By Mark Buck
AI can retrieve an organisation’s accumulated knowledge almost instantly; ensuring that knowledge remains valid is a slower, more demanding management task.
The clause looked entirely ordinary. For years, solicitors across England routinely inserted versions of it into married couples’ wills to solve a genuine inheritance-tax problem. Then the law changed and, for many routine estates, the clause was no longer necessary solely for that purpose.
An AI-assisted drafting system recently offered the wording again. The solicitor removed it, the client received a suitable will, and by any ordinary measure the review process had worked.
But only for that file. Historic wills using the clause remained perfectly valid records of past practice; that is not the same as being current authority for a new one. The clause was the immediate issue. The knowledge system that could not tell the difference was the real one.
Memory is easier to scale than learning
Every organisation stores more than it actively remembers. Precedents, templates, procedures, prior drafts and standard letters are all codified decisions, retained rather than reasoned through afresh each time. This is organisational memory, and it is not the enemy of learning[1]. It is the residue of learning, and often one of its greatest assets: consistency, efficiency, and the accumulated judgement of people no longer in the room.
Organisational learning is a slower, different process. Someone must notice that circumstances have changed, interpret what that means, agree what should replace the old approach, and then alter documents, systems and behaviour to match[2]. Memory only has to be retained. Learning must be contested, decided and propagated, and none of that happens by itself.
When a firm acquires or closes a business, or absorbs a change in the law, the same decision must be carried through a dozen disconnected places: letterheads, website footers, supplier records, internal systems, public listings and the people who must explain the change to a client. None of it happens automatically. AI can rewrite the footer in seconds once instructed; it does not know the footer is wrong until the organisation has recognised the change and told it where that change must be applied.
Retrieval-augmented systems, which pair a language model with an organisation’s own documents so it answers from proprietary material rather than general training alone, make the first of these dramatically easier[3]. A precedent that once required someone to find, recognise and copy it can now be retrieved and recomposed in seconds, dressed in fluent, apparently deliberate language: old judgement, presented in newly generated prose, is harder to recognise as old. Call it synthetic freshness.
Institutional theorists call this layering: new rules are added while old ones remain[4]. AI did not create the habit; it makes every layer searchable at once.
The precedent that outlived its reason
Before transferable nil-rate bands existed, an unused inheritance tax allowance on the first spouse’s death could not ordinarily be carried forward to the survivor[5]. Nil-rate band discretionary trusts were a rational response: they preserved that allowance rather than losing it. The Finance Act 2008 introduced the transferable nil-rate band. Where the surviving spouse or civil partner dies on or after 9 October 2007, their nil-rate band can be increased by the proportion unused on the first death; for married couples, that first death may have occurred before or after that date.[6]
For many straightforward estates, that removed the reason for the trust. It did not remove the trust from every precedent bank, and it did not make the structure obsolete everywhere. HM Land Registry’s own guidance notes that such a trust may subsequently be found to offer no tax advantage, while also confirming it can retain genuine value: protecting assets for children of an earlier relationship, for instance, or avoiding a forced sale of the family home[7]. The honest position is not that the clause became wrong. It is that a solution designed for particular circumstances had, in many routine cases, hardened into a default.
That is the harder thing to police. A clearly incorrect clause is easy to catch. A clause that was once right, remains technically available, and simply no longer suits most clients requires judgement on every file, exactly the judgement a fluent, confident draft makes easy to skip. The same problem appears wherever a firm maintains controlled knowledge: contractual terms, clinical protocols, compliance scripts. Each can remain in active use long after the assumption beneath it has changed.
Why internal grounding solves only half the problem
Professional firms are increasingly adopting private, organisation-controlled AI systems to manage confidentiality and data-protection risk, often grounding them in internal documents rather than relying solely on a model’s general training. That has real advantages: better traceability, more relevant answers, less reliance on uncontrolled sources, material that can, in principle, be inspected.
It solves one problem, not two. Privacy determines who may see the knowledge. It says nothing about whether the knowledge deserves to be used. A private system can retrieve an outdated precedent just as confidently as a public one can retrieve a wrong one, and confidentiality controls catch neither.
In a retrieval-based system, source material can often be updated or removed without retraining the underlying model. The question is whether the firm treats its knowledge base as an archive, added to indefinitely, or as a governed system, with clear ownership of what remains in it.
Correction without learning
Return to the solicitor. The firm now handles a basic will in one of two ways: with the clause and a reminder to remove it, or without the clause and a prompt to add it back in. Either way, someone must make an active decision on every file. Removing the clause where it doesn’t belong is output correction: the immediate document is put right. The forced prompt itself is process control: the firm has recognised a recurring risk and placed a human decision in its path, whichever direction the default runs. Both matter, and both protect clients. The firm has learned enough to build a safeguard. What has not happened is the conversion of that learning into the knowledge system itself.
The system does not treat repeated removal as evidence that the clause should stop appearing by default. The reviewer’s decision is not used to reclassify the historic material, alter the retrieval rules, or change what the next file is offered. The same judgement must be supplied manually, every time. Organisational learning would mean converting those repeated professional decisions into a change upstream: distinguishing historic records from current precedents, altering the system’s default behaviour, and deciding when the clause should appear only by exception.
The document was corrected. The process adapted. The knowledge system did not learn.
That is not evidence the firm was careless. If anything it is the opposite: the firm noticed a recurring problem and changed how people work in response, which is organisational learning by any reasonable definition. A control can prevent an error without changing the knowledge that keeps producing it; human oversight can make the output safe while leaving the system’s organisational memory untouched. Organisational memory was available to the AI automatically. Organisational learning remained dependent on people.
The case against the argument
None of this makes AI the villain. It would be dishonest to pretend it does. The same systems that retrieve an old precedent can compare versions, flag a reference to withdrawn legislation, and push an approved correction through an entire knowledge base in one pass. Nor is old knowledge inherently inferior: organisations rely on precedent because much of it remains sound, and one that reopened every settled judgement as a matter of habit would drown in its own caution.
The honest claim is narrower than “AI preserves bad practice.” It is that AI currently makes retained knowledge easier to reuse than it makes the assumptions underneath that knowledge easier to challenge. Closing that gap is a governance choice, not a limitation of the technology.
The Institutional Memory Audit
Six questions, applied to any precedent a knowledge system offers with confidence.
What is the source and status? Where did it come from, who signed it off, and is it current, superseded, or nobody’s checked?
What decision originally justified it? Every precedent begins as an answer to a real problem. What was the problem?
Which assumptions must remain true? A precedent is only as sound as the conditions it was built on. Naming them makes it obvious when one breaks.
Who owns review and retirement? Firms routinely appoint people to create precedent. Very few appoint anyone to retire it.
How are overrides converted into learning? When professionals repeatedly remove, reject or amend the same output, is that pattern merely recorded, or does it trigger review of the source material, the retrieval rules and the default behaviour?
How are corrections propagated? Once a judgement has been validated, how does it change what the system offers to everyone else?
Every knowledge object should therefore have an owner, a status and a review trigger. Most firms know who wrote a precedent; few know who must decide when it has expired.
Conclusion
Artificial intelligence does not force an organisation to repeat its past. It changes the cost of doing so. Old precedents can now be found, adapted and applied with unprecedented ease, while the work of questioning them remains stubbornly human and organisational.
The challenge, then, is not merely to teach AI the company way. It is to decide which parts of the company way still deserve to be taught.
Disclaimer
This article uses a UK will-drafting example to illustrate a broader management argument about AI and organisational knowledge. It does not constitute legal, tax or financial advice. The treatment of nil-rate band trusts should be assessed against current law, guidance and each client’s circumstances.
AI acknowledgement
AI tools were used for research support and sentence-level refinement. The article’s thesis, examples and final editorial decisions are the author’s own.
Mark Buck
