AI agents for small business: my 30-day human-controlled pilot.

For a small business, the best first AI agent is one narrow workflow with a named human owner, restricted data, a reversible action, approval before external communication and one measurable result. My 30-day pilot maps the workflow, tests historical examples, runs in shadow mode and only then allows limited human-approved use.

The point is not to install the maximum amount of AI. The point is to learn whether one business handoff can become faster, clearer or more consistent without giving up accountability.

Start with one recurring job. Define what “done” means. Keep the risky actions behind a human approval gate.

First decide whether you need an agent at all

“AI agent” is often used for any software with a chat box. I separate three different systems because they require different controls and testing.

System	What it does	Good small-business use
AI assistant	Responds to a request but does not control the workflow.	Drafting, summarizing and analyzing information for a person.
Fixed automation	Follows predictable rules and predefined steps.	Form routing, reminders and consistent record updates.
AI agent	Uses a model, instructions and tools to decide how to complete a bounded task.	Work that includes unstructured information, judgment, exceptions and a clear stop condition.

OpenAI’s agent guide recommends validating that a workflow actually needs contextual decision-making, difficult-to-maintain rules or unstructured data; otherwise deterministic automation may be enough. Anthropic similarly recommends using the simplest solution that meets the requirement and adding agentic complexity only when it improves the result.

My workflow-readiness test

I do not start a pilot until the business can answer these questions:

1. Value: What recurring delay, mistake or missed handoff matters to the business?
2. Trigger: What exact event starts the work?
3. Finish state: What structured output means the task is complete?
4. Sources: Which approved documents, public pages or records may the system read?
5. Permissions: Which tools are read-only, and which actions can change an external system?
6. Approval: Who is responsible for reviewing a customer message, commitment or record change?
7. Stop rules: Which conditions require the system to pause, suppress follow-up or escalate?
8. Measurement: Which baseline and outcome will decide whether to keep, revise or stop?

If the trigger, finish state or owner is missing, the first job is process design—not agent deployment.

The 30-day pilot

Days 1–5: establish the baseline

Observe the workflow before changing it. Record how many tasks arrive, how long they wait, where corrections happen and which downstream result matters. A pilot without a baseline can demonstrate activity, but not improvement.

Days 6–10: map the workflow and permissions

Write the approved input, required output, allowed tools, forbidden actions, escalation path and owner. Give the system the minimum permission needed for the test. A research lane does not need access to payments, production infrastructure or an entire inbox.

Days 11–15: test historical examples

Create a small test set that includes ordinary tasks, missing information, wrong-contact data, a prior reply, an opt-out and at least one ambiguous case. Record the complete trace, not just the final answer. The failure path is part of the product.

Days 16–22: run in shadow mode

Let the system prepare the work while the existing human process remains authoritative. Compare the recommendation, evidence, correction and escalation with what the owner actually decided.

Days 23–30: allow limited, human-approved use

Only after the test and shadow results are reviewed should the pilot influence live work. Keep external communication manual, preserve logs and stop the pilot if its evidence, suppression logic or approval boundary fails.

What I built in the Jeff Revenue Desk

My implementation lab uses a lead-follow-up lane because it exposes the controls clearly. The production Builder now separates Opportunity Scout, Revenue Pipeline, Jeff Revenue Desk and Outreach Desk. The system can organize evidence and prepare work, but it cannot silently advance a deal or send a message.

The current rules are intentionally narrow:

• No more than three operator-confirmed touches in one Chicago calendar day.
• Tony approves the exact business, recipient and message before every reach.
• Sending opens a manual email draft; Jeff does not autonomously send email.
• One manual follow-up can become eligible after two Chicago business days.
• Any reply, opt-out, wrong-contact signal, suppression or prior follow-up blocks another follow-up.
• Outcomes are recorded as awaiting, positive, neutral or negative reply, meeting, proposal, won, lost or no response.
• The desk’s records remain in the operator’s browser unless Tony intentionally exports a backup.

You can inspect the live Revenue Pipeline, the Jeff Revenue Desk and the public permission model.

Measure completed work and downstream outcomes

I track the smallest useful chain of evidence:

• Tasks received, completed, corrected and escalated.
• Cycle time before and during the pilot.
• Missing or invalid evidence caught before approval.
• Suppression and stop-rule events.
• Owner approvals and rejections.
• Replies, qualified conversations, meetings, proposals and customers—only when attribution is documented.
• Operating cost and incidents.

The end-of-pilot decision is not automatically “expand.” It is one of three choices: keep the workflow, revise the design or stop because a simpler process works better.

Questions I would ask any AI-agent provider

1. What exact job will the agent complete, and what is the finish state?
2. Which data and accounts can it read or change?
3. Which actions are reversible, and which require human approval?
4. What happens when information is missing, conflicting or malicious?
5. Can the owner inspect the evidence and complete execution trace?
6. What test set and acceptance criteria will be used?
7. How do replies, opt-outs and exceptions stop automated work?
8. Who owns the system after launch, and what does it cost to operate?
9. How will a claimed business result be attributed?

Sources and methodology

This is an implementation lab note based on the Builder workflow I built and tested, combined with current primary guidance. The “1–50” range describes the owner-led businesses I designed the pilot for; it is not a legal definition of a small business.

• OpenAI: A practical guide to building agents
• Anthropic: Building effective agents
• Anthropic: Demystifying evals for AI agents
• U.S. Small Business Administration: AI for small business
• NIST AI Risk Management Framework Core
• Tony Beal field note 01: useful and controlled small-business agents