Marcus Andrey Study
Ler em português Library

AI for UX · For leads · 5 min read

Building an AI agent roadmap for your UX team

A method and template for deciding which AI agents your design team should build first: inventory the real work, find the parts that are gathering and checking, score candidates on value, risk, data sensitivity and effort, choose their form, and sequence them. With a worked example of a fictional team.

Most teams start with "what can this tool do?". A better question is "what does our work actually consist of, and which parts would we happily hand over?". The most valuable agent is rarely the most impressive one. It's the one that removes frequent, low-craft work that nobody enjoys, while a designer still decides. This lesson gives you a method, a scoring template and a worked example. For what agents and skills are, see UX agents and skills: a guide for designers.

What you'll be able to do after reading:

  • Inventory your team's real work from evidence, not opinions
  • Turn work into candidate agents with a one-line description
  • Score candidates on value, risk, data sensitivity and effort
  • Choose the right form for each and sequence them into a roadmap

1. The method in five steps

  1. 1Inventory

    What the team really did in 3 to 6 months.

  2. 2Candidates

    The gathering, checking and formatting parts.

  3. 3Score

    Value, risk, data sensitivity, effort.

  4. 4Form

    Skill, sub-agent, scheduled, orchestrator.

  5. 5Sequence

    Now, next, later; calibrate each one.

Every quarter: back to step 1.

From real work to an agent roadmap, in five steps. Diagram: Marcus Andrey, CC BY 4.0.

2. Step 1: inventory the real work

Look back over the last three to six months:

  • Tickets assigned to designers (types, titles, labels).
  • Wiki pages designers created or edited.
  • Recurring meetings and what they produce.
  • A fifteen-minute conversation with each designer: "What did you do last week that felt repetitive?"

Group the work into types and estimate each type's rough share. Also note hidden work: decisions buried in long meeting notes, rules that live only in people's heads, questions asked again and again. These are often good targets.

3. Step 2: find candidates

For each type of work, ask which part is gathering, checking or formatting rather than judging, what the agent would read, what it would produce, and where a person must decide. Write each candidate as one line:

"Given X, it reads Y, and produces a draft Z for a designer to approve."

4. Step 3: score them

×2Value

How often, how long, how disliked?

1 rare and quick · 5 weekly, slow, tedious
×1Risk

What if it is wrong and nobody notices?

1 reaches users · 5 obvious, caught in review
×1Data sensitivity

What data, and is the tool cleared for it?

1 personal data, not cleared · 5 internal docs, cleared
×1Effort

How hard to build and keep running?

1 new integrations · 5 existing connectors

Total = value × 2 + risk + sensitivity + effort

Score each candidate 1 to 5; higher is always better. Diagram: Marcus Andrey, CC BY 4.0.

Value counts double because a safe agent nobody needs is wasted work. Adjust the weights with your team: the point is to make the trade-offs visible, as in any Weighted scoring exercise. See AI governance and data protection for the data questions.

5. Steps 4 and 5: choose the form, then sequence

Form Use when Example
Skill The work is interactive: needs Figma or a browser, asks questions, applies fixes Screen review against the design system
Sub-agent The work is heavy reading Interview synthesis
Scheduled agent The work is periodic and only adds what's new A weekly decisions log
Orchestrator (later) Several agents exist and routing between them is itself a chore "Look at the ticket type and start the right step"

Then sequence:

  1. Start with one high-value, low-risk agent that a few designers will use weekly. Early trust matters more than ambition.
  2. Build agents that feed each other: a brief builder feeds a screen review; a decisions log feeds both.
  3. Calibrate every agent against past human work before trusting it.
  4. Package what's worth sharing (see UX agents and skills: a guide for designers).
  5. Revisit the inventory every quarter. The work changes, and so do the tools.

6. Worked example: a fictional team

Inventory (rough shares from a few months of tickets):

Type of work Share Repetitive?
Reviewing developers' builds against Figma about 40% yes
Designing redesigned screens (search, detail dialogs) about 25% partly
Metrics and analytics about 25% no
User research about 10% partly
Decisions buried in weekly meeting notes hidden yes

Candidates and scores (1 to 5, higher is better):

Candidate One line V R S E Total
Design brief builder Given a ticket, reads it, its epic and linked pages, and drafts a brief with conflicts and questions 4 5 4 4 21
Screen reviewer Given a Figma frame, audits it against the design system and suggests fixes 4 4 5 4 21
Build reviewer Given a build task, screenshots the test environment and drafts a design-vs-built table 5 4 4 2 20
Decisions keeper Weekly, adds newly agreed rules from meeting notes to a decisions log, with links 3 4 5 4 19
Interview synthesiser Given a transcript, drafts an anonymised research page 3 3 2 3 14

Sequence:

Phase Agents Why
Now Brief builder, screen reviewer High value, low risk, easy to calibrate
Next Build reviewer The most valuable, but blocked on test-environment access
Next Interview synthesiser Only after the approved tool is cleared for transcripts, with anonymisation built in
Later Decisions keeper Makes the first two cite team decisions, not only library docs

7. Common mistakes

  • Starting with the hardest, most visible problem.
  • Automating work that's rare.
  • Agents that need data you aren't allowed to send to an AI tool.
  • Building something only you can run.
  • Skipping calibration because the first output "looked right".

8. Your roadmap template

# AI agent roadmap: <team name>
## 1. Inventory (last 3–6 months)
| Type of work | Examples | Share | Repetitive? | Craft-heavy? | Supported? |
## 2. Candidates
- <Name>: Given <input>, it reads <sources>, and produces a draft <output> for <who> to approve.
## 3. Scores (1–5, higher is better; total = value×2 + risk + sensitivity + effort)
| Candidate | Value | Risk | Sensitivity | Effort | Total | Notes |
## 4. Form
| Candidate | Skill / sub-agent / scheduled / orchestrator | Needs (connectors, access) |
## 5. Sequence
| Phase | Agents | Why | Blocked on |
## 6. Principles
Read-only by default · cite sources · gaps in the output · anonymise · approved tools only
## 7. Review date
<next quarterly review>

Download: the template as a Markdown file, ready for your wiki.


9. Vocabulary

Term Meaning
Inventory A rough map of what the team actually spends time on
Candidate A piece of work that an agent might take over, written in one line
Weighted scoring Scoring options on several criteria, some counting more
Calibration Comparing an agent's output with past human work before trusting it
Now-Next-Later roadmap A roadmap by phase instead of by date

10. Try it (no code needed)

  1. Fill in the inventory for your team, using real tickets and a quick survey.
  2. Write five candidate one-liners and score them with one or two designers.
  3. Pick the top one, find one past piece of human work to calibrate against, and pilot it for four weeks (see Choosing and piloting AI tools).

Reflection questions: Which of your team's work is frequent, disliked and mostly gathering? What would you say if your manager asked "How would you bring AI into the design team?"

Checklist

Learn more

UX agents and skills: a guide for designers · Guiding designers through AI adoption · Choosing and piloting AI tools · Agentic design · AI governance and data protection · Weighted scoring · Now-Next-Later roadmap · Module 12: AI and the UX Lead

Saved only on this device.

↑ ↓ to move · Enter to open · Esc to close