Records Management: What AI Does to Everything You Kept

Storage got cheap, so nobody deleted anything. That was fine for twenty years, because the material sat in folders nobody could navigate and the practical protection was that it could not be found. Then somebody connected an assistant to the file share.

What changed

Nothing about the files. They have been there the whole time. What changed is the effort required to surface them.

A folder of draft contracts from 2011, performance notes about somebody who left in 2016, and a spreadsheet of customer details nobody remembers exporting was inert when finding it required knowing it existed. A question asked in plain language across the whole share does not require knowing it existed.

This cuts both ways and that is the point. The same capability that makes the material useful makes it discoverable in a dispute, exposed in a breach, and reachable by any staff member whose access was never scoped because nobody could find anything anyway.

The inventory comes before the policy

Plenty of businesses have a retention policy. Far fewer know what is actually on the drive, which is why the policy is a document rather than a practice.

Start by finding out what you hold. Not perfectly, and not file by file, but enough to size the problem.

File dates deserve suspicion. A migration between systems can rewrite every modified date to the day of the move, so a share that appears to contain nothing older than 2019 may simply have been moved in 2019. Check that before trusting age as a retention trigger.

What the schedule has to say

A retention schedule lists each record type, how long it is kept, what triggers the clock, and what happens at the end. Triggers get left out, and without one the schedule cannot be operated. Seven years from what? Creation, last activity, end of the contract, or the end of the fiscal year in which the relationship closed are four different dates.

Periods are not a drafting exercise. Tax, employment, corporate and sector-specific requirements set minimums, some federal and some provincial, and there are records you are obliged to keep for longer than you would choose to. Get the periods from an accountant and a lawyer rather than from a model, which is the same division as in drafting an employee handbook.

Then write down what happens when the period ends, including who approves it and where the record of the deletion lives. A schedule without a disposal step is a filing system with a theory attached.

Classification, not deletion

The work that stops records projects is the inventory, because twelve thousand untitled files is a year of somebody reading documents. Classifying at that volume is a genuine fit, and it is the thing that makes the project finishable at all.

Deletion is the line. A classification is a suggestion with a confidence attached, and a confidently wrong one that disposes of a document under legal hold cannot be undone. Have it propose, sort by certainty, and route anything uncertain to a person rather than letting the uncertain cases fall into whichever bucket scored marginally higher.

That uncertainty routing is the whole safety design, and it is the same argument as preferring a model that knows when it is unsure over one that is confidently wrong more often.

Access scoping belongs in the same project. If search is about to work across the share, the permissions that were theoretical become real, and that review is cheaper before the tool goes live than after, which is the point of keeping it on a current risk register and in your document control practice.

Frequently Asked Questions