Eliciting requirements from people and documents — questions

The blind-spot column is the reason nobody uses one technique alone: each is blind exactly where another one sees. An interview cannot tell you what people do without noticing they do it; observation cannot tell you why; a document cannot be asked anything at all; a questionnaire cannot follow up.

They also differ in what kind of statement they produce. Interview and observation are qualitative — they find out what is going on. The questionnaire is quantitative — it says how widespread it is. Document analysis is the only one that yields structure: field names, types, ranges.

Points the answer must contain:

Before: read the documents first, write down what you want to know at the end of the hour, turn that into six to ten questions, and decide who has to be in the room. An hour is the maximum that stays productive.

During: ask open questions and follow what turns up. The semi-structured form — prepared questions, free follow-ups — is the normal case.

The left column is suggestive: it offers the answer, and a polite person agrees. What you take home is then your own idea with somebody else’s name on it — which is worse than having no answer, because it looks like evidence.

Afterwards: write the protocol the same day — after two days you remember your interpretation rather than their sentences — and send it back for confirmation. The corrections that come back are usually the most valuable sentences of the whole exercise, because they are the points where you misunderstood without noticing.

A recording needs consent, and personal data from the interview stays out of the repository and out of any prompt.

Points the answer must contain:

It delivers structure, and it delivers it before you have spoken to anybody: field names, data types, value ranges, mandatory fields, volumes, the official vocabulary, and the rules that are baked into an Excel formula which nobody in the organisation can state out loud any more.

The procedure: collect and catalogue (what exists, who owns it, from when, still in use?); extract the fields — every column is a candidate attribute with name, type, range and example values, which is the first free draft of your data model; build the glossary in the client’s words, because those are the words the acceptance will use; note the contradictions — a column that is always empty, two forms with different fields for the same thing — as questions for the interview; and record what stayed unclear as an assumption with its source.

Two further advantages are structural. It costs nobody else’s time, so it can be done before any appointment exists. And it has no Hawthorne effect: a form does not behave differently because you are reading it.

Its limit is equally clear: a document shows the rule, never the exception somebody makes every Friday, and it cannot be asked why. That is precisely the division of labour with the interview — which is why the analysis comes first and supplies the questions.

Use anonymised copies or the structure only; real documents carry personal data that you do not need and should not hold.

Points the answer must contain:

Document analysis first, because it is the only one that costs nobody else anything, and because it makes every later question better. Walking into an interview already knowing the fields, the vocabulary and the volumes means the hour is spent on what only a person can answer.

Interview second, for the reasons, the exceptions and the contradictions the documents produced.

Observation where the description and the reality are likely to differ — that is, wherever people have worked around the official process.

Questionnaire last, to find out how widespread what you found actually is. First it would ask the wrong questions very efficiently.

The order is a default, not a law. If there are no documents — a process that exists only in people’s heads — you start with an interview and build the field list from it. What stays fixed is the reasoning: use the cheapest technique that can still answer the question you have now.

Points the answer must contain:

"The trainers say the paper list works fine, but three of them keep a private Excel sheet." — Beobachtung. The statement and the behaviour disagree, and only watching resolves that. An interview will get the description again, because people do not experience their workarounds as part of the process. Stay long enough for the first twenty minutes of best behaviour to wear off.

"We need to know how many of the 400 members would use online booking, and how often." — Fragebogen. It is a question about quantity across a population that is far too big to interview, the answers are closed (yes/no, a number, a scale), and no follow-up is needed. Pretest it with three people first.

The justification always has the same shape: name what you need to know, then name the technique whose output is that kind of statement — reasons, numbers, behaviour, or structure — and check that its blind spot is not exactly your question.

Points the answer must contain:

Three artefacts, all of them in the repository:

The third is the one teams skip and the cheapest insurance in the project. An assumption that is written down can be corrected in a five-minute conversation ("we assumed a member belongs to exactly one group — is that true?"). The same assumption held silently is discovered at the acceptance, when it has a data model built on top of it.

The check that elicitation is actually done: every goal in the charter can be traced back to something somebody said, did, or wrote down — and nothing in it came from the team’s imagination alone.

Points the answer must contain: