Skip to article

Your First AI Agent Pilot: A Playbook for One Department

One department, one recurring task, five things written down before the start, and a condition that ends it.

By Jacqueline V. TwillieNovember 20, 20266 min readGuide
Quick answer

Pick one recurring task in one department and give it to a draft-only AI agent. Write down five things before anyone builds: the finish line, the gate, the credit rule, the stop path and the retirement criteria. Check the laptops a week ahead and split ownership with IT.

A department head asked to run an AI agent pilot is usually handed a tool and a date. The decisions that decide whether the pilot holds up are about people and permissions, and nobody has made them yet. This playbook puts those decisions in the order they come up. It is assembled from what BNEDai has published about its own builds and sessions, and each part says where it comes from.

It is written for the person asked to approve, run or sponsor a first pilot: a department head, an operations lead, or the executive above them. No technical background is assumed. The method comes from three places: BNEDai's published guides on change management and workshop design, the AI Readiness Index, and the October 7, 2026 session Jacqueline V. Twillie taught with Women Applying AI.

Three decisions come before the tool

Three, and all three come before the tool. Which one recurring task goes first. Who in the department owns it and knows what good work looks like. And what result, cost or broken rule ends the pilot.

The third decision is the one that gets skipped. The guide to applying FLOW to a technology roadmap makes the case for writing the walk-away condition before the pilot starts, while you can still write it with a clear head. A pilot with a written end condition can be approved quickly because its downside is bounded. Without one it turns into a standing commitment nobody voted for.

Start with a recurring task whose output is a draft

A recurring task whose output is a draft. A prepared brief, a first-pass triage, a summary a person still signs off on. The guide to custom AI for business gives the reason: draft-only work needs no send permission and no data-boundary negotiation, so it clears the approval bar on its own.

The department pages in this collection list published builds by function, each with what the AI agent drafts, what a person approves and what it reads: sales and account teams, customer success, executives and leadership, strategy, finance and investor work, legal, procurement and compliance and operations and project teams. They are builds you can set up.

Check that the platform opens on your team's laptops

Find out a week ahead, with IT in the room. BNEDai's Impact Report No. 1 records a private company in the Atlanta metro that brought a team to one of the June 2026 build sessions. Managed-device restrictions on the corporate laptops blocked the platform, and the team spent the hour on a workaround plan instead of an AI agent. Every BNEDai session now opens with a device check.

Before an internal kickoff, ask four questions in order: does the tool open on the real hardware, were the accounts provisioned, were the integrations approved, and has anyone told the team which existing work they may point it at. The guide to AI change management treats what is left after those four as the cultural part.

The team owns the logic and IT owns the credential

Split it in writing before the first build. The team running the process owns the logic, the judgment criteria, the review rule and the decision to stop. IT owns the credential, the device policy, the data boundary and the undo path. The guide on reducing IT dependency lists both sides. When the split is unwritten, people automate quietly on personal accounts with real passwords, which is what a governance team most wants to avoid.

Write five things down before the pilot starts

Five items, from the change management guide. Write them before the pilot, not after the incident.

  1. 01

    The finish line.

    What each participant can show at the end. A working AI agent on a task they own, not attendance.

  2. 02

    The gate.

    Which AI-assisted actions may reach the outside world without a human yes. The honest first answer in most organizations is none, stated as one line in every tool's instructions: draft, never send.

  3. 03

    The credit rule.

    Whose numbers change when the AI agent works, and what happens to the time it frees. Unanswered, every team assumes the answer is headcount.

  4. 04

    The stop path.

    A named person to ask when output looks wrong, written where the work happens.

  5. 05

    The retirement criteria.

    The result, cost or rule condition that ends the pilot. The guide to applying FLOW to a technology roadmap shows how to write the three.

Ask for the list of what it can read and what it can send

Ask for the list. In the October 7, 2026 session Jacqueline V. Twillie taught with Women Applying AI, the permission screen separated read-only tools from write and delete tools, and she showed how to restrict each group. BNEDai starts every AI agent on one rule: draft, never send. The page on what an AI agent can touch and what it cannot walks through the published rules.

A sponsor asks six questions after the first test

Six questions, each answerable by showing a screen or the words the team gave the AI agent: what did you tell it and why, can I see how it reasoned, what can it send and what can it only read, how is its access removed, when are its instructions reviewed, and what did you change after the first result. The sponsor review page has the one-page sheet.

Three checks show whether the pilot is holding at 90 days

Three checks, again from the change management guide. Take the Index again ninety days later and look for movement between bands, because a same-band bump of a few points can be a good quarter. Ask someone who did not set up the tools to explain the rules. And count the AI agents still running on real work a month after they were built.

The AI Readiness Index is BNEDai's free 18-question self-assessment. It scores how a team runs AI across eight dimensions and places it in one of four bands: Ad hoc, Watched, Gated or Governed.

The second pilot is where cohorts quietly end

The first AI agent gets built because a facilitator is in the room. The second is where cohorts quietly end. The guide to AI workshops for business gives three things that carry it: book the second session before the first ends, let the participant choose the second build, and give the group a place to show what ran. Keep a named person available for the month after.

Frequently asked questions

What should the first AI agent pilot in a department be?

One recurring task whose output is a draft, such as a prepared brief or a first-pass triage, that a person signs off on before anything goes out.

What has to be written down before the pilot starts?

Five things: the finish line, the gate that keeps AI-assisted actions from reaching the outside world without a human yes, the credit rule, the stop path and the retirement criteria.

Who owns what in a department pilot?

The team running the process owns the logic, the judgment criteria, the review rule and the decision to stop. IT owns the credential, the device policy, the data boundary and the undo path.

Reviewed on November 20, 2026.

More on Leadership for Corporate Executives