The only person who understands the process is leaving in three weeks and nothing is written down. How do you get what you need?

What you are actually trying to save

Three weeks is not enough to document a process properly, so any answer that starts with a plan to document the process properly is the wrong answer. The first move is to work out what is genuinely irrecoverable once this person walks out, because the categories differ enormously in how expensive they are to reconstruct.

The routine path is usually the cheapest thing to recover. It is visible in the system, other people have seen parts of it, and if you lose it you can rebuild it in a fortnight of watching whoever inherits the work struggle. Spending your three weeks on a beautiful swimlane of the happy path is the most common way this goes wrong, because it produces the most impressive artefact and saves the least.

What is expensive is judgement: the rules that decide which exception path applies, the tolerances they apply without thinking, the reasons behind decisions that look arbitrary, and the informal network they use when something breaks. A colleague at a supplier who will fix a mismatched reference by phone is not in any system and cannot be rediscovered. Neither can the reason a particular customer group is always processed manually, which is usually a decision taken years ago for a reason that is still valid.

So triage explicitly, and say the criterion out loud: what would cost the most to work out again after they have gone. That single sentence usually earns the round.

Watch the work, do not only ask about it

Interviewing someone about a process they have performed for eleven years produces a sanitised summary, because expertise has become automatic and automatic steps are invisible to the person performing them. They will not mention the spreadsheet on their desktop, the check they run before the check that is written down, or the four cases they route differently, because none of that registers as part of the process any more.

Sit beside them while they do a real day's work and narrate. Ask why at every pause, every alt-tab, every second look at a screen. The most productive question available is not "what do you do" but "when did that last go wrong, and what did you do about it". A war story contains the exception, the decision rule and the escalation route in one telling, which is three of the four things you are trying to capture.

Also collect their artefacts on the first day rather than the last. The personal spreadsheet, the saved email folder, the printed crib sheet in the drawer, the browser bookmarks. These are the actual process documentation and they disappear with the laptop. Copy them before you understand them.

Verify against the system

Anything you are told is a memory, and memories under time pressure round towards the typical. Check the important claims against records, because the discrepancies are where the real process lives.

A worked case. They tell you that around 5 per cent of applications need manual review. You pull twelve months of records: 24,000 applications, of which 4,100 have a manual review flag. That is 17 per cent, more than triple what you were told. Take that back to them, and the answer is usually revealing rather than embarrassing — the flag is set for two different reasons, one of which is a batch process that marks everything from a particular channel and which they do not think of as review at all. You have just discovered a rule that was never going to come up in any interview, and you found it in an afternoon with a query. Those figures are illustrative, but the method is the point: ask for the number, then look at the number.

The same check works on timing, on volumes by exception type, and on who actually performs the steps. Systems record who clicked, and it is frequently not who the process says.

Prioritise inside the three weeks

A defensible plan for the time you have, with the reasoning attached rather than a schedule for its own sake.

Days	Focus	Why it comes first
1 to 2	Collect their artefacts; list every exception type they can name; map the interfaces to other teams and suppliers	Cheapest, and it defines everything else
3 to 8	Observe real work, including a month-end or whatever the busiest cycle is	Only obtainable while they are here
9 to 14	Verify claims against system records; interview the two or three people at the other end of the handoffs	Finds the contradictions while there is still time to ask
15 to 21	Write the successor's guide; walk the successor through it with the leaver present	Tests the capture rather than assuming it

That last row matters more than the rest. The only real test of captured knowledge is whether someone else can act on it, so get the successor doing the work while the leaver is still there to correct them. Two days of supervised handover finds more gaps than a fortnight of interviews, and it converts your document from a claim into something verified.

Write it for the person who has to do the job

The artefact should be organised by the situations the successor will face, not by process phase. A section per exception type, each one saying how to recognise it, what to do, who to contact, and what happens if you get it wrong. Named people with contact details, including the ones outside your organisation. A list of the decisions that appear arbitrary with the reason attached, because a successor who does not know the reason will optimise it away. And an explicit list of the questions the leaver could not answer, because a known gap is manageable and a silent one is not.

Keep the outstanding-questions list in the handover document rather than tidying it away. It is the most valuable page, it tells whoever inherits the risk exactly where the thin ice is, and it protects you when something surfaces in November.

Where three weeks quietly become nothing

Two failures do most of the damage here. The first is spending the time on breadth: eleven interviews across the whole process, a complete model of the routine path, and no time on the exceptions, so the capture looks thorough and covers only the part that was never at risk. The second is treating this as a documentation exercise rather than a risk exercise, and therefore never telling anybody what remains unmitigated.

The organisational move that goes with the analysis is to name the residual risk in writing while there is still time to act on it. If the honest position is that four exception types remain undocumented because they occur quarterly and none will occur in the next three weeks, then somebody senior can decide whether to extend the leaver on a short consultancy arrangement, delay a related change, or accept the exposure. All three are reasonable. Discovering it in the quarter-end after they have gone is not.

Spend the three weeks on what cannot be reconstructed: the exceptions, the judgement and the informal contacts. The routine path is recoverable from the system, and it is the part that will tempt you to spend the whole three weeks on it.