Every company has a person everyone calls when something goes wrong. When that person retires, the knowledge goes with them. AI can capture a meaningful share of it, but only if the capture starts while the person is still working. The window is the last eighteen to twenty-four months of their tenure, not the exit interview.
That is the whole problem in one paragraph. The rest of this piece is about what actually walks out the door, why Alberta feels it more sharply than most markets, and what a realistic capture program looks like.
What actually leaves when a thirty year employee retires
Ask a leadership team what their senior estimator knows and you usually get a shrug and a compliment. He just knows. That answer is the reason the knowledge disappears.
Four distinct things are leaving, and they need different handling.
Documented knowledge. Procedures, drawings, templates, price books. This is the easy layer. Most companies already have it, and most companies overestimate how much of the total it represents.
Semi-documented knowledge. The spreadsheet on someone’s desktop with fourteen tabs and no labels. The email folder that functions as the real project archive. The margin notes on a printed scope. This layer exists in writing, but nobody else can find it or interpret it.
Tacit knowledge. Judgment. Why this subcontractor gets called for tight-schedule work and that one does not. Which client always changes scope in week three. How to read a site before the geotech report arrives. This layer is the expensive one.
Relational knowledge. Who to call at the municipality. Which supplier will still take a Friday afternoon order. Thirty years of goodwill that does not transfer through a contact list.
Most knowledge transfer programs handle layer one, gesture at layer two, and ignore three and four entirely. Then the company spends the next four years relearning things it already paid to learn.
Why this lands harder in Alberta
The industries that carry this province run on long-tenure operating expertise. Energy services, heavy construction, agriculture, logistics, municipal operations, industrial fabrication, and family businesses that are now in their second or third generation of ownership.
These are not companies with rotating twenty-eight year old analysts documenting process for a living. They are companies where the person who knows how the plant behaves in a cold snap has been there since the nineties.
There is also a succession overlay. A large share of Alberta’s mid-market is owner-operated and approaching transition. When the founder leaves, the operating knowledge and the customer relationships and the pricing instinct all leave at the same moment, and the incoming buyer or family member inherits an org chart without an operating manual.
Statistics Canada reported that 19.2 percent of Canadian businesses used AI to produce goods or deliver services in the second quarter of 2026, roughly triple the level of two years earlier. Adoption is climbing. Very little of it is pointed at knowledge retention, which is one of the reasons this is an open lane.
The four layers, and what AI does with each
Here is the practical mapping. This is where the Imagination Gap usually closes for leadership teams, because they can finally picture the work.
Layer one, documented. AI indexes it and makes it answerable in plain language. Instead of a shared drive nobody searches, a person asks a question and gets an answer with the source document attached. Low risk, fast payback, good first project.
Layer two, semi-documented. AI helps normalize the mess. Inconsistent spreadsheets, scattered email threads, and legacy naming conventions get structured into something a system can use. This work is unglamorous and it is where most of the real hours go.
Layer three, tacit. This is the one that requires human effort first. AI does not extract judgment on its own. What works is structured interviewing, recorded and transcribed, then organized into decision patterns. You sit the estimator down for ninety minutes a month and walk through real past bids, asking why at every branch point. The transcript becomes a source. The AI system turns it into something searchable and, over time, into decision support that reflects how the company actually thinks.
Layer four, relational. AI helps with the map, not the relationship. It can surface who has spoken to whom, what was agreed, and where the history sits. The handshake still has to be handed over in person.
A labelled hypothetical
A mid-sized mechanical contractor in Calgary has two estimators. One has been there twenty-six years and plans to retire in 2028. The company wins roughly one in four of the bids he prices and one in nine of the bids the junior estimator prices.
The program runs eighteen months. Monthly recorded sessions walking backward through completed jobs, including the losses. Historic bid files get structured and tagged. An internal assistant is built on top of that corpus so the junior estimator can ask why a similar scope carried a fifteen percent labour contingency in 2019.
The realistic outcome is not that AI replaces the estimator. The realistic outcome is that the junior estimator’s win rate moves toward the senior estimator’s, and the company stops repricing the same lesson every year.
This example is illustrative. It reflects how these engagements are typically structured rather than a specific client result.
Where this goes wrong
Starting at the exit interview. Two hours in the final week produces a document nobody reads. The capture has to be spread across months while the person is still doing the work, because the work is what triggers the memory.
Recording everything and structuring nothing. Four hundred hours of transcript is not knowledge management. It is a bigger haystack.
Treating it as an IT project. The people who know what matters are in operations. If operations does not own the program, it produces an archive instead of an asset.
Skipping the person’s consent and dignity. Someone with thirty years of tenure will read this as being made replaceable if it is handled clumsily. Handled well, it reads as legacy. Most senior people want their knowledge to outlast them. Say that out loud, early.
A ninety day starting sequence
Days 1 to 30. Identify the three to five people whose departure would cause the most operational damage. Not the most senior people. The most load-bearing ones. Map what each of them holds across the four layers.
Days 31 to 60. Fix layer one and two for a single function. Pick the function with the clearest financial exposure, usually estimating, service dispatch, or client onboarding. Build search and retrieval over the existing material so the team gets an immediate win.
Days 61 to 90. Run the first structured interview cycle with one person. Two sessions. Review what came out of it with the team who will inherit the work, and adjust the questions.
Clarity before code. Readiness before rollout. Sequencing before spend. The order matters more here than in almost any other AI project, because a knowledge system built on the wrong function is worse than no system at all.
Common questions
How long does a knowledge capture program take? Twelve to twenty-four months for a single critical role, running alongside normal work. Compressing it into a quarter produces documentation rather than judgment.
Does this require the person to be technical? No. The sessions are conversations about real past decisions. The technical work happens after, on the transcripts and the existing files.
What about data privacy and confidentiality? Internal knowledge systems should run on infrastructure your organization controls, with defined access rules by role. Client-confidential material needs a separate classification pass before anything is indexed. This is a governance question and it belongs at the start, not after the pilot.
Is this different from a wiki? A wiki holds what someone remembered to write down. A knowledge system built this way holds what someone was asked the right question about, and makes it retrievable at the moment of decision.
What size of company does this make sense for? Anywhere from roughly twenty-five employees upward, if the concentration risk is real. A ten person firm where one person holds everything has the same problem and usually needs a simpler version of the same approach.
Can AI capture relationships as well as process? It captures the record of relationships. It does not capture trust. Plan a human handoff for the accounts that matter.
The part worth remembering
Alberta built itself on transferred knowledge. Ranching techniques, drilling practice, cold weather construction, irrigation, trades apprenticeship. Every one of those moved from one generation to the next through deliberate teaching, not through documents.
AI does not change that principle. It changes the medium and the reach. The teaching still has to happen, and it has to happen while the teacher is still in the building.
If your organization has someone whose retirement would genuinely hurt, that is the first AI project worth funding. Start with the mapping, not the tooling.
To scope what your organization would actually lose and what can be captured, start a conversation.
