Skip to content

16 November 2026, 7 min read

When AI and people disagree: provenance, proposals and keeping company facts true

AI agents and syncs will contradict what people wrote. Provenance and proposals keep company facts true without overwriting people.

Sooner or later an AI agent will tell your company something that conflicts with what a person wrote. A sync from the HR system will say Dana manages the platform team. The team's own page says Marcus does. An assistant reading a meeting transcript will conclude the launch moved to March. The project owner entered April and has not changed it.

How a system handles that moment decides whether company facts stay true. There are really only two answers. Someone, or something, wins by default, or the disagreement is made visible and a person decides. This post argues for the second, and describes how to build it.

Why this problem is new

Wikis never had to settle this. Pages were edited by people, and conflicts were social: you saw the other edit, messaged the author and agreed. The last edit won and history recorded the rest.

Software changes three things.

Volume. An agent can propose hundreds of changes an hour. No colleague can read that much history.

Confidence. A machine states a wrong value as fluently as a right one. Nothing in its tone says "I am guessing".

Plausible sources. A sync from another system looks authoritative, even when the other system is the one that is out of date.

When writes arrive at that rate with that polish, "last edit wins" becomes "whoever writes last, wins", and the last writer is increasingly a machine.

Three ways to resolve a conflict

Policy How it works What goes wrong
Last write wins The newest value replaces the old A confident agent erases a correct human entry
Machine never writes Agents are read-only Safe, but humans end up doing every routine update
Proposal The new value is held with its source until the owner decides Needs a review queue, which must stay short

Most systems start with the first, because it is the default of every database. A few overcorrect to the second, which is safe but gives up much of the benefit of having agents. The third is more work and, we think, correct for facts people depend on.

The question is not whether AI will be wrong. It is whether being wrong is cheap to catch and cheap to undo.

What provenance means

A proposal only helps if the reviewer can judge it. For that they need provenance. The W3C defines provenance as information about the entities, activities and people involved in producing a piece of data, which can be used to assess how good, dependable or trustworthy it is.

For a company fact, a useful provenance record answers four questions:

  1. Who or what wrote this value? A named person, a named agent acting for a named person, or a sync from a named source.
  2. When? The time of the write, and ideally when the source last confirmed it.
  3. From where? A link to the transcript, document, row or ticket it came from.
  4. What did it replace? The previous value, so the reviewer can compare.

With those four, a reviewer can look at "Dana manages platform, from the HR sync on Tuesday, replacing Marcus" and decide in seconds. Without them, the same reviewer has to investigate.

Provenance also helps downstream. When an AI tool answers a question, it can cite where the value came from and flag if the value is an unreviewed proposal rather than a confirmed fact. A fact that says "owner: Dana, entered by Dana" deserves more trust than one that says "owner: Dana, suggested by an agent, pending".

How a proposal flow should work

The design principles are short.

The person's value stays until a person decides. The record keeps showing the original, labelled with who wrote it. The proposal appears alongside, not instead.

Proposals go to the owner. Every record needs an owner, or the queue has nobody to route to. For a person record, that may be the person, their manager or HR, depending on the field.

Only conflicts need review. If an agent fills in a blank field, or confirms what is already there, no human attention is required. Reserve the queue for cases where machine and person disagree.

Accept, edit or reject, and say why. Rejecting with a reason is valuable, because it tells the source what to fix. A sync that keeps proposing the same wrong value is a bug worth surfacing.

Rejections stick. If the owner has rejected a value, do not re-propose the same one every night. Show it as previously rejected.

Agents do not approve themselves. An agent that can propose must not be able to approve. The approval path should not appear in an agent's tool list at all.

A worked example

Take the Projects collection. Each project has an owner, a status and a launch date. Priya, the owner of the billing migration, set the launch date to 14 April.

On Monday an agent reads a meeting transcript and concludes the launch has moved to 12 May. Under "last write wins", the field now says 12 May. Priya is on leave. The finance team plans around the new date. Nobody remembers that the transcript was about a different workstream.

Under a proposal model, the same event looks different. The record still shows 14 April, with Priya's name beside it. A small marker says an agent suggested 12 May, quoting the sentence it relied on and linking to the transcript. When Priya returns, she sees the marker, reads the sentence and rejects it, noting that it was about the reporting workstream. Or she accepts it, and the date changes with her name attached to the decision.

Either outcome is fine. What matters is that nobody planned around a value that no accountable person had confirmed.

The cost of the proposal model is one extra click on a real conflict. The cost of the other model is a wrong date that looks right. Most teams, asked directly, take the click.

Sharing is a fact too

The same logic applies to who can see what. An AI suggestion to widen access, for example "this page looks like it is meant for the whole company", is a useful nudge, and a risky automatic action. Treat it like any other disagreement: show exactly what would change, name who would gain access and have a person approve it. If the preview is vague, the approval is meaningless.

Sensitive fields need a further rule. Administrators sometimes need to read restricted data, such as salaries, to fix something. A good system lets them, but through a logged break-glass request that notifies the record's owner. That is provenance for reads, and it is the counterpart of provenance for writes.

Where this fits with access control

Keeping facts true and keeping facts private are separate problems that share a mechanism: an explicit record of who may do what, and an explicit record of who did. OWASP's guidance on LLM applications notes that instructions in a prompt are not a dependable control, and that system-prompt restrictions can be bypassed through prompt injection. The practical lesson carries over to writes. Do not ask a model to be careful with a field. Make the system refuse, or make the change a proposal.

We cover the read side in how to give AI agents your company's knowledge without leaking what they shouldn't see, and how an MCP connection fits in what is an MCP server.

Measuring whether it works

A few simple numbers show whether the model is healthy.

  • Proposal queue age. If the oldest proposal is six weeks old, nobody is reviewing, and the system is failing in the quiet way.
  • Acceptance rate by source. A source with a ten per cent acceptance rate is noisy and needs fixing or muting.
  • Repeat proposals. The same rejected value coming back is a bug in the source.
  • Time since confirmation. Facts with no review in a year are candidates for a nudge to their owner.

None of these needs special tooling. A table view of proposals, grouped by source and sorted by age, answers all four.

How Contextely approaches this

Contextely records the provenance of values and treats a contradiction from an agent or a sync as a proposal, not a write. The person's value remains until its owner accepts, edits or rejects the proposal. AI-suggested sharing is shown as an exact preview and is never applied without a person approving it.

Agents connect over MCP, a CLI or a REST API and see only what the person they act for can see. Permissions cover fields as well as records, and administrators cannot read restricted data without a logged break-glass request that notifies the owner.

The short version

AI will disagree with people, and sometimes it will be right. The goal is not to prevent disagreement but to make it visible, attributed and cheap to resolve. Keep the person's value until a person decides, record where every value came from and route conflicts to the owner. Company facts stay true when someone is accountable for each one.

Questions

What happens when an AI agent contradicts what a person wrote?

In a well-designed system, nothing is overwritten. The agent's value is recorded as a proposal with its source, and the person who owns the record accepts, edits or rejects it. The original stays in place until someone decides.

What is provenance in a knowledge base?

A record of where each value came from: who or what wrote it, when and from which source. The W3C describes provenance as information about the entities, activities and people involved in producing a piece of data.

Should AI ever be allowed to edit company records directly?

For low-stakes, easily reversed fields and well-scoped agents, perhaps. For facts that people rely on, such as ownership, policy or compensation, we think machine changes should be proposals reviewed by the record's owner.

Doesn't review slow everything down?

It adds a step only where there is a conflict. Agreeing updates and additions to empty fields can go through. The review queue should hold disagreements, not routine changes.

Start with what you already have.

Free for up to five editors. Unlimited readers on every plan.