AI Discovery Workshop: Questionnaire and Stakeholder Map

What you'll learn: how to run discovery for the Lumen Order and Policy Assistant: map the stakeholders, interview them about real past cases, audit the policy documents, time a baseline, and leave with discovery notes and a problem statement the customer signs.

At a glance

Figure 1. The discovery workshop from pre-read to readout

How to read this: interviews come before the group session because people say different things alone, and the readout can send you back to interview again.

• Discovery decides whether to build, and what.
• Draw the stakeholder map first: who uses it, who owns the data, who can say no.
• Interview one person at a time about the last real case, not an imagined future.
• Every answer is a fact, an assumption or a constraint. Write which, and who said it.
• Your artifact: discovery notes and a stakeholder map.

Scenario

Composite scenario, not a real company. Lumen Industrial Supply, a fictional distributor of industrial equipment, has 24 customer-service agents in two regions handling 9,600 contacts a month about orders and policies. The Head of Customer Operations says: "Agents should answer faster and more consistently." Agents look in three places: the order management system (OMS), the policy library and the ticketing system. All numbers in this lesson are illustrative.

You are the Forward Deployed Engineer for two weeks. You write no code.

What you will be able to do

• Draw a stakeholder map with each person's interest, what they can block and the question to ask them.
• Run a 45-minute interview built around a real past case.
• Sort notes into facts, assumptions and constraints, each with an owner.
• Audit a document set and report a baseline with its sample-size caveat.

Prerequisites

Read the seven-step design method: discovery fills steps 1 and 2. For requirements and decision records see AI Architect Mindset.

Core ideas

What discovery produces

• Discovery produces: a defined problem with exclusions, users and their data, constraints marked hard or soft, and a plan to measure success.
• It does not produce: a prototype, a model choice, a promised accuracy or a price.

The GOV.UK Service Manual says "You should not start building your service in discovery", and that stopping at the end of discovery is not a failure.

At the customer: the sponsor asks, "Can you show me something by Friday?" Say: "I can show you a one-page problem statement and the three risks that decide whether this is worth building. A demo of the wrong problem costs more than waiting."

The stakeholder map

Figure 2. Stakeholder map (illustrative)

How to read this: the last column is your interview plan, and "can block" tells you whom to meet before the group session.

STAKEHOLDER MAP (one row per person)
Name / role:
What they care about (in their words):
What they can block, and when:
What I need from them:
Question to ask:
Interview date, notes link:
Open items they own:

The interview guide

Interview one person at a time for 45 minutes: in a group, an agent will not say in front of a team lead that a rule is bent.

INTERVIEW GUIDE (45 minutes)
0-5    Purpose: we are learning, nothing is decided. Ask to take notes.
5-10   Role: what does a normal week look like?
10-30  Last case: walk me through the last time you did this. Open the
       real order, policy and ticket if they let you. What did you search?
       Who did you ask? What did you not trust?
30-38  Pain and workarounds: where does it go wrong? What do you do instead?
38-43  Constraints: what must never happen? Who else must agree?
43-45  Close: who else should I talk to? What did I not ask?

Past behaviour is evidence. A prediction about using a tool is an opinion.

Spot the flaw. Which of these four questions for agents are defective, and why?

1. "Would you use an assistant that answers policy questions?"
2. "Tell me about the last return request that took you more than 15 minutes."
3. "How accurate does it need to be, 95 percent?"
4. "What do you do when you find two versions of the same policy?"

The questionnaire and the triage

Send a questionnaire first so people can bring documents.

DISCOVERY QUESTIONNAIRE
1. Goal: which task should get better, for whom? How long does it take today?
2. Users: who does it, how often, from which tools? What happens when it is wrong?
3. Data: which sources hold the truth, who owns each, how current?
4. Systems: what must it read or write, through which interface? How do users sign in?
5. Risk: who approves a new system that uses company data? What must never be output?
6. Success: how would we measure better? Is there a baseline?

Figure 3. Sorting what you hear

How to read this: only the "yes" branches are safe to build on, and the assumption branch is your follow-up list.

Worked example

Assumptions: all numbers illustrative; "agent" means a customer-service agent, not an AI agent.

• The contact mix. 5,280 order contacts (55 percent) + 3,360 policy contacts (35 percent) + 960 other (10 percent) = 9,600. The first release targets order and policy contacts: 5,280 + 3,360 = 8,640, or 90 percent.

• The policy-corpus audit. The library holds 62 documents. Nine exist in more than one version: 9 / 62 = 14.5 percent. An old version is still searchable on the intranet. Policy P-RET-01 shows why it matters: version 3 accepts returns within 30 days of delivery, superseded version 2 said 14. For an order delivered 20 days ago, v3 accepts the return and v2 refuses it.

• The baseline. Six agents each timed 4 real contacts: 6 x 4 = 24 timings, median 9 minutes. State the limit with the number: 24 timings are 0.25 percent of one month's 9,600 contacts, they come from 6 of 24 agents (25 percent), and four timings from one person are not independent.

The problem statement from the readout: "Agents handle 8,640 order and policy contacts a month, with an estimated median of 9 minutes (24 timings from 6 agents). Lumen wants consistent, cited answers from current policy only." The recommendation is narrow: order and policy contacts, no write actions.

Practice task

You have notes from two interviews. Write a stakeholder-map row for each person you met or must still meet, and rank your next five questions.

INTERVIEW 1 (agent):
- "I search the intranet and take the first return policy I find."
- "Above 5 percent discount I wait for my team lead, sometimes an hour."
- "I never change an address after the order is picked."
INTERVIEW 2 (Head of Customer Operations):
- Wants handling time down and answers consistent across both regions.
- "IT will be fine with it, it is just search."

Mark each sentence fact, assumption, constraint or a person to meet. "IT will be fine" is an assumption about someone nobody asked, so the IT security lead is a first target. The agent's first sentence points to the policy owner. Keep the result as your discovery notes and stakeholder map.

Decide first, then reveal.

Exercise

Write a 45-minute interview guide for the policy owner (Commercial).

Acceptance criteria:

• At least 8 questions, none naming a solution or asking for a prediction.
• At least 3 ask about a specific past event or a real document.
• At least 1 surfaces documents that exist in several versions, and 1 asks who decides when a version is superseded.
• A closing question asks who else to talk to.

Done when each question passes the spot-the-flaw test and a colleague could run the interview without you.

Common mistakes

• Starting with the solution. "We will build a chatbot" colours every answer.
• Interviewing only the sponsor. Meet the agents too.
• Meeting security last. A reviewer who hears late can stop the project.
• Writing every claim as a requirement. Mark assumptions as such.

Knowledge check

You can now

• Add discovery notes and a stakeholder map to your design pack.
• Run a 45-minute interview centred on a real past case.
• Rewrite a leading question as a question about a past event.
• Triage notes into facts, assumptions and constraints with owners.
• Report a baseline together with its sample-size caveat.

Further reading

• GOV.UK Service Manual: How the discovery phase works, on defining the problem, hard constraints, measuring success and stopping as an allowed outcome.
• GOV.UK Service Manual: How the alpha phase works, on what follows discovery and how it ends in a decision.
• AI Architect Mindset, for requirements and decision records.

Last verified: 2026-11-03

Next

From a vague request to measurable outcomes