignment requiring repeated searches may need a different cap from a short inbox sorting job. Begin with enough room for one ordinary assignment, then examine the cost before increasing it. A spending limit controls consumption. It cannot tell you whether a result is correct.

In Delagents, a person writes a brief, an agent carries out the delegation, and the result arrives as a deliverable. Each delegation has a hard budget cap. The roster includes a support triager and a research analyst, so the team can start with an agent suited to the work rather than inventing a role.

Decide what must stop for a person

An agent preparing a reply and an agent sending that reply are doing different jobs. The first produces text for inspection. The second makes a statement to a customer on behalf of the business. Your operating plan must distinguish them before the first assignment runs.

An agent can prepare a supplier payment list without authorising payment. It can propose a customer record correction without overwriting existing information. It can draft an announcement without publishing it. These distinctions should appear in the briefs, not emerge after a mistake.

Make the rules explicit for your first two assignments:

• The support agent may classify messages and draft replies, but a person approves any message before sending.
• The research agent may prepare an internal brief, but cannot publish findings without approval.
• Neither agent may change its own permissions or connect a new source of business data independently.
• The reviewer may reject a proposed action or request changes when the evidence is insufficient.

Delagents holds every action with an effect outside the workspace in an approvals inbox. Each card shows the exact proposed action, cost so far and deadline. A person approves, changes or rejects it. For operations and support teams, this keeps preparing customer responses separate from authorising what gets sent.

An approval should be specific enough to judge. 'Send this reply to the customer about order 418' gives the reviewer something to inspect. 'Continue working on support' does not. Decide who handles requests when the usual reviewer is absent. Silence must never be treated as permission.

Run the first month as a controlled trial

During the first week, run both assignments on a small set of ordinary work. Compare each deliverable with its inputs. If the support agent labels a billing complaint as a general question, record it. If the researcher misses the release date in an announcement, record that too. Specific errors tell you what to change.

In the second week, revise the briefs where mistakes repeat. If the support agent keeps guessing what customers mean, tell it to flag ambiguous requests. If research briefs include irrelevant announcements, narrow the subjects and specify the evidence each item needs.

During the third week, watch interruptions. Do approval requests reach the right person? Can that person decide without reopening the entire assignment? An agent that saves preparation time but repeatedly interrupts someone who cannot approve its work creates another queue.

In the fourth week, gather the completed assignments and measure what happened. Keep a simple record of:

• Assignments that finished with a usable deliverable.
• Human review time and corrections required.
• Repeated errors and unfinished assignments.
• Cost per delegation and budget caps reached.
• Approvals that caused delays, changes or rejections.

Count rejected work too. A polished report nobody used is not a useful result. An agent that stopped because it lacked a document may have followed its instructions correctly, but the workflow still needs attention. Keep those two findings separate.

Make the next staffing decision from the evidence

At the 30-day review, put both assignments side by side. Ask each reviewer to show a finished deliverable, a corrected one and an approval decision. Examine the work a person actually had to do, rather than the agent's description of its performance.

There are three reasonable outcomes. Keep an assignment that produces dependable work at an acceptable cost. Narrow one whose output is useful but takes too long to correct. Stop one that consumes budget without reducing the team's workload. Your two agents may earn different decisions.

Do not add another agent simply because the first two completed their briefs. The next task needs its own owner, review method and budget. If it depends on another agent's output, decide who checks that material before it moves forward. Otherwise a wrong category or missing date can pass through several assignments unnoticed.

Open your team's task list now. Mark two recurring jobs that end in something a person can inspect. Write a one-paragraph brief for each, name the reviewer, set the spending cap and put a review meeting on the calendar for 30 days from now. Leave the rest of the task list alone until that meeting.