In short: A good judgment system does not ask people to tell it everything. It knows what the next decision will need, notices what it is missing, and asks for exactly that, of the person most likely to know.
Key points
- "Summarise the meeting" prompts produce long, generic text that nobody reads.
- Write a needs template per decision class and compare each record against it to get a gap list.
- Rank gaps by decision dependence, staleness and cost to answer, then ask only the top one.
- "I don't know" is a valid answer that flags an item for verification.
- Gaps that recur across many records usually point to a missing schema field or data connection.
The default capture prompt is "summarise the meeting". It produces long, generic text in which the one thing that mattered is buried or absent. People learn quickly that nobody reads the summary, and they stop writing good ones.
The better pattern is gap-seeking, and it is more tractable to build than it sounds.
How to compute a gap
Every decision class has an implicit list of what a good decision needs. A renewal needs the budget holder, the usage trend, the competitive threat and the relationship temperature. A credit review needs recent financials, loan-condition status, ownership changes and the sector view. Make that list explicit as a template for each class.
Then, after any interaction touching a live decision, compare the record's current state with the template. The difference is the gap list.
Rank gaps by three things: how much the next decision depends on the item, how stale the current information is, and how cheap it is to answer. Ask only the top one. A single, sharp question is answered in seconds. Five questions feel like an interrogation and get skipped.
Good gap questions are specific and answerable
- "Did the budget holder change after the reorganisation?"
- "Was the objection about price, or about timing?"
- "Is the March waiver of the loan condition still in force?"
Each can be answered in a few words. "I don't know" is a valid answer and itself useful: it marks an item the next person should verify, not assume.
Route the question to whoever knows
Gap-seeking does not have to ask the person who just finished the interaction. The system often knows who touched the account last, who owns the relationship, or who analysed the sector. Route the question to the person most likely to answer it well. This turns the system from a form into something closer to a well-informed colleague who knows who to ask.
Gaps that recur are schema problems
Track which gaps appear most often across records. A gap that recurs across many decisions is usually a field the schema lacks, or a source the system is not connected to. Recurring gaps are the cheapest product research you will ever get, because the system is telling you, in aggregate, what it does not know.
Mind the tone
A system that only ever asks what is missing can feel like an auditor. Pair every gap question with a give: the precedent, the warning, the drafted brief. And never ask the same person more than one gap question per interaction.
In practice
- Write a needs template for one decision class: the six to ten things a good decision in that class requires.
- After each interaction, compute the gap list and ask the single highest-ranked question.
- Accept "don't know" and flag the item for verification.
- Review the most frequent gaps monthly and turn the top recurring one into a field or a data connection.
What would prove this wrong. Records built from free-form summaries proving as complete and as useful for later decisions as records built from targeted gap questions, at similar cost in time. Then the summary is good enough.
Common questions
What is gap-seeking in a knowledge or decision system?
Instead of asking for a summary, the system compares a decision record with a template of what that decision class needs and asks for the single most important missing item.
How many follow-up questions should a system ask after a meeting?
One. A single sharp question is answered in seconds, while five feel like an interrogation and get skipped. Never ask the same person more than one gap question per interaction.
Who should a gap question go to?
The person most likely to answer it well, such as whoever last touched the account, owns the relationship or analysed the sector, not necessarily the person who just finished the interaction.
What do recurring gaps in records tell you?
A gap that recurs across many decisions usually means the schema lacks a field or the system is not connected to a source. Review the most frequent gaps monthly and turn the top one into a field or data connection.