Precedents: show approvers how similar requests were decided
How DueReason ranks similar past decisions for the same action, what approvers and requesters see, and how agents can read precedents.
Updated · 7 min read
In short
- Before an approver decides, DueReason shows the decided requests of the same action that resemble this one most, with who decided, what they wrote and whether it was approved, rejected or left to expire.
- A requester sees the same panel while typing, with nothing filed yet, so a request that looks like three earlier rejections can be rewritten before it is sent rather than after.
- Similarity compares the action's own payload fields plus the free text of the title and reason; it never adds a field the action does not define and never invents a formula beyond closeness and overlap.
- A precedent can never show a request the viewer could not already open directly, and it only ever informs a decision - nothing in DueReason approves or rejects a request because of what precedents say.
Most approvers answering a request are not the first person to see one like it. A payment to the same vendor, a deploy to the same environment, a publication with the same kind of claim - someone in the organization decided one before, and wrote down why. Precedents put that decision in front of the next approver and the next requester, so the house rule an org already has lives in the record of what it actually did, not just in a policy document nobody reopens.
What the approver sees
Opening a request that has history shows a Similar decisions panel above the decision form: a one-line summary of how many of the decided requests of this action actually resemble this one (and how many were approved, rejected or expired), among how many decided requests of the action exist in the last year, followed by up to the highest-scored similar ones (5 by default, 10 at most). Each one is labelled a close or partial match, shows its outcome, who decided it and when, an excerpt of their written reason (up to 300 characters), any evidence filenames they attached, and which payload fields differ from the request now being decided - with both the earlier value and the current one, not just the field's name, so an approver can tell at a glance whether the difference actually matters (e.g. "Amount: 12 000 then → 4 800 now"). If the most recent SIMILAR one was a rejection, its reason is called out separately above the list, since a recent "no" on the same kind of ask is the single most useful thing to read first; a rejection that has nothing in common with this request is never surfaced that way, even if it is the newest decided request of the action. An approval decided automatically under a mandate is labelled as such rather than attributed to the person who configured the mandate, since nobody looked at that particular request. A request with history but no similar match still says so explicitly ("None of the N decided requests of this action is similar to this one") rather than showing nothing, so an approver can tell the panel ran and found nothing, instead of wondering whether it loaded at all.
What the requester sees while filing
The request form shows the same panel while someone is still typing, before anything is submitted: as they pick the action, fill in its fields and write a reason, a collapsible "Similar past decisions" section updates (debounced, so it does not refetch on every keystroke) against exactly the same decided requests an approver would later see. A requester who is about to re-ask for something that was rejected three times for the same reason sees that before filing, not after - the same console path the getting-started guide walks through for a first request.
What counts as similar
Only decided requests of the same action, in the same organization, are ever compared - a payment is never scored against a deploy, and one org's history never crosses into another's. Within that pool, each candidate is scored against the request (or the draft) on the action's own payload fields, using the schema the org defined for that action, plus the free text of the title and reason:
| Field kind | How it counts |
|---|---|
| Text or choice field (vendor, environment, channel...) | Matches exactly or not at all - an identity field either names the same thing or it does not |
| Yes/no field | Same idea, with only two possible values, so a chance match says a little less |
| Number field (amount, quantity...) | Rewards closeness scaled to the size of the values being compared, not just exact equality |
| Title + reason together | The significant words they share, so similar phrasing scores higher than coincidental overlap |
A field only one side has is skipped entirely - "not stated" is treated as neither a match nor a difference, not as evidence either way. The result is a single score from 0 (nothing alike) to 1 (identical on everything comparable), plus the list of fields that matched and the list that differed. Only candidates scoring at least 0.4 are shown as a precedent at all - sharing nothing but a vague resemblance in wording does not make a past request "similar", it just makes it the same kind of ask, and the panel says so by naming how many of the decided requests actually cleared that bar versus how many exist in total. A precedent scoring 0.75 or higher is labelled a close match; anything from 0.4 up to that is a partial match. For each one, the fields that differ are shown with both values side by side (field label, then the earlier value, then the current one), which is what lets the panel say "Amount: 12 000 then → 4 800 now" instead of just "Differs: amount".
Privacy and visibility
A precedent is never a way to see more than the viewer already could. Reading precedents needs exactly the same approval:read permission as opening the request directly, scoped to the same organization - so a precedent list can never surface a request the caller would be refused if they tried to open it on its own. Someone who has since left the organization still shows up as the decider on an old precedent, labelled "Former member" rather than disappearing from the record.
Using precedents from an agent
An agent reading an approval through the API, the SDK or the MCP server can read its precedents the same way the console does, before or while it asks for one of its own - see the API reference for authentication and the full request/response shape:
- API:
GET /api/v1/approvals/{id}/precedents, with an optional?limit=up to 10, returning the same items and summary the console renders. - SDK (
@duereason/sdk):client.getPrecedents(id, { limit }). - MCP server: the
get_precedentstool, described to the model as informational only - a precedent, however similar, never authorizes anything, and an agent still has to wait for an actual approval before acting on it.
Limits
Precedents only ever compare requests of the same action, only from the last year, and only ones that were actually decided - approved, rejected or expired. A cancelled request does not count: it was withdrawn before anyone decided it, so it carries no reasoning to learn from. And whatever the panel shows, it only informs: nothing in DueReason's approval flow reads a precedent to approve, reject or route a request automatically. Every decision still comes from a person (or an explicit mandate they configured) answering the request in front of them.
Questions
- Do precedents approve anything automatically?
- No. Precedents are shown for the approver and requester to read; nothing in DueReason's approval flow uses them to decide, route or auto-approve a request. Every outcome still comes from a person answering, or from a mandate the org explicitly configured.
- Can an approver see requests they otherwise could not?
- No. Reading precedents needs the same approval:read permission, in the same organization, as opening the request directly - a precedent can never show something the viewer would be refused if they opened it on its own.
- Why is a cancelled request missing?
- A cancelled request was withdrawn, not decided, so it carries no reasoning to learn from. Only approved, rejected and expired requests count as precedents.
- How far back does it look?
- Decided requests of the same action from the last 365 days, newest first, up to 500 candidates scored - of which the 5 highest-scored are shown by default, or up to 10 with a higher limit.
See precedents on your own requests
Create an account, install a template, and the second request you file for an action already has one to compare against.
Create your accountSources
This guide explains the regulation as we read it; it is not legal advice. Check the official text and your national law, which may add obligations.