The thesis
AI accelerates individuals and fragments organisations.
The acceleration is real and per-person. The fragmentation is silent and collective — and it happens because every AI reads from a shared pile and writes into a private void.
The evidence
Everyone got faster. The organisation did not.
This is not a vendor argument. It is what the published research already says about what happens when AI meets work that is not fresh.
Developers with an AI pair-programmer finished a fresh task faster. The gain is real, and it is not in dispute.
Peng et al., 2023
The same tools, experienced developers, real issues in code they had known for years — while they believed they had been made 20% faster.
METR, 2025
For every 25% rise in AI adoption, teams reported feeling more productive while delivery throughput fell 1.5% and stability fell 7.2%.
DORA, 2024
Why it happens
The industry built the feed and never built the return.
Every time someone works with an AI, they hand it the organisation’s understanding of itself — what a customer means here, which of four systems called billing is real, why the migration stopped in 2023. Almost all of it comes out of their own head, because that is where it lives.
Then the session ends, and that understanding is destroyed. The next person rebuilds it from scratch. The colleague two desks away never had it.
Three years have gone into optimising the reading half — more context, bigger windows, better retrieval. The return path was never built. It is half a loop, shipped as though it were whole.
The consequence
Two colleagues, each with an AI, diverge inside a week into two versions of the same company — and neither can tell.
That is a silo: private knowledge, invisible to peers, lost when the holder leaves. With one new property. It assembles in an afternoon.
The wall
Cost rises, efficiency falls, and the output cannot be trusted.
That is the wall enterprise AI hits at scale. Everything we build ladders up to those three.
Cost
Retrieval that costs more the more you know
Everyone is moving to RAG, and flat retrieval means every question drags more text through the model. The bill tracks usage, and the more knowledge you add the worse it gets. Cited facts break the link between how much you know and what a question costs, so the same budget goes further as the knowledge grows.
Efficiency
A pile of documents still leaves the hard part to you
Handing a model twenty documents and asking it to sort them is not retrieval, it is delegation of the hard part. The right typed facts, cited, mean less to read, less to get wrong and less to redo.
Trust
Nobody can tell you why it said that
An answer with no provenance cannot be audited, corrected or defended. When a fact turns out to be wrong, you need to find every decision that rested on it — and that is only possible if the facts were typed and cited in the first place.
At enterprise scale
Finding the right document is not the same as being right.
This is the failure that does not show up in a demo, and the one that decides whether you can put AI in front of a regulated process. Both answers below came from the same question and the same documents.
One question, asked two ways
“Do we need in-region key management for the new analytics service we are standing up in the US?”
A frontier model reading your files
158,023 tokensYes. Your encryption policy requires customer data to be encrypted at rest with keys held in the same region as the data, rotated every 90 days. You will need in-region key management in place for the US deployment before it can go live.
Confidently wrong
It found the right document. The in-region key requirement is scoped to EU regions only — a qualifier further down the same policy than the clause it quoted. Nothing in the answer tells you an EU rule has just been applied to a US service.
Cornerstone
1,288 tokensNo. In-region key management applies to EU regions only. For US regions the standing requirement is encryption at rest with 90-day rotation, and no in-region key management is needed.
Why it holds
The scope is part of the fact, not a sentence buried near it. A rule that applies to the EU cannot come back as though it applied everywhere — along with who set it and when, so you can check.
123×
fewer tokens for the same question
100%
of required documents retrieved by file search
3 of 8
tasks it still got wrong
Measured across 2,201 public documents, plus a separate set of eight realistic briefs with deliberate traps in them. The failure shown is the one the run recorded: file search was handed the encryption policy in full and still applied an EU-only rule to a US service. No method in that run fabricated a citation — the reliability problem is not invented sources, it is confident answers with the qualifier dropped.
What we do about it
Build the return path.
What any AI learns about the business becomes structured, attributable knowledge the whole organisation owns — cited to what the company actually wrote, versioned as it changes. Every session repays what it borrowed.
Measured, not asserted
Seven methods. Same documents. Same model.
How much less you spend for each answer you can actually use. Query tokens only — the one-off cost of building the index is excluded, because on your own model it is not billed per token. Each bar is what that method costs you that Cornerstone does not.
- Cornerstone vs File searchReading whole documents99.7%
- Cornerstone vs Vector RAGSearching by meaning20.2%
- Cornerstone vs Hybrid + rerankKeywords + meaning, re-sorted18.2%
- Cornerstone vs Hybrid RAGKeywords + meaning · best alternative13.8%
Not on this measure: GraphRAG. A knowledge graph answers from its own graph rather than by handing back a source document, so “did it find the right document?” has no answer in these terms — and without that there is nothing to divide the spend by. It is on the other three measures, and on the trap corpus it captured 7 of 10 planted facts against Cornerstone’s 10.
13.8% cheaper per correct answer than the best alternative, 20.2% cheaper than searching by meaning, and 99.7% cheaper than reading whole documents.
How it wins
Per question these methods look level — Cornerstone is 1.5% cheaper than vector search and nobody should buy anything on that. The gap opens once you count only the questions that were actually answered. Vector search finds the right document 78.3% of the time, so a fifth of what it spends buys nothing; Cornerstone finds it 96.7% of the time. Same spend, more of it landing: 2,421 tokens per correct answer against 2,810 for the best alternative, 3,033 for vector search, and 861,225 for handing the AI whole documents to read. Measured directly rather than derived, the three-method report puts the vector gap wider still, at 26.0%.
Why it matters to you
This is the number that behaves like a bill. You are not buying questions, you are buying answers somebody can act on, and a method that is fractionally cheaper per question while missing one in five is charging you for the misses twice over — once in tokens, and again in the time it takes somebody to work out the answer was never there. One caveat stated plainly: building the index is a real one-off cost, and on a hosted model it is a real token bill. Run your own model, as Foundry customers do, and it is an overnight job on a GPU you already own — 3.8 hours for 2,201 documents — with no per-token charge at all. That is why it is excluded here, and it is the only thing excluded.
Knows when there is no answer — 41 of 49
Every other method scored no — and not narrowly. A search always has a closest match, even when nothing in your documents answers the question at all. So it hands back its best guesses, and the AI writes them up with exactly the same confidence it would use if the answer were really there. Nothing in what you read tells you which just happened.
The defining property is compounding.
Today each session is a write-off and tomorrow’s begins at zero. It does not have to. Each session can leave the next one better — across people, across time.