A Knowledge Base Needs Skill Provenance Before AI Makes It a Taxonomy

Somebody wrote up the outage.

They filed it under LESSONS_LEARNED_DO_NOT_DELETE.xlsx, which is already a difficult name to defend in a board paper.

It contains a customer escalation, a workaround that stopped working after a product release, a handful of tags supplied by the team that wrote it, and a cheerful comment saying “Useful for onboarding.”

The comment is probably correct.

The tags are probably useful too.

Then a generative assistant searches the page, finds configuration management, and adds it to your organisation’s emerging skills list.

A local retrieval label has just become a workforce classification.

A lesson learned is evidence of work in context. A skill is a capability claim.

They meet.

They are not the same object.

This became urgent because AI can retrieve, summarise and recombine pages before the person who remembers the incident has finished finding the old ticket.

A label that once helped a colleague find a narrow procedure now looks like an authoritative statement of what every engineer is expected to know.

The model has not become unreasonable.

Your data contract has become ambitious.

A lesson begins as a local event

The useful flow starts with a concrete surprise.

An outage.

A near miss.

A successful handoff.

A customer escalation.

Somebody leaving.

Take an ordinary one: two services interpret the same configuration value differently and delay a customer launch.

The first record captures a date, a system scope, the participants, a problem statement, and why the knowledge might matter elsewhere.

That record deliberately separates three things.

Evidence of what happened.

A working explanation.

A proposed change.

This is not decorative colour-coding for people who enjoy incident reviews.

It lets a later reviewer correct an explanation without editing away the history.

NASA describes lessons learned as knowledge from successful or unsuccessful work.

Its Shuttle closeout captured about 112 of them.

The point is not that you need a Shuttle programme, which would create several other administrative concerns.

The point is that knowledge gets captured while the people, timelines, logs and decision context still exist.

At this stage a tag like config or partner-integration has a proper job.

It helps retrieval.

It connects a searcher to the incident, the services, and the people who understood it.

It does not yet mean that configuration management is a canonical capability.

It does not mean everyone with the tag has it.

It certainly does not mean every role now requires it.

Your wiki thinks labels are facts, because the search index needs nouns.

The durable object is the applied judgment

The durable thing beneath the page is not its title, its tag or its folder.

It is the capability evidenced by applying a reviewed procedure or judgment in a defined work situation.

Recognising a configuration-contract risk.

Checking the relevant conditions.

Changing the control accordingly.

The lesson has a practical action and an owner: add an explicit unit field to the interface schema, and a cross-service test using mismatched fixture values.

It also records what would falsify it.

That last part is the one everybody skips.

If later evidence shows both systems already enforce units, then the problem was never “schemas lack units.” The problem may be that review was bypassed.

So the incident supports a claim about a mechanism, under conditions.

It is not a clean universal noun, waiting to be imported into a taxonomy because a model found it beside a code block.

Public taxonomies are useful here precisely because their labels arrive with a purpose, a structure and a version.

O*NET, ESCO, Lightcast and SFIA all publish what their identifiers mean and when they changed.

A wiki tag arrives because somebody wanted to find Thursday’s answer before lunch.

Keep the knowledge base. It is doing real work.

None of this is an argument against knowledge bases, procedure libraries, lessons-learned programmes, or the people who maintain them with an alarming number of browser tabs open.

Knowledge management preserves continuity after projects end.

It stops a later team rediscovering the same failure.

NASA describes a manager receiving links to relevant lessons within 15 to 30 minutes, for a funding decision on a high-cost project.

Retrieval changed the decision, because the records supplied concrete precedent.

That is real value.

To do it, the knowledge base needs searchable titles, synonyms, source links, access controls, owner information, and enough context to distinguish versions.

A published lesson should carry system and product versions, a date, an owner, an access group, source links, and whatever triggers its review.

ISO 30401 is a knowledge-management systems standard, and its value here is not that a standard number makes a wiki wise.

Its management-system framing requires intent, accountable people, resources, identified risks, and performance review.

That is what separates a controlled technical standard from an approved decision record, a working note, or an unverified community contribution.

Keep the lesson.

Keep the tags.

Keep the highly specific search path that lets a new employee find the correct procedure instead of inventing one.

Just do not let the page’s local finding aids become your skills ontology by accident.

Authority is not a formatting choice

A controlled record carries a responsible owner, a purpose, an effective date, a status, a source, and a sensitivity classification.

Those fields tell a reader whether a search result is instruction, history, or a hypothesis wearing a confident title.

Authority also has more than one owner-shaped concept behind it.

A platform administrator can restore a deleted page.

A subject expert can explain a system.

A designated owner is accountable for factual accuracy.

A decision authority can commit the company.

In a small team one person may do all four. The distinctions remain, rather rudely, even when the org chart is compact.

The access boundary matters as much as the accuracy boundary.

Search snippets, semantic ranking and AI answers must not disclose restricted text, or infer facts from documents the requester cannot open.

Where several sources contribute to one answer, the strictest relevant boundary applies. A transfer or a departure changes access in the source system, and therefore in search.

An assistant that reads a restricted incident and produces a cheerful general skills recommendation from it has not improved knowledge discovery.

It has invented a permission bypass with a better interface.

Supersession is provenance with a calendar

Knowledge decays.

Product versions end.

Supplier contracts change.

Teams reorganise.

Laws change.

A last-updated date alone is weak evidence.

A reviewer has to look at the sources, the owner, the scope and the recommendation.

The result is one of four states: confirmed current, revised, superseded, or archived.

A superseded lesson points to the current guidance.

An archived record stays historical and non-operational.

Neither should vanish because the old recommendation became awkward in a search result.

The National Archives describes the same pattern for a federal office updating a file plan.

Review the functions.

Consult the inventory.

Match records to the schedule.

Develop the plan.

Your records regime may differ.

The operational shape travels: inventory what exists, identify change, decide disposition, communicate it.

This is the part AI systems find inconvenient.

A generated answer prefers a neat sentence.

Provenance occasionally replies, “That sentence applied to version four, before the supplier change, for a team with a particular access boundary, and it has now been superseded.”

Excellent.

That is what a trustworthy answer sounds like.

The extraction pipeline needs a human stop

The shortcut is understandable.

Your knowledge base already contains procedures, incident analyses, decision records, training notes and field language.

An AI system can retrieve them and propose skills for a capability model. Your workforce platform wants recommendations, gap analysis, learning pathways and mobility matches.

Nobody wants to convene the committee again.

But a generated extraction needs a record of what it found.

The source object.

The owner.

The status.

The context.

The applicable version.

The access classification.

The proposed relationship to a canonical capability.

And when the label reaches a public taxonomy, it needs mapping direction, scope, confidence and review status too.

That is not paranoia.

A residual filing category in a respected occupational classification is useful for filing and unsuitable for automatic inference.

If a national statistical system can contain one of those, your wiki tag deserves the same humility.

The human reviewer does not need to perform mystical taxonomy rituals.

The reviewer asks one manageable question.

Does this lesson evidence a capability in the stated conditions, and does the proposed skill keep the uncertainty, scope and provenance of that evidence?

Sometimes the answer is yes.

A reviewed procedure can evidence a narrowly defined capability.

Sometimes the answer is no.

The lesson may be a local workaround, a deprecated product feature, a historical warning, or a perfectly good search term.

That is not an extraction failure.

That is the system declining to promote a page tag into a company-wide requirement because the demo is on Thursday.

What survives the pipeline

Knowledge management makes organisational learning retrievable across time.

Skill taxonomies make capability classifications legible across contexts.

AI can help connect the two.

It cannot make the boundary between them disappear without making a claim on your behalf.

The durable object is the demonstrated capability: reviewed judgment applied in defined conditions.

The lesson, the page, the tag, the search result, the taxonomy identifier and the generated answer are all wrappers.

Around evidence, or retrieval, or classification.

Related objects, all of them.

And none of them the same object.

And if LESSONS_LEARNED_DO_NOT_DELETE.xlsx has quietly become your skills architecture, the spreadsheet has acquired responsibilities it did not apply for.