System of record
A decision log keeps the option you rejected
A decision log is the one record that keeps the options an organization considered and refused. A policy states what applies and says nothing about what was ruled out, so a retrieval assistant reads that silence as permission.
Cognatum Team · Nov 20, 2026 · 6 min read
Someone asks whether the team can use a new vendor. The assistant reads the policy. The policy does not mention that vendor at all.
Cognatum governs the entry
source · version · approver · permissions
So the answer comes back clean. Nothing in here says you cannot.
The vendor came up in the spring. It was looked at closely, and it was turned down.
That got written down. It just did not get written into the policy.
What a decision log is actually for
A decision log is a running record of the choices an organization has made. What was decided, who decided it, when, and on what reasoning.
Most teams keep one somewhere. A spreadsheet, a wiki page, a board register, a column in a project tool. The format varies and the purpose does not.
Project management guidance treats it as an input to audits and as the handover artifact when a project changes hands. Both uses assume somebody will read it later.
Written for people who were not in the room
Lucid Meetings puts the purpose plainly. A decision log makes information available to people who were not present when the decision was made.
It even names the question the log exists to answer, the one that surfaces every few months and costs an afternoon each time. Did we not already decide that?
An assistant is always the reader who was not there
A retrieval assistant has no memory of a meeting. It reads what is written and assembles an answer from the nearest matching text.
Which makes it, permanently and by construction, the reader a decision log was invented to serve. It is also the one reader that never opens it.
The column nobody else keeps
Look at what a decision log template asks for and one field stands out, because no other document in the organization carries it.
Alternatives considered
The Microsoft engineering playbook lays the columns out as decision, date, alternatives considered, reasoning, and who made it. The alternatives get a column of their own.
Monday is more direct about what that column holds. It is the rejected options, recorded with the reason they lost, so nobody repeats the same analysis later.
Why a policy is quiet about what you refused
A policy states what applies. It is not a list of everything that was ruled out on the way to writing it.
So the rejected option leaves no trace in the document an assistant will read. The policy is silent on it, and silence has an obvious reading.
Nothing here prohibits it. That reading is fluent, confident and exactly backwards, because the quiet is not an oversight. It is the shape of a decision.
This is one piece of institutional knowledge the organization deliberately wrote down, and the one piece retrieval cannot use, because a refusal does not look like a rule.
Reasons expire before records do
A document goes out of date when somebody edits the source. A decision goes out of date when the world behind the reasoning moves.
The review trigger admits it
Good decision logs include a review trigger: a date, a milestone, or a threshold that says come back and look at this one again.
That field is an admission in plain sight. The organization knows its own decisions are contingent, and it schedules the moment to check them.
Nobody edits a decision into staleness
The record itself does not change. The vendor gets acquired, the rule is amended, the system it was chosen to work with is retired.
The entry still reads as current because it was never wrong. It was right in a world that has since moved, and no revision marks the difference.
When the decision and the document disagree
The more common failure is not staleness. It is two live sources saying different things, both reachable, both plausible, both apparently in force.
Two answers, one retrieval
A survey of knowledge conflicts in language models names this case directly. Inter-context conflict is what happens when the retrieved sources contradict one another.
The deck from two years ago says one thing. The decision recorded last quarter says another. An assistant reading both has to pick one.
Both are real records. Neither is labelled as the one that governs, because labelling them was a job nobody assigned to either document.
Picking is not retrieving
What it picks on is similarity to the question, or recency, or whichever passage reads more definitively. None of those three is authority.
A system choosing silently between two contradictory sources is not finding an answer. It is making a decision, and keeping no log of its own.
The law logs the system, not the deliberation
Regulators have noticed that AI systems need records. What they have specified is a different kind of record from the one this problem needs.
Events, not reasons
The EU AI Act requires high-risk systems to allow the automatic recording of events over their lifetime. That gives you a log of what the system did.
It does not give you a log of what the organization had already settled, or of whether the answer the system produced respected that.
Provenance is the right shape
The NIST glossary defines provenance as the chronology behind something: where it came from, who has owned it, and how it has changed since.
Applied to a document, that is well understood. Applied to the answer a person acts on, almost nobody has it.
Give the answer the furniture the decision already has
The fix is not another log. It is making the decision reachable from the answer, carrying the same fields the log already keeps.
A person signs off, and that is recorded
In the Cognatum Knowledge Loop, AI runs seven of the eight steps and a person holds Approve. The approver, the date and the source are recorded on the entry.
Every answer served from that entry carries the record back to a dated, attributable sign-off. The chain runs from the answer to a named human.
Conflicts are held apart, not averaged
Where two approved entries disagree, Cognatum surfaces the disagreement rather than settling it by recency. A genuine conflict routes to a named person.
While that flag is open, no contested version is served as approved. The question comes back as a conflict, which is the honest answer to it.
Notification, not quiet rewriting
When a source behind an entry changes, Cognatum flags the entry and routes a revision for approval. It notifies. It does not rewrite the decision for you.
The last approved version keeps serving, with its approval date visible, until a person signs off on the new one. Nothing goes current in silence.
What the log was always trying to do
None of this requires moving your content. Entries stay tied to the systems they came from, and the governance sits above those systems.
A decision log was always an attempt to stop an organization relitigating itself. It just stopped at the log, while the answer kept going.
Your company's knowledge isn't missing. It's unusable. Cognatum changes that. See the Cognatum Knowledge Loop, or book a thirty minute demo.
Sources
- What is a Decision Log? (lucidmeetings.com)
- Decision Log, Engineering Fundamentals Playbook (microsoft.github.io)
- Decision log: build a system for better decision tracking (monday.com)
- How to Use a Decision Log for Optimal Results (projectmanager.com)
- Knowledge Conflicts for LLMs: A Survey (arxiv.org)
- EU AI Act Article 12: Record-Keeping (artificialintelligenceact.eu)
- Provenance, Computer Security Resource Center glossary (csrc.nist.gov)
- The Cognatum Knowledge Loop (cognatum.ai)
- Book a time with Cognatum (cognatum.ai)
Common questions
Questions this raises.
What is a decision log?
A decision log is a running record of the choices an organization has made, capturing what was decided, who decided it, when, the reasoning, and the alternatives that were considered and rejected. Teams keep one in a spreadsheet, a wiki, a board register, or a project tool. Its purpose is to make past decisions legible to people who were not in the room.
How is a decision log different from meeting minutes?
Minutes are a chronological account of a discussion. A decision log is the outcome and the reasoning, pulled out and made findable. Minutes tell you how the conversation went. The log tells you what was settled, by whom, and which options lost. For retrieval the difference matters, because a log is structured and a transcript is not.
Why can an AI assistant not use our decision log?
Two reasons. The log usually sits outside the systems an assistant reads, and even where it is reachable, a rejected option does not read as a rule. A policy that is silent about an option looks permissive to a retrieval system, when that silence may record a deliberate refusal written down somewhere else entirely.
What should happen when a decision and a document disagree?
Neither should be served as the answer until a person resolves it. Where the record shows a dated supersession, the conflict can be reconciled and queued for approval. Where two sources genuinely disagree, they should be held apart and routed to a named approver, rather than settled silently by recency or by similarity to the question.
Does Cognatum replace our decision log?
No. Cognatum governs the entries your people and AI systems answer from, so an approved answer carries its approver, date, source and version. Your decision log stays where it is. What changes is that the reasoning behind a settled question can be reached from the answer, instead of living in a record nothing retrieves.