Governance · 22 Sept 2026
What will your company’s AI actually know?
As businesses give AI more responsibility, much of the reasoning behind past decisions sits with people and advisers who may move on. Russell looks at why gathering records is not the same as establishing what happened.
Russell Gould, Founding CEO
Before we ask AI to take on more responsibility in a business, I think we need to ask a fairly ordinary question: how well can that business explain itself?
Why was a particular corporate structure chosen? Which assumptions supported a financing decision? Was the lawyer's recommendation adopted, modified or left on the table? Does advice given three years ago still apply?
These questions matter because the next generation of business AI is being built to do much more than summarise documents.
In YC's official Summer 2026 startup call,Tom Blomfield,Monzoco-founder and then aY Combinatorgeneral partner, argued for a “company brain”: a way of making scattered business knowledge usable by AI. He has since taken leave from YC to joinAnthropic's compute team. His argument invites us to examine what knowledge is available to assemble in the first place.[1][19]
There is significant investment behind this direction. In June 2025, Glean announced $150m in Series F funding at a $7.2bn valuation. Weeks earlier,Coworker.ailaunched with a $13m seed round and an organisational-memory architecture called OM1. Those announcements show commercial interest; they do not establish that any platform has solved the whole problem.[2][3]
The major platforms are moving too.Microsoft'sJared Spatarodescribed Work IQ in November 2025 as combining work data, memory and inference, with permissions and compliance controls. Anthropic's Applied AI team has explained how agents can retain useful context through structured notes outside a model's immediate context window.[4][5]
This is substantial progress. It would be wrong to argue that the industry has overlooked memory, context or access control.
What interests me is the part of company understanding that crosses organisational boundaries.
A business's history is shaped by founders, employees, boards, finance teams, lawyers, accountants, tax specialists and strategic advisers. Each may see a different part of the decision. Their records may serve different purposes.
Illustrative scenario, not a documented case.
Consider a company reorganisation. The legal documents may establish the resulting structure. The tax advice may explain the circumstances under which it was recommended. A finance model may contain assumptions about future funding. The board may approve only part of the original proposal.
Finding all four records would be useful. Understanding their relationship would be more useful still. But if the records disagree, or a subsequent decision was never documented, assembling them does not by itself settle what happened.
While the people involved remain available, gaps can sometimes be filled by picking up the phone. The finance director can explain an assumption. The accountant can recall why an alternative was rejected. The lawyer can clarify which conditions mattered.
But leaders leave, advisers change, firms merge and people retire. Documents may transfer without all the reasoning behind them. A new team can inherit the outcome of a decision without the understanding needed to interpret it. The person who could explain why may no longer be available.
That becomes commercially important when the same history is examined again. Due diligence for a sale, a fundraise or new financing can revisit overlapping documents, advice and decisions. Each exercise may involve reconstructing context that someone assembled previously.
A data room can organise the material for a transaction, and it can be maintained afterwards. But creating one does not establish who will keep the explanations current, record subsequent changes or preserve what was learned. Without that continuity, the next event can trigger another round of questions and reconstruction.
These gaps can cost money: management time, repeat professional fees and delays while uncertainty is resolved. The company may pay again to establish something it previously understood.
Retrospective reconstruction is not automatically a red flag. But if a company cannot substantiate why a material decision was made, what advice it relied upon or whether the necessary approval was obtained, a buyer, investor or lender may reasonably ask what else remains unresolved.
That is how an internal knowledge gap can become a commercial problem. The company is trying to establish the facts at precisely the moment someone else is deciding how much confidence to place in it.
Access is another part of the problem, although it needs careful description.
It is perfectly possible for advisers to share information appropriately. ICAEW expressly discusses authorised access to working papers, including in due diligence. It also distinguishes records belonging to a firm from those belonging to its client. The SRA's guidance similarly addresses disclosure with consent or legal authority.[6][7]
So the issue is not that a company's AI can never communicate with an adviser's systems. It is that installing an integration does not itself establish the authority, scope or purpose for sharing. Nor does it make every record relevant, current or conclusive.
There are practical examples of progress. Morgan Stanley announced Debrief in June 2024: a tool that captures client meetings with consent, prepares follow-up material for adviser review and stores notes. It illustrates how AI can help preserve context during the relationship, before someone later has to reconstruct it.[8]
The distinction between having information and using it properly is equally important.
In its findings on the Patisserie Holdings audits, the Financial Reporting Council identified failures involving evidence, documentation and professional scepticism. This was an audit failure, not a test of AI or proof of a missing-memory problem. Its relevance is narrower: information supplied by management still required challenge and corroboration.[9]
Knight Capital offers a different lesson. The SEC found that 97 automated emails identified an error before the market opened on the day of its 2012 trading incident, but the firm did not act on them. The messages were not designed as system alerts. Again, the lesson is about controls and response, not proof that a memory tool would have prevented the loss.[10]
Both examples should make us cautious about assuming that collecting more material automatically produces sound decisions.
A recent survey, Memory in the Age of AI Agents, describes a field with differing definitions, approaches and evaluation methods. It is a research preprint, rather than proof that corporate history cannot be reconstructed. Its relevance is that “memory” covers several distinct technical problems.[20]
Research gives us reasons to be careful, without requiring pessimism. The peer-reviewed RAGTruth work documented unsupported or contradictory outputs even when models were supplied with retrieved information. Earlier long-context research also showed that where information appeared could affect performance. These studies examined particular systems and tasks; they are not a ceiling on future capability.[11][12]
Indeed, subsequent work demonstrated improvements in long-context retrieval. Better models, better evidence and better workflows can all reduce uncertainty.[13]
The harder limit is an evidential one. If two materially different accounts fit everything currently available, the evidence has not yet established which account is correct. A system may reasonably assign probabilities or recommend further investigation. It should not quietly turn that uncertainty into a settled history.
That applies to people as well. A confident recollection from a former director is not automatically definitive. Human review is useful because someone can test an interpretation, seek missing material and take responsibility for a decision—not because people are infallible.
For business leaders, I think the practical response begins with deciding which history matters enough to preserve. We do not need to record every conversation. We do need to be clearer about consequential decisions: the advice considered, the assumptions relied upon, the action authorised and any later change.
AI can help with that work. It can prepare summaries, connect related material, flag inconsistencies and ask questions that expose gaps. Advisers and companies can agree what should be handed over and maintained as part of their normal relationship.
As automation takes on more responsibility, being able to explain the basis for its actions becomes more valuable. The question I would put to a board is simple: if your AI gives an answer about an important past decision, can the company explain why that answer deserves to be relied upon?
Sources
[2]Glean Series F announcement
[3]Coworker.ai launch and seed funding
[5]Effective context engineering for AI agents
[6]Disclosure of confidential information
[7]Confidentiality of client information
[8]Morgan Stanley Debrief launch
[10]Knight Capital enforcement
[20]Memory in the Age of AI Agents, v2 full text
[11]RAGTruth
Russell Gould, Founding CEO · 22 Sept 2026