For consultancies, "own your AI" is the pitch of the year. You hear it from model providers, cloud vendors and the new wave of AI services firms. The usual package has three parts:
- open-weight models fine-tuned on your data
- your data kept under your control
- a bench of experts to build it all for you
Part of this is right, and we offer open-weight models to our own customers. Enterprises should not build their AI on a single rented model they can't move. But ask what you own once the project ends, and the answer is often a set of model weights and a services contract. Neither is where your company's knowledge lives.
This post is about the part of ownership that the pitch tends to skip: company context with its sources and permissions intact.
Market Signal
The shift to control is real
The move away from single-model dependence shows up in enterprise spending. Andreessen Horowitz's January 2026 survey of 100 Global 2000 CIOs found that average enterprise spend on large language models rose from about $4.5 million to about $7 million, with respondents forecasting a further 65% rise this year. More than 80% of those enterprises now test or run three or more model families in production, while open-source models account for 11% of enterprise LLM spend.
At the same time, buyers feel boxed in. In the IBM Institute for Business Value's survey of 1,000 senior executives, 71% said switching their primary AI provider, vendor or model would be difficult if required, and 81% reported that a week-long vendor outage would cause severe disruption.
Open weights answer one half of that worry. If you can take the weights with you, no single model provider can raise your prices overnight or retire a model your workflows depend on.
The other half is harder, and gets less attention.
Where Lock-In Hides
What lock-in looks like in practice
A model is the easiest part of an AI system to swap. The hard part is everything the model needs to know about your company to do useful work: which documents are current, which decisions still stand, who is allowed to see what, and how your teams talk about their clients and projects.
Analysts writing in Tech Policy Press and for the Digital Markets Act portability debates have pointed to the same gap between raw file exports and working memory. You can download every file, ticket or transcript you ever put into a platform and still lose the relationships, connectors and decision history that gave those files meaning. Getting raw records out of a tool takes hours; rebuilding the structure around them is what consumes weeks of engineering time.
Harvard Business School's Jeffrey Bussgang made a matching point from the investor's side in his July 2026 essay on company AI brains: frontier models have become interchangeable, while "what compounds is the proprietary context of the firm: customer knowledge, operating history, decisions made and unmade, workflows, judgment, corrections, and (painful) mistakes." Once a platform holds your history and decisions, replacing it means starting from cold.
That is a strong moat if the platform is yours. It is a costly trap if the platform belongs to a vendor, even one running an open-weight model inside your cloud.
Weights vs. Live Memory
Why fine-tuning is a weak place to keep company knowledge
The standard "own your AI" pitch leans hard on fine-tuning open-weight models on company data. Fine-tuning has clear uses: teaching a model a house format, a domain vocabulary, a tone of voice, or a narrow repeatable task so a smaller model can run it at lower cost.
Keeping your company's factual knowledge inside model weights has four problems:
- Models struggle to learn new facts this way. A Microsoft research study led by Oded Ovadia compared fine-tuning against retrieval for adding new information to large language models. Retrieval won across the board, both on facts the model had met during training and on brand-new ones. The model struggled to absorb new facts through unsupervised fine-tuning alone.
- Weights can't keep a secret from the wrong person. Once a document is baked into a model's weights, the model cannot check at query time whether the person asking is allowed to read it. An intern and a board member querying the same fine-tuned model draw on the same weights. If the board minutes went into training, you have a problem.
- Weights go stale the day after training. A price changes, a customer churns, a policy is updated, a deal team reverses an earlier call. Updating weights means another training run, another round of checks, and no clean way to remove the old fact.
- Weights can't show their work. When a colleague asks "where did that figure come from?", a model answering from its weights cannot point to the paragraph in last quarter's board pack. A retrieval system linked to the original passage can.
Fine-tune for style, format and specialist tasks. Keep company knowledge in a live, permissioned knowledge base that any model can read at the moment someone asks a question. We run open-weight models for customers who want them, and we pair them with that layer so the model stays replaceable while the knowledge stays current and traceable to its source.
Why Pilots Stall
Projects stall on context
Look at why enterprise AI projects fail to reach day-to-day use, and the same gap shows up.
MIT's Project NANDA studied 300 public AI initiatives, interviewed 52 organisations and surveyed 153 senior leaders. Only 5% of custom enterprise AI tools had reached production with measurable return. Their diagnosis was blunt: "The core barrier to scaling is not infrastructure, regulation, or talent. It is learning. Most GenAI systems do not retain feedback, adapt to context, or improve over time."
McKinsey's 2026 survey of 1,719 organisations tells a milder version of the same story. About two in ten respondents report scaling AI agents across their organisation, and 37% say AI has added to EBIT, unchanged from the year before.
Better models haven't fixed this, because the models were never the bottleneck. An agent that doesn't know what your company decided last Tuesday, or who on the team is allowed to see a contract, will stall no matter whose weights it runs on.
Living Structure
A map drawn in a workshop is out of date by Friday
Some platforms do treat context as the core problem. The older way to build it is to send in a team of engineers for a few months to interview staff, agree a fixed schema of "objects" and relationships, and wire your systems into it by hand.
When that work is done well, the result is useful. It also creates a different dependency. Your company changes every week: a new product line, an acquisition, a reorganised sales team, a new way of pricing. If updating the map takes another statement of work from the people who drew it, they own the pace at which your AI keeps up with your business.
A company's context should grow from the documents, messages, meetings and records your teams already produce, and update itself as new ones arrive. Every extracted fact should point back to the exact passage it came from, and when a newer source contradicts an older one, the newer one wins and the older one is marked as replaced. No six-month schema workshop before the first question gets answered.
The Six-Part Standard
What ownership should mean
I would apply one test to any AI project, ours included. Suppose the vendor, the services partner and the model provider all disappeared tomorrow. What would you still have?
Real ownership means keeping six things:
- The context, in your cloud. What the AI knows about your company (its clients, decisions, timelines and relationships) stays inside your own AWS, Azure or GCP cloud environment, in a documented format your engineers can query themselves.
- The source of every fact. Every claim in that graph links back to the file, message or call transcript it came from, with the date it was written and whether a later source has replaced it.
- The permissions of every source. A person or an agent asking a question sees only what their access in Google Drive, SharePoint, Slack and your CRM already allows. Someone without access to a file never sees facts drawn from it.
- Freedom to change models. Your teams can use the same context from Claude, ChatGPT, Cursor, an internal tool, or an open-weight model running on your own GPUs, through open protocols such as MCP. Swapping models means changing a setting. Nothing has to migrate.
- The models and the workspace, on open weights. Open weights only take you so far if your people still do their daily work inside a closed assistant tied to one lab. You should be able to run open-weight models for both the company brain and the agent workspace your teams work in every day, inside Your Cloud when data rules demand it, or on 60x Cloud when convenience wins, and mix in frontier models where a task calls for them.
- What your people know, kept apart from what your company knows. Each employee can connect their own inbox and notes to a personal brain that belongs to them, while shared company knowledge sits in the company brain. Neither leaks into the other without a clear choice.
None of this needs a large bench of specialists. We install a company brain for one use case in two weeks, then add the next department. Your team runs it from there.
Buyer Checklist
Questions to ask before you sign
If a vendor or adviser pitches "own your AI" this quarter, ask seven questions:
- Where do our company's extracted facts, relationships and decision history live, and what format do we get them in if we leave?
- When a document is updated or deleted, how fast does the AI stop repeating the old version?
- How do you stop a junior staff member's query from drawing on a restricted document?
- Can every answer point to the exact passage it came from?
- If we switch from one model family to another next month, including to an open-weight model we host ourselves, what breaks, and does our team's workspace keep working?
- How much of our budget goes on something that keeps running after your team rolls off, versus hours billed to set it up?
- How long until one real team is using this on live work?
Open weights are a good start, and we will run them for you. They make the model yours. Your company's context, kept in your cloud with its sources and permissions intact, and a workspace that runs on whichever model you choose, are what make the AI yours.
Sources
Sources
- Sarah Wang, Justin Kahl and Shangda Xu, Leaders, Gainers and Unexpected Winners in the Enterprise AI Arms Race, Andreessen Horowitz, 30 January 2026.
- IBM Institute for Business Value and Oxford Economics, The Calculus of AI Sovereignty, survey of 1,000 senior executives.
- Chris Riley, We Need to Control Personal AI Data So Personal AI Cannot Control Us, and Eliot Bendinelli, Can the Digital Markets Act Free Users' Data in the AI Age?, Tech Policy Press.
- Jeff Bussgang, The 10x Organization: Building Your Company's AI Brain, Flybridge, 30 July 2026.
- Oded Ovadia, Menachem Brief, Moshik Mishaeli and Oren Elisha, Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs, arXiv:2312.05934 (Microsoft, December 2023; EMNLP 2024).
- Aditya Challapally, Chris Pease, Ramesh Raskar and Pradyumna Chari, The GenAI Divide: State of AI in Business, MIT Project NANDA.
- The state of AI in 2026: On the road to ROI, McKinsey & Company, 25 August 2026.
60x builds a company brain and an agent workspace for knowledge-heavy teams. Both run on open-weight models, in Your Cloud or on 60x Cloud.