AI Library "Exploration"

Step

Briefing

Explanations

"Read the charter"

Six fields: Motivation, Playground, Scope, Questions, Resources & Stakeholders, Schedule. The Questions are the sponsor's guiding questions only. If something essential is missing, ask once, specifically.

Without this step AI may fill gaps confidently. Where the charter is thin, the pull is to invent something plausible rather than to flag the hole. The charter is the only real context AI has — one precise question costs less than a fluent fabrication nobody notices.

"Break the playground apart"

Name both components and say where the uncertainty sits:

1. Opportunity or challenge — new technology, changed regulation, societal trend (you act), or competitive pressure, scarce resources, inefficiency (you react).

2. Element of the Picture of the Future it bears on — a customer segment, an offering, a step in the value chain, a capability.

Left alone, AI would answer the topic in general ("customer loyalty") instead of this specific framed field. Naming the two components keeps every later question anchored to the playground, and locating the uncertainty is what makes focus, scope and priorities derivable instead of arbitrary.

"Set focus and scope"

Focus (which areas): technological, organisational, market-driven.

Scope (how broad or deep): take it from the charter if it is there, and only check it fits the guiding questions. If missing, propose one and ask back; when signals conflict, choose T-shaped.

State in one sentence what this scope does not deliver. What does not fit becomes a follow-up exploration, not a bigger project.

Scanning:

• several research areas touched (4–8)

• 2-3 questions per area

• mostly desk research, a few interviews, some expert talks

T-shaped:

• 1 deep research area + 2–3 areas touched

• 5–9 questions (deep) and 2–3 questions (broad)

• fieldwork in the deep area, desk research elsewhere

Deep-dive:

• 1 deep research area

• 5-9 questions

• real fieldwork, ~15 conversations

AI has no sense of cost. Nothing in generating text tells you that fifteen interviews fill six weeks, so without the scope it will write an excellent plan that needs six months. The scope is also the antidote to its pull towards completeness: covering everything a little is a decision with a price, not a free bonus.

"Prioritise the research areas"

Choose from: Customers & Users · Market Structure & Dynamics · Technology · Competitive Landscape · Rules & Regulations · Data & Resources · Today's Business Model · Practices & Processes.

For a market focus: customers first, then competitors, then market. Name the areas you leave out with half a sentence each.

Given a free hand AI invents fresh categories every run, so two plans never compare. A fixed list of eight makes outputs comparable across projects and teams. Naming what it leaves out matters just as much — otherwise omissions look like oversights rather than choices, and nobody can contradict them.

"Write the questions"

Number consecutively. Prefer many concrete questions over few general ones.

• aim at today's real behaviour and today's state, not opinions or a wished-for future

• ask about behaviour, needs and emotions, not only pain points — and on loyalty topics also ask directly what a customer values and what puts them off

• separate buying and using in B2B

• stay open enough to let something surprising through

• park solution questions ("What should we offer?", "Should we open up to Y?") and say which research questions will feed them later

AI defaults run the wrong way here. AI tends towards questions about opinions and wishes ("What would customers want?") because they are easy to phrase, and it slides into solution mode because proposing is what AI is good at. Both produce material that cannot be researched. The numbering is the thread the second pass depends on.

"Produce the output"

Open with a visible note on the draft's limits: this comes from an AI that knows only the charter — not the history, the failed attempts, the people, or what is already answered internally; it cannot judge access or feasibility; it writes what is plausible, and the plausible is rarely surprising; the judgement sits with the team; this is a proposal to argue with, not to sign off.

Then three parts:

1. Playground & focus — both components with uncertainty, then focus and scope with reasoning and what the scope does not deliver.

2. Research areas & questions — numbered per area; areas left out with half a sentence each.

3. Revision brief — tell the team to cut what they already know (name two or three candidates yourself), add what only they can know, make questions more concrete (give an example from your draft), argue with the prioritisation (name where you are least certain), and check whether these questions can find something surprising in six weeks. Close with the decisions they can overturn: focus, scope and its origin, areas left out, parked solution questions.

AI output reads more authoritative than its substance warrants — fluent, structured, evenly confident whether well founded or guessed. Without the limits note and a concrete revision brief, a draft becomes a decision simply because it looks finished. The structure also lets the team find the parts it disagrees with, instead of reacting to the whole.

"Stop"

One sentence on what happens next — the revised questions come back and the approach is developed from them. Then stop.

AI is built to be maximally helpful, so the strongest pull in this whole briefing is to deliver the work packages too — unasked, "just as a first idea". That single move hands the team a finished plan and takes the decision away from them. Stopping is the instruction most likely to be broken, which is why it is a step of its own.

Step

Briefing

Explanations

"Take the revised questions"

Use the team's wording verbatim. Renumber if questions were cut or added, and say in one sentence what changed against your draft. Name real contradictions — two questions meaning the same thing, a question nobody can answer — instead of quietly resolving them.

AI is drawn to improve text you are given. Rewording the team's questions looks like polish and is in fact a takeover: they just spent a workshop deciding those formulations. Silently merging two overlapping questions also hides a disagreement the team may need to have.

"Bundle the questions into work packages"

One method sometimes answers several questions. Cluster by: which questions can be answered through the same access, the same occasion, the same method? No fixed number — roughly: Deep-dive 2–4, T-shaped 3–5, Scanning 4–6.

Each package: title (the activity, not the topic) · method(s) · answers (question numbers, gladly across areas) · procedure (one or two sentences: with whom, what, which sources) · scale · documentation format.

A question-by-question table is the tidier structure, so that is what AI may produce by default — and it is useless in the field, because nobody runs, e.g. one interview per question. Bundling also forces AI to think about who is actually being approached, which is where plans break.

"Apply the cutting rules"

One package = one access. Different counterparts means different packages, even inside the same research area.

1. Packages crossing research areas are a quality mark, not an error.

2. Check access before planning a method. Observation is often unrealistic in regulated environments. Ask when unsure.

3. Every question must appear in at least one package.

4. With T-shaped and deep-dive, the analysis is its own package — attributes, segments, representatives, development paths. It costs real days.

5. Access to non-customers is a work step. Five to eight approaches per acceptance; start in week one; say it must not be the first thing dropped.

AI cannot see a door. Observation, site visits and "talk to ten non-customers" cost you nothing to write and can be impossible to arrange, so AI plans them freely. AI also under-plan analysis, because synthesis feels like something that happens by itself — for AI it does, for a team of four it is e.g. two sessions.

"Set scale and documentation format"

Interviews: one person holds the conversation, recorded with consent, AI-transcribed — no second note-taker. Post-processing into structured form takes about 60 minutes per conversation, so one interview costs two to two and a half hours all in. At one day per week per person that is about two a week.

Documentation: with interview packages the structured interview documentation comes first. Insights matrices, personas and segment matrices are condensations on top, and optional at scanning scope.

Numbers in generated text carry no cost, so "15 interviews" and "5 interviews" feel equally cheap to write. An explicit rate is the only thing that keeps a plan inside six weeks. And naming a persona as the deliverable skips the step it is derived from — the documentation is the working result; the polished artefact is not.

"Triangulate the important questions"

Have the most important questions answered by two packages from different angles — talking about it, watching it, doing it yourself.

AI treats a question as closed once it has been assigned somewhere, because coverage is what a checklist rewards. But the gap between what people say and what the data shows is usually the actual finding, and it only appears if the same question is asked twice by different means.

"Produce the output and stop"

Open with the limits note: this comes from an AI that knows only the charter and the questions — not the history, the people, or what access is realistic; it writes what is plausible, and the plausible is rarely surprising; the judgement sits with the team.

Then four parts: 1 Playground and focus, carried over · 2 the revised questions per area, with areas left out · 3 the work packages, most important first · 4 review point, triangulation and bias, what is left out with the bet behind it and parked questions.

A fluent, complete plan reads as authoritative whether it rests on the charter or on your guesswork, and the parts you invented look exactly like the parts you were told. Saying what AI left out — and what AI bet on by leaving it out — is what lets the team find the places to argue with, instead of accepting the whole because it looks finished.