Five of the world's largest software companies shipped company memory in eight days. Most businesses still do not have one.

Something happened in the second week of October 2026 that is easy to miss because it arrived as five separate press releases from five unrelated firms. Every one of them was about the same problem, and none of them used the same word for it.

Dell called it trusted context. SAP bought a company for its context graph for work. Google shipped four kinds of memory. Meta shipped ActiveMemory. Workday shipped an agent harness with built-in context, memory and guardrails. Five of the largest enterprise software firms on earth, inside eight days, all deciding that the thing their products were missing was memory of the customer's own business.

This matters more for a company of fifteen people than for a company of fifteen thousand, and for an unglamorous reason. The reason is on the next line.

What actually shipped in eight days

Take these in date order, because the clustering is the signal. Large vendors do not converge on one architecture by accident.

Two honest caveats before this is used as evidence. Dell's three components are scheduled for release in the first half of 2027, not shipping today. SAP's announcement is an agreement to acquire, expected to close in the fourth quarter of 2026, and TechWolf is not yet part of SAP. So of the five, two are announcements and three are shipped products. That does not weaken the point, all five are large commitments to the same architecture, but it does mean this is not a week in which five products became available.

Dell states the underlying problem more plainly than anyone else, and it is worth quoting because it is not a marketing line: “In most cases the model isn't the bottleneck to achieving enterprise scale for AI. The data is.” It goes on: most company detail “was created long before anyone expected an AI agent to read it. It was never labeled, connected or organized with a machine reader in mind.”

That sentence describes almost every small business in the country. The knowledge is all there. It has never been written down in a form a machine can consult.

The arithmetic nobody adds up

Here is where it stops being a technology story and becomes a payroll story.

Glean's Work AI Institute surveyed digital workers and published its findings on 10 June 2026. The headline figure is that 75% say AI makes them more productive, saving roughly 11 hours each per week through automation alone. 87% of digital workers now use AI at work. Both numbers are real and both are good news.

Then the report asks what happened to those hours, and the answer is that a large share of them were spent getting the AI to work at all. Glean names it botsitting: feeding it missing context, checking its outputs, debugging its mistakes, rerunning prompts, cleaning up confident-but-wrong answers. The average is 6.4 hours a week, which the report notes is more time than workers spend using the tool to produce work. The largest single component is 2.3 hours a week loading context the AI should already have had.

Do the subtraction. 11 hours saved, 6.4 hours spent making it usable, leaves 4.6 hours. That is 42% of the advertised gain. No one is lying; the productivity is real. It is being paid for twice.

Scale it to a business. A ten-person company, at a loaded cost of GBP 25 an hour:

Per person, the context feeding alone is 119.6 hours a year, or three working weeks. The full botsitting load is 333 hours, or better than eight weeks. Those are order-of-magnitude figures from one survey than a measurement of your business, and the survey's own sample weights matter, but the shape of it is unlikely to surprise anyone who has watched a colleague paste the same background into three different tools before lunch.

Why the gains never reach the company

The same Glean report contains the statistic that should end every AI productivity argument: only 13% of workers say their firm is performing significantly better as a result. Eighty-seven percent of them use it. Seventy-five percent of them are individually more productive. Thirteen percent say the company got better.

McKinsey's 2026 global survey, fielded from 4 May to 8 June 2026 with 1,719 respondents across 97 nations, finds the same shape from the other end. Eighty percent of respondents say AI has improved their own productivity. But the share reporting enterprise-level financial impact has not changed since the previous year: 37% attribute at least some EBIT impact to AI, and the proportion of genuine high performers, those attributing at least 5% of EBIT to AI and calling the impact large, has stayed flat at about 6% of all respondents.

Individual productivity is spreading fast. Company productivity is not. The bottleneck is not intelligence and it is not access. It is that each person's AI session starts from nothing.

Glean quantifies that. More than half of workers. 53%, say the critical detail they need to do their jobs is not accessible through their AI systems. They then close the gap themselves, by hand, every day, at the cost measured above. And the report separates workers into context-rich and context-poor groups by that single question. The gap between them is stark. Workers in context-rich firms are 64% less likely to feel worn out by AI, 52% less likely to ship work they cannot explain, and 31% less likely to ship AI output they cannot defend.

An independent source says the same thing from outside the AI industry. BCG's AI at Work 2026 report, published 3 June 2026 and based on 11,749 workers across 14 markets, found that “nearly half (47%) report spending more time managing and directing AI than doing the work itself.” Glean measures the tool; BCG measures the same phenomenon without selling the tool. Both point at the same missing layer.

Which produces a specific, unglamorous conclusion. The constraint on AI value in most firms is not the model. It is that the company's own knowledge is not in a form its tools can read.

What it costs when the knowledge walks out

There is a second cost, and it is the one that arrives without warning. It is what happens when the person who holds the knowledge leaves.

A survey cited by Nationwide found that “71% of small firms reported that they were dependent on one or two key people for ensuring firm success.” That is the normal case, not the exception. And when that person goes, the knowledge goes with them, because it was never written down anywhere a replacement could read.

The replacement cost has been measured properly. Oxford Economics, in a February 2014 report for Unum, found that across five UK sectors the loss of an employee earning GBP 25,000 or more carried an average financial impact of GBP 30,614. GBP 25,181 of lost output while the replacement reaches full productivity, plus GBP 5,433 of logistical replacement cost. The range runs from GBP 20,113 in retail to GBP 39,887 in legal firms.

Then the harder number, which is the one that explains why:

Source: Oxford Economics, "The cost of brain drain", February 2014. Fifteen weeks is nearly four months of a new person operating below their eventual output, and that is the best case.

Two further surveys put a figure on what the rest of the team loses in the meantime. Panopto's 2018 Workplace Knowledge and Productivity Report, run with YouGov, found that “42 percent” of institutional knowledge is unique to the individual holding it, if an employee leaves, colleagues cannot perform 42% of that person's job. The same survey found 81% of employees say personal and professional experience is the hardest kind of workplace knowledge for an firm to replace once lost. And during their first month in a new position, new hires spend 12.7 hours a week asking coworkers for help.

That last figure is the bridge between the two halves of this article. Twelve point seven hours a week is one person rebuilding, from scratch, the context that used to live in one head. Multiply it across a team that turns over, and it is the same problem as the 2.3 hours of context feeding, with a different cause. One is a person leaving. The other is a machine that never knew.

The headline figures on turnover cost are real but should be quoted carefully. Gallup states the cost of replacing an employee “can range from one-half to two times the employee's annual salary, and that's a conservative estimate”, and puts US voluntary turnover losses at roughly $1 trillion a year. SHRM gives a range of “50% to 200% of their annual salary, depending on their level”. Both are published ranges than measured averages, and Gallup's own article cites no underlying study for its band, so cite them as the estimates they are than as findings.

What a company brain actually is

Google's 8 October announcement is the clearest public description of the thing, because it names the four parts. A company brain is not a chatbot. It is four distinct kinds of memory, and only one of them is a chat.

Read that second row again, because it is the one that matters commercially. Semantic memory is your company's approved facts. Procedural memory is how your work gets done, which is almost never identical to the documented process. Episodic memory is why no one has to re-explain the client history every January.

Workday's announcement of 13 October adds the piece that turns memory into something a business can run. Its Agent Harness runs every agent, whether built by Workday, a customer or a partner, under the same permissions, approvals and audit trails. That is governance, and it is the difference between a knowledge base and a liability. SAP's acquisition is aimed at the same layer from the workforce side, buying a continuously updated model of the work people do than the job titles on an org chart.

None of these five products is what a fifteen-person firm should buy, and none claims to be. The point is the architecture. The largest software firms in the world have now agreed, in public, within one week, that a business's own knowledge is the missing layer, and that it has to be governed, not stored.

Why adding more context makes it worse

There is a tempting shortcut here that fails, and it is worth being explicit about because it is the shortcut most firms take first.

The shortcut is: dump everything in. Every policy, every deck, every email thread, pasted into a tool that searches it. It feels like progress. Chroma's context rot research, published 14 July 2025, explains why it is not. Their finding: “LLMs do not maintain consistent performance across input lengths. Even on tasks as simple as non-lexical retrieval or text replication, we see increasing non-uniformity in performance as input length grows.”

Glean's workers report the same thing from the other side: “the more they load, the worse the output gets.” And for every 10% more time a worker spends feeding context, they are 25% more likely to report feeling worn out by it. The effort is real, the result degrades, and the person doing it is tired.

So volume is not the answer. Which file is the current version is the answer. Whether a spike on the dashboard means something or is a known artefact is the answer. What the operations team does when the official process breaks down is the answer. What “Q3” meant last quarter versus this one is the answer. That is the layer Glean describes as sitting between the raw data and the usable answer, and it is what none of us write down.

Dell's approach is the same insight expressed as infrastructure: define a term once, in a semantic layer, so that every agent gets the same meaning instead of each one deriving its own. The knowledge is not the pile of files. It is the agreement about what the files mean.

The trap: an unverified brain is worse than none

This is the part most conversations about company knowledge skip, and it is the part that decides whether the whole idea helps or quietly makes things worse.

Consider what happens when you put every meeting summary into a system that answers questions from it. Someone says in a meeting that the delivery cut-off is Thursday. The summary records it. Three months later the system states, with total confidence and no visible source, that the delivery cut-off is Thursday. It was wrong when it was said. It has now been promoted into company fact by a machine that cannot tell the difference between a record of a talk and a decision.

The distinction is simple and non-negotiable. A meeting summary is evidence of what was discussed. It is not policy, and it does not become policy by being written down. A system that cannot hold that distinction will answer a customer with something no one approved.

Which is why the governance properties are not extras. A company brain has to be able to do six things a chatbot structurally cannot:

1. Say “I don't know.” A system that guesses when it has no basis is worse than a system that tells you to ask a human.
2. Show where an answer came from. Which document, which version, which date, whose approval. An answer you cannot trace is an answer you cannot defend.
3. Know what was true when. Prices and policies change. The brain must track versions and supersession, not silently overwrite.
4. Respect permissions, per item. Salary detail is not in the same permission class as the holiday rota, and a folder-level setting is not good enough.
5. Be corrected. Someone must be able to mark an entry wrong and have the correction propagate. Otherwise errors are permanent.
6. Be deleted on request. Including a specific memory, not an account. If it cannot forget, it cannot be trusted with anything sensitive.

And one more, which is the commercial reason underneath all of this: the brain has to survive its model. If your institutional knowledge is welded to one vendor's assistant, then changing vendor means rebuilding the most valuable thing you own. Google's own announcement frames this as choosing the best model per job than one model for everything, which is only possible if the memory underneath is portable. Knowledge that cannot leave the tool that stores it is not an asset. It is a lease.

The test that tells you if you have one

Most of this can be checked in a ten-minute talk. Ask these six, in this order.

1. “Show me where this answer came from.” Not a confidence score. A document, a version, a date, an owner. If it cannot cite, it is guessing.
2. “What does it not know?” Ask it a question about your business that is genuinely undocumented. A real brain admits the gap. A demo invents an answer.
3. “Someone told you something that was wrong. What happens?” The answer should describe a correction that propagates and a record that it happened, not a retrain.
4. “Two staff members have different access levels. Show me that working.” Per-item permissions, not a shared folder. Ask about something confidential.
5. “Delete one specific memory.” One fact, not the whole account. If the honest answer is that you cannot, that is your answer.
6. “If you changed AI provider tomorrow, what do we lose?” The only acceptable answer is “nothing but the tool.”

Products that fail three or more of those are a chat window with a database behind it. That is not useless, it is genuinely helpful for some work, but it is not a company brain, and it should not be priced or described as one.

If you want to know what the gap costs in your business, the Operations Audit measures it in your own hours than a survey's average, and you keep the diagnosis whether or not you go any further with us. The first thing it looks at is not your tools. It is how many hours your team spends re-explaining what the business already knows.