Discovery

The idea in one minute#

Discovery is the work of finding out what the problem actually is before proposing anything. People arrive with a solution — “we want a chatbot”, “we need agents” — and the architect’s first task is to walk backwards from that solution to the problem behind it, the people who have it, the numbers that describe it and the data that exists.

Good discovery produces one page: a problem statement with a baseline measured in real units — hours, days, dollars, errors — and a list of what is known, what is assumed and what must still be found out.

It is the cheapest phase of a project and the one that decides most of the outcome. An hour spent watching someone do the job teaches more than a week of workshops about what they say they do.

A picture#

flowchart LR
  ASK[":i-message-square: <b>The request</b><br/><small>“we need a chatbot”</small>"] --> WHY[":i-search: <b>Walk back to the problem</b><br/><small>why, for whom, how often</small>"]
  WHY --> OBS[":i-eye: <b>Watch the work</b><br/><small>sit with the people who do it</small>"]
  OBS --> NUM[":i-gauge: <b>Baseline numbers</b><br/><small>volume, time, cost, error rate</small>"]
  NUM --> DATA[":i-database: <b>Data reality check</b><br/><small>exists, reachable, allowed, good</small>"]
  DATA --> PEOPLE[":i-users: <b>Stakeholder map</b><br/><small>sponsor, users, blockers</small>"]
  PEOPLE --> PS[":i-file-text: <b>Problem statement</b><br/><small>one page, with unknowns listed</small>"]
  class ASK neutral
  class WHY,OBS queue
  class NUM,DATA memory
  class PEOPLE warn
  class PS io

How it really works#

Walk back from the solution#

When someone asks for a technology, ask “what would that let you do?” until you reach a sentence with no technology in it.

"We need an AI agent for claims."
   What would that let you do?
"Process claims faster."
   What is slow today?
"Handlers spend most of the day reading documents to find a few facts."
   How many handlers, how many claims, how long each?
"120 handlers, about 900 claims a day, around 25 minutes of reading per claim."
   What happens because it is slow?
"Customers wait nine days for a decision. Our target is five."

The last two answers are the problem. Everything before them was a guess at a solution.

The five things to find out#

Watch the work#

Ask to sit beside two or three people doing the task, and write down what they actually do. You are looking for four things that never appear in a process document:

• The tools they really use — the spreadsheet beside the official system.
• The judgement calls — where two experienced people would decide differently.
• The exceptions — the one case in ten that takes half the time.
• What they would never trust a machine with, and why. That list is your risk register.

The data reality check#

For each input the solution would need, answer five questions. One “no” is a project inside your project.

The last row is the one beginners forget. Without examples that have known correct outcomes you cannot evaluate anything, and a solution that cannot be evaluated cannot be approved.

The stakeholder map#

Write one line per person or group. Do it in the first week and keep it updated.

Two rules. Talk to the people with a veto early, when they can shape the design, rather than late, when they can only refuse it. And find out who loses something if the project succeeds — they will not say so in the workshop.

Questions that produce numbers#

Vague questions get adjectives. These get figures.

• “How many of these did the team handle last month?”
• “Take the last one you did. How long did it take, start to finish?”
• “Out of ten, how many are simple and how many are awkward?”
• “When one goes wrong, what happens, and what does it cost?”
• “If this took half the time, what would you do with the other half?”
• “What would have to be true for you to trust the result without checking it?”

The last question returns in Requirements and Success Metrics: the answer is the quality bar.

Signals that it is not an AI problem#

Discovery often ends with “do not use a model”. Say so plainly; it earns more trust than any design.

The problem statement#

One page. If it does not fit, you do not understand it yet.

PROBLEM STATEMENT — Claims document review              v0.3, 13 Oct 2026

Who          120 claims handlers, property line
Task         Read submitted documents to extract the facts needed for a decision
Today        ~900 claims/day · ~25 min reading per claim · 9 days to decision
Target       Decision within 5 days (regulatory service standard)
Cost today   ~375 handler-hours/day spent reading
Why now      Volume up 18% year on year; hiring freeze

Known        Documents are PDFs and photos in the document platform (API exists)
             Two years of decided claims available as examples
Assumed      Extracted facts can be checked by the handler in under 3 minutes
             Customers consent covers automated extraction (legal to confirm)
Unknown      Share of documents that are handwritten or photographed
             Error cost: how often does a missed fact change the decision?

Out of scope The decision itself. A person decides every claim.
Not AI       Missing-document chasing — fixed by a form change, not a model

Next         Sample 200 claims for document types · Legal review · Shadow 3 handlers

Notice three sections people leave out: Assumed, Unknown and Not AI. They are what make the page honest, and they are what the next two weeks of work come from.

Remember this#

• Walk back from the requested technology to a sentence with no technology in it.
• Get the baseline in real units: volume, time, cost, error rate.
• Five data questions: exists, reachable, allowed, good enough, examples with right answers.
• Meet the people with a veto in week one.
• “This is not an AI problem” is a valid and valuable finding.

Try it#

1. Choose a task you or a colleague do every week. Run the walk-back until you reach a sentence with numbers and no technology.
2. Answer the five data questions for that task. Which is the weakest?
3. Write the one-page problem statement, including Assumed, Unknown and Not AI.

Check yourself#

1. Why watch people work instead of reading the process document?
2. Which data question is most often forgotten, and what can you not do without it?
3. Give three findings that mean the answer is not a language model.
4. Why speak to security and legal in the first week?