What happens to company knowledge when every process gets its own AI agent?

Every organization has a memory, and it is not the one stored on its servers. It is the clause in a standard contract that exists because of a lawsuit in 2011, the precise wording a regulator once required for a product description, the sentence a chief executive insisted on rewriting before a crisis statement went out, and the tone that makes a customer in Madrid feel they are dealing with the same company as a customer in Chicago. Nobody wrote this memory in one sitting. It accumulated over decades, through corrections, arguments and hard lessons, and it now lives in validated documents, style guides, approved terminology and the heads of a few long-tenured employees who can still say, with authority, "we never phrase it that way, and here is why."

It is among the most valuable assets a company owns. It appears on no balance sheet. And the rapid spread of AI agents across corporate processes is about to test it as nothing has before.

A company of a thousand authors

The scale of what is coming is easy to underestimate. Gartner projects that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024, and that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024.

When a single process acquires an agent, the organization gains a productive assistant. When every process acquires one, it gains something far less obvious: hundreds of new authors, each writing in the company's name. The sales agent drafts proposals, the support agent answers customers, the HR agent writes policies, and the product agent produces documentation. Across markets, other agents translate and adapt that output, multiplying every inconsistency across languages, terminology, regulation and cultural context. Few of these agents have been shown what the company has already said, and fewer still know which words were fought over and why they matter.

This is a less familiar version of what might be called the agent tax. It accumulates when unvalidated agent output is copied, translated, stored and reused until it becomes difficult to distinguish from authoritative knowledge. The visible costs of deploying agents at scale, such as compute, supervision and integration, are already catching up with many organizations: Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. The subtler cost is harder to budget: the gradual dilution of an organization's accumulated knowledge by a growing volume of content that looks authoritative but was never validated.

When errors stop looking like errors

For most of corporate history, a poor document announced itself. Awkward phrasing, clumsy structure and visible hesitation served as warning signs. Generated content has removed that signal. Every sentence is fluent and every paragraph is confident. A term carefully negotiated with the legal department is replaced by a more natural-sounding synonym. The same product acquires three different names in three markets. A cautious commitment becomes a slightly broader promise.

The legal system has already begun to price this risk. In 2024, a Canadian tribunal ruled on a case in which an airline's website chatbot gave a grieving passenger wrong information about bereavement fares. The airline argued the chatbot was a separate entity; the tribunal rejected this, holding that as part of the airline's website, the company was responsible for all information provided, including that from the chatbot. Just as telling was its answer to the airline's claim that the correct policy was published elsewhere on its site: the tribunal found no reason why a customer should know that one section of the company's webpage is accurate and another is not.

That sentence deserves to be read in every boardroom. From the outside, a company does not have a marketing voice, a legal voice and a support voice. It has one voice, and it is liable for all of it.

A memory that overwrites itself

There is a deeper risk that receives less attention. What agents produce does not simply leave the building. It flows back into shared drives, knowledge bases and repositories, the same sources that will ground the next generation of agents.

Research on model collapse offers a warning, although it does not directly test corporate knowledge systems. A 2024 study in Nature found that models trained recursively and indiscriminately on model-generated data can lose information from the tails of the original distribution. In that research, the "tails" are rare events in the training data. The enterprise parallel is an analogy, but a useful one: repeated reuse of unvalidated agent output may crowd out the unusual clause, the regulatory nuance or the term that matters in only one market. Those are often the details an organization can least afford to lose.

The broader information environment is moving in the same direction. In a sample of 55,400 English-language articles and listicles from Common Crawl, Graphite classified 50.9% as primarily AI-generated in Q4 2025 and 49.9% in Q1 2026. Detection at this scale is imperfect, but the result shows how quickly synthetic material can approach parity with human-written material in a major source of training data. The multilingual picture is starker still. Researchers at Amazon found that web content is often translated into many languages and that the low quality of many multi-way translations indicates machine translation. They also found evidence consistent with low-quality English content being translated at scale into lower-resource languages.

An organization's memory is not stolen in this process. It is overwritten by its own copies.

The disappearance of a useful filter

The agent tax also has an absurd dimension. When content was expensive to produce, every document had to justify its existence. Cost was a crude filter, but a real one. Once production becomes nearly free, that filter disappears. Reports nobody reads are generated weekly, documentation for discontinued products is maintained with perfect diligence, and every page exists in every format, for every audience and in every language.

This matters because finding the right information was already one of the most expensive activities in corporate life. In 2012, McKinsey estimated that the average interaction worker spent nearly 20 percent of the workweek looking for internal information or tracking down colleagues who could help with specific tasks. The estimate is old, but the underlying problem predates generative AI. More documents do not make knowledge easier to find. They make the validated version harder to recognize.

An architecture nobody chose

Perhaps the most uncomfortable aspect of this shift is that almost no company deliberately decided to organize its knowledge this way. It accumulated, largely from the bottom up. Microsoft and LinkedIn's survey of 31,000 knowledge workers found that 75% now use AI at work, and that while 79% of leaders agree AI adoption is critical to remain competitive, 60% worry their company lacks a vision and plan to implement it. In the meantime, 78% of AI users are bringing their own tools to work, missing out on the benefits of strategic AI use at scale and putting company data at risk.

The result is less a strategy than a reflection of the existing organizational chart, translated into software, an outcome the computer scientist Melvin Conway predicted in 1968 when he observed that organizations design systems mirroring their own communication structures. Silos that were once cultural problems become embedded in tools, prompts and integrations. Once each part of the company speaks through its own agents, restoring a single, coherent voice is no longer an editorial decision. It becomes something closer to an excavation.

The controls that come before the tools

Most corporate conversations about AI begin with technology: which model, which platform, which integration. The controls that protect an organization come earlier. Four capabilities matter. First, authoritative knowledge: the company must identify which documents, terminology and decisions are validated. Second, ownership: named people must be able to approve or change content with legal, regulatory or brand weight. Third, traceability: the organization must preserve the source and rationale behind important wording. Fourth, controlled reuse: unvalidated agent output must not become authoritative merely because it has been copied, translated or stored. Together, these controls let a company decide what content should exist, in how many versions and where human judgment remains necessary.

Gartner's own advice points in the same direction: in many cases, organizations should rethink workflows with agentic AI from the ground up rather than insert agents into legacy processes. The hard decisions concern how the organization works, who owns its knowledge and what evidence its agents may treat as authoritative. Those decisions are considerably easier to make before a thousand agents begin writing than afterward.

The asset nobody insured

Companies invest heavily in protecting their brands, patents, data and code. Very few give comparable attention to what they have learned to say and how, even though it is the asset every customer encounters directly. The commercial stakes are not abstract. In a survey of 8,709 consumers across 29 countries, CSA Research found that 76% of online shoppers prefer to buy products with information in their native language, and 40% will never buy from websites in other languages. The same research found that 75% of respondents are more likely to purchase the same brand again if customer care is in their language. A company's voice, in every language it speaks, is not a cosmetic layer. It is where trust is won or lost.

Artificial intelligence will not destroy that asset overnight. Its effect will be quieter: it will make content cheap to produce, easy to copy and increasingly difficult to keep true to its source. That is the real agent tax: the cumulative cost of losing authority, traceability and meaning as unvalidated content spreads. The organizations that emerge ahead are unlikely to be those that produce the most. They will be those that still know what they mean and still sound like themselves, wherever they speak.

References

1. Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027," press release, June 25, 2025.
2. Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia Civil Resolution Tribunal, February 14, 2024.
3. Shumailov, I., et al., "AI models collapse when trained on recursively generated data," Nature 631, 755–759 (2024). doi:10.1038/s41586-024-07566-y
4. Graphite, "AI Hasn't Overtaken Human Writers Online," Five Percent research, May 2026.
5. Thompson, B., et al., "A Shocking Amount of the Web is Machine Translated: Insights from Multi-Way Parallelism," Findings of ACL 2024, pp. 1763–1775.
6. McKinsey Global Institute, "The social economy: Unlocking value and productivity through social technologies," July 2012.
7. Microsoft and LinkedIn, "2024 Work Trend Index Annual Report: AI at Work Is Here. Now Comes the Hard Part," May 2024.
8. Conway, M., "How Do Committees Invent?", Datamation, April 1968.
9. CSA Research, "Can't Read, Won't Buy – B2C," July 2020.