Feature: Agents That Suggest, and a Human Who Decides

Sapeum's mapping agent is tuned to capture everything and invent nothing, and everything it proposes beyond the map arrives as a suggestion card a person accepts or rejects.

The Sapeum team

3 min read

The instinct in 2026 is to let the agent write. For process mapping that instinct is wrong, and the reason is specific rather than philosophical. An invented step in a compliance map is worse than a missing one. A gap is visible to the person reviewing it. A plausible step that nobody actually performs is not, and it will be signed off, and it will be what the auditor reads.

So the rule the mapping agent is tuned against is capture everything, invent nothing. That is a real constraint with a real cost, and Sapeum measures it. Prompt changes go through a batch of scored tasks across several evaluation templates, graded on around twenty dimensions: whether relevant content was discarded, whether invented content appeared, whether interview questions were wrongly treated as steps, whether existing steps were updated in place instead of duplicated. The guidance now in production was chosen off that scorecard rather than off a demo.

The suggestions panel on version 2 of a medical device manufacturing process. Above it, a summary of the changes just made. Below, suggestion cards each carrying a type label, a title, a description and a Benefit line, with a tick box beside each. One card, labelled Improvement, reads "Add triage decision before QC audit" and explains that it prevents low-severity defects from entering the most expensive remediation path.

What the change is, and what it buys you. The label says whether the agent thinks it is an improvement or a cleanup.

What the agent produces beyond the map itself arrives as suggestions rather than edits. Each card is labelled with the kind of change it is, an improvement or a cleanup, and then states what the change is and what it buys you under a Benefit line, so it can be read in about two seconds and accepted or rejected in one click. Cards can be ticked and actioned together. The agent is also told what it has already suggested, so it will not re-offer an idea you rejected, and it can follow a reference back to an earlier card when you say "that one about the handoff" instead of treating it as a new instruction.

A Gate Review step on the canvas outlined in orange and labelled SUGGESTION, with a red reject button and a green accept button floating beside it. A tooltip above shows the proposed field change, from empty to "Grouped with the delivery phase as the completion gate before launch validation."

A proposal sitting on the map, showing exactly what it would change, until somebody clicks one of the two buttons.

The pattern worth noticing: the interface is doing the safety work. An agent that proposes into a review queue can be wrong cheaply. An agent that writes directly into a document somebody will sign cannot. Everything the agent contributes lands in the same Approvals sidebar as everything a colleague contributes, with the same approve and reject controls, which is the only arrangement most process owners will accept.

Sapeum lets an agent do the tedious half of process mapping while a human keeps the decisions. Try Sapeum.

Share

Keep reading

More from the blog

All posts

See it on your own process

Bring a workflow to a 30-minute session and watch it become a living process map — captured, redesigned, and ready to govern.

Request a demo