For business purposes, the "is it really intelligent" debate is over: the models read, reason and draft well enough for most office work. What surprised us in client projects is how little that changes on its own. Your competitors bought the same models in the same quarter, sometimes on the same subscription tier.
Capability stopped being a differentiator the moment it became purchasable. What can't be purchased is the one thing no general model knows: how your company actually works.
Two things worth writing down here:
- why context, not the model, became the scarce part
- where that context is quietly accumulating right now (this part worries us more)
What the smartest model doesn't know
Not "doesn't know yet": can't know. Things like:
- the discount a customer negotiated years ago and still expects
- which supplier says yes and delivers no
- why the March numbers look wrong but aren't
- the handshake agreement that overrides the contract
- the rule that exists because of an incident nobody wants to repeat, etc
Little of this is written down anywhere. It lives in inboxes, in the heads of two long-tenured employees, in the gap between the process document and what the team actually does. A model with encyclopedic knowledge of the world has none of your company in it, and everything useful it could do for you depends on exactly that missing part.
So the practical equation inverts. The scarce input is no longer intelligence, it's context. Same model plus your context produces work; same model without it produces plausible text.
Where the context is accumulating right now
Vendors understood this, which is why every AI product suddenly has "memory". The assistant remembers your preferences, the copilot learns your codebase, the agent platform collects your workflows and corrections.
Convenient. And worth a second look, because each of these memories is being built inside someone else's product. Knowledge fed to an assistant accumulates in that vendor's silo, in their format, under their terms. Repeat that across the five or six AI tools your departments adopted this year, and the company's operational knowledge is spreading into five or six containers that don't talk to each other, and that you don't control.
At that point the switching cost stops being the licence and becomes the memory: leaving a vendor means losing what your own people taught it. And the models themselves churn fast (this quarter's best is next quarter's second-best), so the sensible posture is picking the strongest model per task and switching without ceremony. That posture is only available if the context doesn't live inside any single one of them.
Treating knowledge as infrastructure
The alternative we keep arriving at: treat company knowledge the way companies long ago learned to treat data: as infrastructure you own, with everything else connecting to it.
Concretely, one governed place where facts, rules and history live:
- customer terms, approval rules, entity records
- decisions and the reasons behind them
- readable by every process and every model the company chooses to use
- access revocable per scope, stored where the data already lives, not in a vendor's cloud
Two properties matter more than any feature list. The knowledge must be shared: one memory for the company, not one per tool, otherwise the silo problem just reappears internally. And it must be portable: kept in a form that outlives any model vendor, so switching engines stays a procurement decision.
The natural home for this is wherever processes already pass through. An execution layer sees every process run and can write what was learned back into shared context, which is roughly the posture flow8 is built around.
Three questions for the next exec review
This topic tends to hide inside "AI strategy" discussions, where it doesn't really belong: it's an ownership question. Three questions usually surface it:
- where, physically and contractually, does the context our AI tools have accumulated live?
- if we dropped our main AI vendor next quarter, what knowledge walks out with them?
- which system is accumulating what we learn, and do we own it?
We keep meeting companies with excellent models and nothing of themselves to feed in, and the difference a year of accumulated context makes is bigger than any model upgrade in that period. Curious how others handle this: where does your company's context live today, and who could take it away? Happy to compare notes!