Lost in the Chat Scroll: The Audit Trail AI Leaves Behind

The model answered. The employee copied the answer into the deck, the ticket, or the ledger. Next quarter someone asks why the number moved. The person has changed teams. The chat, if it was retained at all, sits in a personal account the company does not administer. The system of record shows the new value and not the proposal, the policy version, or the name that let it through. Nothing in that sequence required a bad model. It required a place that forgets by design.

That place is the chat scroll. Consumer assistants, and the enterprise chat windows modelled on them, are excellent at a single sitting. They are a poor system of memory. Organisational intelligence — which rule was in force, who had already signed a similar exception, which source the figure came from — is trapped in threads that do not join. People become the join. They paste, re-explain, and re-ground a fresh session because the last session cannot see the ledger and the ledger cannot see the last session. Seventy-eight percent of AI users bring their own tools to work. Each of those tools is another scroll the company will not be able to subpoena from itself.

A lifecycle graph is the record built for the question the scroll cannot answer: what caused this outcome? Not a transcript of tokens. A chain — the brief, the sources in scope, the action proposed, the person who released it, the system that changed, and when. W3C’s provenance model has described this shape for years: entities, activities, and agents, linked so derivation can be reconstructed. A lifecycle graph is that instinct applied to work a model helped do. It is not a claim that a chat product has implemented a standards stack. It is a claim that without some chain of that kind, the company does not know why its own systems moved.

How context fragments

Fragmentation is not a metaphor. It is a Monday.

A finance manager asks a personal assistant to explain a revenue line, pasting an export because the official tool cannot see the close file. A seller, in another product, drafts a concession from a battlecard that was replaced in January. Support, in a third, answers the same customer from a macro the model ranked as relevant. Three sentences now exist about one commercial fact. None of them knows about the others. The human who notices, if anyone does, is doing integration work the stack refused to do.

The pattern scales with adoption. McKinsey’s 2025 survey puts use in at least one function at 88 percent of organisations, with nearly two-thirds not yet scaling. Local chats multiply faster than any shared memory. Each chat is a small truth with a confident tone. The company does not have a contradiction report. It has employees who have learned to check the assistant against a colleague, which is the old process with extra steps.

There is a legal version of the same fracture. If a customer relied on one of those sentences, the company will be asked which sentence was the company’s. A tribunal has already declined to treat the chatbot as a separate legal person. Context fragmentation is how you arrive at that hearing with three versions and no chain.

The last row is the point of the architecture. It does not automatically ingest the personal scroll. It makes the personal scroll unnecessary for the jobs that matter, because the sanctioned path can see the file and will keep the chain.

What retrieval does not do

Most “memory” projects in the last two years have been retrieval: chunk documents, embed them, fetch the nearest passages, hope the model behaves. Retrieval is necessary and it is not a record of work. It answers “what text looks similar to this question?” It does not answer “what did we decide last quarter, under which policy, with whose signature?”

Similarity is indifferent to time. A page from 2022 and a page from this morning can be equally near the query. The model then blends them, which is how a bereavement rule and a current policy page both speak. Similarity is indifferent to authority. A draft in a shared drive and the certified close can both be retrieved. The draft is often better written. Similarity is indifferent to permission. A chunk the searcher should not see is a chunk the index may still return if the index was built for the whole corpus.

Temporal context, dependency, and prior sign-off are not passages. They are relations. The revenue line depends on a contract. The contract was amended. An exception was signed in March and must not be re-litigated by a model that has not been shown the signature. A vector store can be made to hold some of this with enough metadata discipline. At that point you are no longer “doing RAG.” You are building a graph and using search as one way to enter it. Enterprise retrieval remains the right tool for “find the clause.” It is the wrong tool for “why did this field change?”

Four questions, one chain

People reach for “4D” because a flat log feels one-dimensional: a list of messages in time. The record operators actually need has to answer four questions at once.

What. The entity: a customer, a contract, a journal line, a policy version, a workstream.

What happened. The action: a draft, a refusal, a proposed write, a released write.

When. Not only a timestamp on a message. Which version of the rule and which state of the system were in scope. A bi-temporal habit — when the fact was true in the world, and when the company recorded it — is what stops a late correction from rewriting history.

Under whose authority. The person or role that released the action, and the rule that made it theirs to release. A model is an actor in the chain. It is not the authority.

The lifecycle graph is this chain kept as the company’s, not as a feature of one chat vendor. It points at systems of record. It does not become a second ledger. The official number stays in the financial system. The graph says how an AI-mediated job touched it. That distinction matters. A shadow ledger that “becomes the truth” is a new fragmentation, only more confident.

Provenance of this kind is also what NIST’s framework is asking for when it says map and measure, and what a data-protection inquiry means by a record of processing. The ICO’s AI guidance does not require a branded graph. It requires you to be able to say why this data was in this system. A scroll that the employee selected because the tab was open is a weak answer. A chain with a purpose, a source, and a time is a serious one.

A close that can be reopened

Consider a quarterly close in which revenue recognition depends on a contract, a delivery milestone, and an exception finance signed earlier in the year. The work is familiar. The new risk is that each input now has a fluent twin.

Without a graph, the sequence is a tour of tabs. Someone exports the contract list. Someone else asks an assistant whether a performance obligation has been satisfied, pasting a paragraph. A third person drafts the commentary. The commentary is lucid and slightly wrong about the exception, because the exception lives in an email the assistant did not see. The pack goes to the audit committee. Six weeks later the auditor asks why a line was treated as recognised. The company can produce the journal. It cannot produce the chain. The hours spent reconstructing it are the cost of the scroll.

With the chain, the job is a workstream. The contract and the milestone are reads against systems the team is allowed to see. The earlier exception is an entity on the graph, with the person who signed it and the date it still binds. The assistant may draft commentary that cites those objects. It may not introduce a treatment that is not among them. If a proposed journal entry would change the books, a person releases that payload. The graph keeps the draft, the refusal if there was one, the sources, and the release. When the auditor asks, the answer is a path, not a search through inboxes.

Nothing in that picture requires the model to “understand accounting.” It requires the company to keep the decision in a place a successor can open. Standards such as revenue recognition remain the finance team’s. The graph is how an agent-assisted close stays inspectable. The same shape fits a claim, a price exception, or a hiring decision: prior authority in, proposed action, human release, outcome stored.

The person who became the join

Context fragmentation has a job description, and someone on the team already holds it. They are the colleague who “knows where that lives.” They keep a private note of which assistant gave a usable answer, which export is current, and which email contained the exception finance will ask about. When a new hire needs the same fact, the colleague pastes it again. When they are on leave, the fact is on leave with them. The organisation has automated the draft and left the integration to a person.

That labour is easy to miss because it looks like diligence. The manager who re-grounds a fresh chat with three attachments before every meeting is performing retrieval the stack would not do. The analyst who checks a model’s number against a colleague is performing reconciliation. The counsel who asks “which version did it see?” is performing provenance by interview. None of these tasks appear in a pilot’s success slide. They appear as calendar load, and they grow as more teams adopt more windows. McKinsey’s finding that most organisations are still piloting is consistent with a world in which local fluency is high and shared memory is still a person.

The cost shows up when the person leaves or when two people glue different versions. Sales has a concession in one scroll. Support has a warmer cousin of it in another. Finance discovers both when the credit posts. The reconstruction is then an email search, which fails if one of the scrolls was a personal account the company cannot open. Institutional memory is the property that survives that departure. A lifecycle graph is the mechanism: the exception is an object with an author, a date, and a scope, and the next job starts from the object.

What the next run is allowed to inherit

The first close proves the chain can be built. The second close is where it earns its cost. A later run preparing commentary should meet the earlier exception as a relation it can cite: who signed it, which contract it binds, whether the date still holds. If the exception remains in force, the draft quotes it. If it has lapsed, the run says so and stops, rather than extending a treatment because last quarter’s prose sounded settled. Inheritance here is a relation a person can expire, narrow, or refuse to apply. It is not a model’s recollection of a thread.

Departments compound when they share that relation. Legal’s sign-off on a clause helps finance only once finance’s workstream can see it without asking legal to paste the email again. Support can see that a concession was refused and can stop drafting a near-copy for the next customer. The human join shrinks because the next brief starts from the chain, not from whoever remembers the meeting. Collaborative work is this handoff: several functions, one operational story, a model that drafts inside it.

There is a failure mode that looks like compounding and is only a larger pile. If every chat is ingested into an index and treated as authority, last quarter’s wrong draft becomes this quarter’s source. The graph has to be selective. Released outcomes, current policy versions, and explicit refusals are inheritable. Unreleased brainstorming is not. A refusal is as useful as an approval: it tells the next swarm which treatment was already declined, and by whom. Teams that store only the happy path will relitigate the same bad idea every cycle, fluently.

Bi-temporal discipline is what keeps a correction from laundering history. When a milestone is later found to have been met on a different date, the company records both the new fact and the moment it learned the new fact. The commentary that shipped in March remains explicable: it was written against what was known in March. A scroll that is edited, or a vector index that quietly re-chunks, cannot show that distinction. Auditors and successors ask it constantly. The ICO’s expectation that you can explain why data was used is the privacy version of the same question. The operational version is “why did we recognise this line?”

What to stop storing only in the thread

Pick the decisions that already cause arguments: revenue, concessions, coverage, access, anything a customer can screenshot. For each, require the chain before the outcome is allowed to stand. If the work happened in a personal chat, it has not happened as far as the record is concerned — which means you either bring it into the workstream or you accept that you cannot explain it. Shadow use is already a breach-cost problem, not only a knowledge problem. The same paste that fragments context is the paste that leaves.

Do not fund a “memory” programme that is only a larger index. Fund the relations: version, authority, scope, and the link to the system that actually holds the fact. Search remains how people enter. The graph is what they are entering.

A call to chief information officers

The chat window will not get less convenient. People will keep it for drafting that does not touch a decision. The decisions themselves have to leave a chain your successor can open. That is not a knowledge-management fashion. It is how you keep a single operational story once every team has a model.

If the only answer to “why did this change?” is a transcript nobody can find, the company does not yet have institutional memory. It has a collection of sittings. The graph is the difference.

References

• Microsoft and LinkedIn, 2024 Work Trend Index
• W3C, PROV-Overview
• McKinsey, The State of AI: Global Survey 2025
• Moffatt v. Air Canada, 2024 BCCRT 149
• NIST AI Risk Management Framework
• ICO, guidance on AI and data protection
• IBM newsroom, 30 July 2025