An AI Agent Rollout Plan for Small Companies, in Phases

A good AI agent rollout plan moves through phases, and every phase has an entry condition you must meet before starting and an exit condition you must prove before moving on. For a small company the usual sequence is preparation, shadow mode, drafts with human approval, limited authority, and then controlled expansion. The goal is that your team's normal work never depends on an agent that has not yet earned the trust.

Key takeaways

• Roll out one role at a time, and move to the next phase only when the exit condition of the current one is demonstrated, not when a calendar says so.
• Start with the agent observing or drafting, and grant authority to act only after its work has matched human judgment on real tasks.
• Write down, before each phase, what the agent may read, what it may change, and who approves what.
• Keep a way back at every phase: the manual process must stay alive until the agent has proven it can replace it.
• Expansion is a new rollout, not an extension of the old one. A second role gets its own phases.

Why does a small company need phases at all?

Large companies can absorb a bad first week with an agent. They have a quality team, a spare analyst, a change manager. A small company has none of that, so a mistake lands directly on the owner or on the one person who handles customers, invoices or scheduling.

Phases turn a risky leap into a series of small, reversible steps. Each phase answers one question. Can the agent do the work correctly? Can the team review it without extra burden? Can it act on its own inside a narrow boundary? Can we keep it running without surprises? If the answer is no, you stay where you are or step back. Nothing breaks, because the previous way of working is still in place.

There is also a people reason. Staff accept a new colleague more easily when they can watch it work before it touches anything they are accountable for. A phased plan gives them that time and gives them a say in the exit decision.

The plan below uses five phases. They are deliberately unglamorous. Most of the value comes from the discipline of the entry and exit conditions, not from the tools.

Phase zero: what must be true before any agent runs?

Phase zero is preparation, and it is the phase small teams most often skip. The entry condition is simple: someone in the company has decided that a specific role is worth an agent. The exit condition is a one-page description that the team can read and challenge.

That page should cover the following:

• The role in plain words: what the agent does, for whom, and what it explicitly does not do.
• The one workflow to start with. Your first AI workflow should be small enough to prove, which means frequent, bounded and easy to check.
• The sources the agent will rely on, and which of them is the source of truth when two disagree.
• The tools it needs, with the smallest access that lets it work.
• The approval boundary: which actions need a human decision, and who that human is.
• One observable success criterion, such as drafts accepted without rewriting, or requests routed to the right person on the first try.

If you cannot fill in one of these lines, that is useful information. It usually means the process itself is not yet clear enough for a person to follow consistently, and an agent will inherit that ambiguity. Fix the process first, or pick a different workflow.

Phase zero also includes tidying the documents the agent will read. Outdated price sheets, duplicate policies and unlabeled folders become the agent's mistakes later.

Phase one: can the agent shadow the work without touching it?

The entry condition for phase one is the finished phase zero page. In this phase the agent receives real inputs and produces its proposed output, but nothing leaves the building. A person does the work as usual, and the agent's version is compared afterwards. This is the idea behind running an AI agent in shadow mode before giving it authority.

Shadow mode gives you evidence at almost no risk. You see where the agent misreads a request, where it lacks context, where your documents contradict each other, and where it is simply better or faster than the habit it replaces.

Keep the review light but structured:

• Collect a set of real cases, including awkward ones, and note the human outcome for each.
• Compare the agent's output to the human outcome and label each difference: agent wrong, human wrong, or both acceptable.
• Record why the agent failed, in categories you can act on: missing context, wrong source, unclear instruction, tool failure.
• Fix the cause, not the symptom, and rerun the same cases.

The exit condition is that the agent matches acceptable human judgment on the cases you chose, that its remaining failures are understood, and that the people doing the review agree the output is worth reviewing. If the review feels like redoing the work, the agent is not ready, or the task is too broad.

A team can describe tests like these as a short written task contract and a test set, and Anthropic's guidance on [building effective agents](https://www.anthropic.com/engineering/building-effective-agents) recommends starting with the simplest solution and adding complexity only when it demonstrably helps.

Phase two: when should the agent draft and a human approve?

The entry condition is a passed shadow phase. Now the agent's output goes into the real workflow, but only as a draft or a proposal. A person approves, edits or rejects it before anything reaches a customer, a record or a payment.

This is where the approval boundary from phase zero becomes concrete. Decide where your AI agent should stop and ask by looking at consequences: an action that is hard to undo, visible to a customer, or tied to money or personal data deserves a human decision. A summary for internal use may not.

Three practices keep this phase from becoming a bottleneck:

• Make approval fast. The reviewer should see the proposed action, the evidence behind it and the alternatives, not a wall of text.
• Log every edit. The changes people make to drafts are the best signal of what the agent still gets wrong.
• Name an owner for review. If everyone may approve, in practice nobody does it on time.

Permissions matter here too. Give the agent read access and the ability to create drafts, and withhold the ability to send or write final records. The idea of least privilege for AI agents is to start with the smallest access that makes the work possible and widen it only for a reason you can state. The OWASP [authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) supports the same principle for any system that acts on behalf of a user.

The exit condition is a stable pattern: reviewers accept most drafts with little or no change, the changes that remain are minor, and no draft has caused harm that review failed to catch. Equally important, the reviewers say the process saves them effort.

Phase three: how much authority is safe to hand over?

Phase three is limited authority. The entry condition is a record from phase two showing consistent quality on a particular class of task. The agent now acts on its own, but only within a narrow, named boundary, and only for that class.

Choose the first autonomous action for its low risk and high repetition. Filing a document in the right folder, tagging an incoming request, or sending a routine acknowledgement are typical candidates. Pricing changes, refunds and anything legally binding are not.

Before granting authority, settle these points:

• The exact actions allowed, written as a short list, with everything else denied by default.
• A limit on volume or value, so a mistake stays small.
• A trigger for escalation, such as low confidence, missing data or an unusual request, that routes the decision to a named person.
• A way to pause the agent quickly, without a technical meeting.
• A review rhythm: someone samples the agent's autonomous work on a schedule and reports what they found.

The exit condition is that the agent has operated within its boundary, the sampled work is correct, escalations arrive with enough context for a quick decision, and nobody has needed to undo its actions in a way that surprised them. If it fails, step back to phase two for that class of task. That is a normal outcome, not a defeat.

Phase four: how do you keep it running and expand safely?

Phase four is operation. The entry condition is a successful phase three, and the work is no longer about proving the agent but about maintaining it. Connections expire, documents change, staff turn over, and the business itself shifts. An agent that was correct last quarter can drift quietly.

Set a recurring routine: review the agent's logs and exceptions, re-run your original test cases, check that permissions still match the role, and update the sources it relies on. Record changes to instructions so you know what altered behavior and when.

Expansion comes only after this routine is working. A new workflow, a new channel or a new tool goes back to phase zero for its own page and its own shadow period. Reusing what you learned makes it quicker, but the evidence still has to be gathered. Skipping phases because the first one went well is the most common way small rollouts fail.

The NIST [AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) describes this as continuous governance: map the context, measure, manage and keep reviewing. You do not need its full structure, but its habit of treating oversight as ongoing is the right one.

An illustrative scenario: a small services firm and its inbox

This scenario is illustrative, not a real case. A small consultancy wants an agent to help with incoming client requests. In phase zero, the owner writes that the agent will read the shared inbox, classify each request by type, and propose a reply. It will not send anything, and it will not touch the billing system.

In phase one, the agent runs on incoming mail for a stretch while the office manager keeps working as usual. At the end of the stretch she compares the agent's classifications with hers and finds that it confuses scheduling changes with new project requests whenever a client writes informally. The team adds a clearer definition of each type, with examples, and reruns the same cases.

In phase two, the agent's proposed replies appear as drafts. The office manager approves or edits them. Edits cluster around tone for two long-standing clients, so the team adds that context to the agent's notes. In phase three, the owner lets the agent send only a short acknowledgement for scheduling confirmations, capped to a small daily volume, with anything unusual escalated. Phase four is a weekly look at what the agent did and what it escalated.

At no point did the office manager lose the ability to do the work herself. That is the real test of a phased plan.

How does Be Human run a rollout like this?

Be Human designs, implements and operates personalized AI infrastructure for companies, and a first engagement follows the same logic as the phases above. It defines one useful role, the memory and tools that role needs, the approval boundaries and an observable success criterion. Agents keep their identity, memory, channels, tools and permissions under human control. Managed operations can then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, so a small team does not have to carry phase four alone. Work is remote, in English and Portuguese.

Frequently asked questions

How long should each phase last? It depends on how often the workflow happens and how costly mistakes are, not on a fixed schedule. A phase ends when its exit condition is met, so a frequent, low-risk task can move quickly while a rare, sensitive one needs more time in shadow mode and review.

Who should own the rollout in a company with no technical staff? Name one business owner who understands the workflow and can approve or reject the exit conditions. Technical work can be handled by a partner or a managed service, but the decision to grant authority should stay with someone inside the company who is accountable for the outcome.

What should we do if the agent fails an exit condition? Stay in the current phase or step back one, find the cause, fix it, and test again with the same cases. A failure is useful evidence about your documents, instructions or tools, and finding it during shadow mode or draft review is exactly why the phases exist.

Do we need a separate rollout for every new agent or workflow? Yes, but later ones go faster because you already have a phase zero template, review habits and an approval structure. Each new role still needs its own boundary, its own test cases and its own evidence before it earns more authority.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.