Our position · AI and your company's DNA
Our AI series episode walked through the five ways data leaves for AI assistants, and the rules that fit on one page. Those rules are necessary, and they stay necessary. Here we propose a new angle of protection: next to your data and your operations, you now need to protect your company's DNA.
Our thesis comes in three steps. Protection was first built around data: the file that leaks, the database you encrypt. It then extended to operations: continuity, recovery, the vendors you cannot do without. With generative AI, a third perimeter appears: what makes customers choose you, your way of doing things, which we hand to AI models every day, one prompt at a time, without any data loss prevention tool ever seeing it go.
Three perimeters, three questions:
- Data: who can read it? That is confidentiality, and it stays essential.
- Operations: how long can you run without them? That is resilience.
- DNA: who else is learning to do what you do? That is the new angle, and almost nobody measures it.
This is a position, not settled fact. We wrote it to debate it: tell us in the comments where you disagree.
The control we are still building
For three years, the corporate response to generative AI risk has followed the playbook of data protection fifteen years ago: classify documents, block sites, filter what gets pasted, ban personal accounts. It is useful, and we recommend it. But the playbook rests on two assumptions that no longer hold.
The first: that you can list the tools. A new assistant, extension or agent ships every week, and AI now arrives inside the tools you already approved: the office suite, the CRM, the mailbox, the developer environment. The blocklist is always a week behind.
The second: that what has value is always a file. A data loss prevention tool recognizes a card number, a contract, a customer list. It does not recognize a way of pricing a bid, sequencing a diagnosis or handling an angry customer. Yet that is exactly what you pass on when you work with an assistant.
Your company's DNA is not a document
Files matter: a leaked customer list or medical record is still an incident, with its obligations. But ask an executive why customers choose them. They rarely answer "our files". They answer: the way we estimate a job, our method for selecting a supplier, our grid for accepting or declining a file, the way our best salespeople run a call. That is your company's DNA: know-how, most of it written down nowhere. The philosopher Michael Polanyi put it in a sentence that stuck: we know more than we can tell.
That is precisely what generative AI is good at capturing. Not the final document, but the path to it. The law already recognizes part of this: in its factsheet on generative AI and intellectual property (2024), the World Intellectual Property Organization (WIPO) warns that a business putting confidential information into prompts may disclose a trade secret or waive its confidentiality, since the provider keeps a copy. Status: guidance from an international body, not binding, but it describes the state of the law. A trade secret is only protected if it is reasonably protected; know-how narrated to a tool whose terms nobody read is less and less so.
The TRACE: what every use leaves behind
The word isn't ours. In AI systems, a trace already means a precise technical object: the full record of a request, broken into linked units of operation. It holds the user's input and the detected intent, data retrieval (documents pulled from a store, their relevance scores, the assembled context), the tool and API calls the agent ran, the successive model calls (parameters, tokens consumed, latency, errors) and the final answer. Teams running agents use it to understand, debug and monitor what the system does.
That technical trace already says a lot: step by step, it records how your company asks a question and what it fetches to answer it. But it speaks to engineers. To make the stakes clear to leadership, we prefer to present the TRACE through our own acronym: everything an AI use leaves with the provider and in the model, beyond files, seen through the lens of your company's DNA. Five components, one per letter.
- Tasks. What you delegate says what you do. A department's prompts over a month describe its processes better than its documentation.
- Reasoning. How you break a problem down, the order of your questions, the assumptions you make: your method, written in plain text, prompt after prompt.
- Approvals. What you accept, reject and correct. Every "no, more like this", every regenerated answer, every thumbs up or down teaches the tool what an expert at your company considers good. It is the most valuable part, and the most invisible.
- Context. Customer, supplier and competitor names, prices, lead times, margins, given "so the answer is relevant". Not a document marked confidential; a map of your business.
- Expression. Your tone, your vocabulary, the way you negotiate or reassure. What makes a customer recognize your company from the first email.
Where does that trace go? It depends on the settings, and that is where the position becomes defensible rather than speculative.
- It is retained. History, assistant memory, provider logs: even when nothing is used for training, a copy sits with a third party, for as long as the contract (or the consumer terms) says.
- It can train the models. At the major providers, consumer-account conversations are used by default to improve the models unless the user opts out; business offerings generally commit to the opposite. Feedback given to the tool (rating an answer, flagging it) sometimes follows separate rules. Status: policies published by the providers, which change; check them in the contract, not on the marketing page.
- It settles into the tools. Saved prompts, custom assistants and agents that chain your work steps are codified know-how. When they live in an employee's account at a provider, they don't belong to you, in practice and not always in law.
NIST lists intellectual property and data privacy among the twelve risks specific to generative AI (NIST AI 600-1 profile, 2024), and OWASP ranks sensitive information disclosure second among the risks of large language model applications (LLM02:2025).
Why it is controversial
Two objections are legitimate, and we take them seriously.
"A model doesn't regurgitate one company's know-how." Largely true, and we don't claim otherwise. The risk is not a competitor reading your method word for word. It is slower: by learning the practices of a whole sector, the model makes good practice available to everyone, including those who didn't have it. What set you apart becomes the default answer. You have lost nothing visible; you have lost your lead.
"Then we should stop using AI." No. A company that refuses AI loses its lead faster than one that uses it. The question is not to stop leaving a trace, it is to decide where it goes and who owns it. Same logic as with public assistants: a ban moves the problem, governance shrinks it.
Protect the DNA, not just the file
1. Name your five pieces of know-how. One hour with leadership: the five to ten ways of doing things that explain why customers choose you. An estimating method, a selection grid, a sales script, a recipe, a process. That is your inventory of intangible assets, and it fits on one page.
2. Add a fourth class. The three data classes from our episode (free, anonymize first, never) describe content. Add a row for know-how: which of it can be worked on with an approved assistant, and which stays in a tool you control, or between people.
3. Set training, memory and feedback. On the approved tool: training turned off by contract, retention written down, assistant memory governed, and clear rules on feedback given to the tool. On personal accounts, know-how stays out, like data.
4. Bring home what you codify. Prompts, custom assistants, agents and automation flows are company assets: an owner, a versioned internal repository, an export path. In the contract, a clause on who owns what you configure and on portability if you switch providers.
5. Treat AI providers as suppliers. They now hold part of how you work. They belong in your third-party inventory, with their subcontractors, their commitments and an exit plan.
Put it in the register
The scenario is written with its four parts. A fictional example: "Estimators use a consumer assistant on personal accounts to price bids; the estimating method, margins and supplier prices are retained outside any contract and may be used to train the model; the company can neither show it protected its trade secret nor recover the method when an estimator leaves." The business consequence is not a fine: it is an eroding lead, and know-how that walks out with the person. Like every scenario, it needs an owner.
This week's test
Pick your best piece of know-how, the one that wins contracts. Then ask the person who masters it one question: if you left tomorrow, where would the method be? In your head, in a company document, or in your AI assistant's history?
Three answers should worry you. "In my history": your method lives with a provider. "On my personal account": in practice, it no longer belongs to you. "I have a custom assistant that does it": your know-how is codified, but not at home.
What you should have in a month
A page naming your five pieces of know-how, a fourth class in your AI use policy, training and retention set by contract on the approved tool, the company's prompts and agents in a repository you own, and your AI providers in the third-party inventory. Data still needs protecting, and so do operations. Your company's DNA becomes the third perimeter, and it doesn't look like a file.
What do you think?
We know this position divides. Some see a real, underestimated risk; others a non-issue as long as business contracts exclude training. Tell us what you think in the comments under our LinkedIn post: does your know-how already go through an AI assistant, and have you written it down anywhere? The best objections will feed a future article.
Where FortaRisks helps
Forta Compliance holds your AI use policy and inventory and maps them to ISO 42001 and the NIST AI RMF among more than 40 frameworks: the training rule, the agent review and the "know-how" class become dated controls, with their evidence. Forta Suppliers tracks your AI providers like any other third party, with their commitments and subcontractors. Forta Cockpit puts the know-how erosion scenario in the register, in front of leadership. To discuss your case, thirty minutes with an expert is enough to map your know-how and your AI use.
Sources: WIPO, Generative AI: Navigating Intellectual Property (2024) · NIST, AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (July 2024) · OWASP, LLM02:2025 Sensitive Information Disclosure