Jonathan Dickerson
Case study · Bank of America

Seven agents and a playbook

I built seven AI agents in the Microsoft 365 Copilot agent builder. They cover a content designer's workflow, from project kickoff to final presentation. Then I started a working group and led a Playbook, so the rest of the department could work the same way.

Where it started

I went through my own workflow at Bank of America's Experience Design group, looking for anything annoying or slow. Then I built an agent for each one.

  • Checkers

    Project readiness · Accessibility reviewer · Design system standards reviewerHold work against a standard and say what is missing.

  • Synthesizers

    Research synthesizer · Feedback synthesizerTurn a pile of input into something a person can act on.

  • Makers

    Content first draft generator · Presentation developerGet a person past the blank page.

KICKOFF Project readiness CHECKS DISCOVERY Research synthesizer SYNTHESIZES DRAFTING Content first draft generator MAKES REVIEW Accessibility reviewer CHECKS Design system standards reviewer CHECKS CRITS + FEEDBACK Feedback synthesizer SYNTHESIZES PRESENTATION Presentation developer MAKES
Seven agents, each placed at the stage of the content design workflow where it does its job, from kickoff to the final presentation.

The agents apply rules the organization already has: process documentation, content standards, brand guidelines and WCAG 2.2. They do not improvise their own.

The presentation developer began as four separate prompts. I chained them into one agent that runs in order: outline, slides, gap analysis, speaker notes. It critiques its own draft before it writes the notes.

From tools to a practice

Tools one person built are not a practice. I started a voluntary group of content designers that meets every two weeks to compare how we use AI, and I run it. The Playbook came out of that group, and I lead it.

It holds twelve prompts in four groups, organized by task: Communicate, Analyze, Review & Refine, and Presentations. I wrote nine. Colleagues wrote the other three, plus a best-practices section.

The format was my idea: click and play. Each prompt gets one slide that says what you need and what you will get. One click launches the prompt in Copilot.

Sample slideIllustration · generic prompt
Prompt
Readability scoring
What you need
Your draft copy and the audience it is written for.
What you get
A reading-level score, the sentences that raise it, and suggested rewrites.
Start
One click opens the prompt in Copilot.

In use

Other content designers have used these on real projects.

  • Three studies at once. One team ran three user studies at the same time. They used the research synthesizer to get through all three under deadline.
  • A second opinion that agreed. A colleague ran the research synthesizer on her own study. The output aligned with what she saw when she reviewed the results herself.
  • A first pass before formal review. A team used the accessibility reviewer ahead of its formal accessibility review. The review identified no high-severity issues.

I built the agents and presented them with my team at a department promptathon. A junior content designer on the team contributed an eighth agent, an error message reviewer.

The Playbook is complete and in review. Launch is planned for the end of October 2026.

Back to all work