// AI Deep Dive
The Switching Cost You Can't See
Lock-in used to live in the integration layer. It's moving into what your assistants remember, and most of that belongs to the seat, not the company.
// Executive Summary
For three years, AI switching costs meant plumbing: APIs, integrations, prompt formats. That problem is shrinking. The new one is piling up quietly inside the memory features every major assistant has shipped, the preferences, project history and inferred profile that make an assistant useful on day 300 and hard to replace on day 301. It sits where most procurement checklists don't look, and on the business plans we checked it belongs to the seat, not the company.
Enterprises already rarely switch. In Menlo Ventures' mid-2025 survey of 150 technical decision-makers, 66% upgraded within their existing provider and only 11% switched vendors. Andreessen Horowitz's CIO survey found agentic workflows are making switching harder still, because the prompts and guardrails get tuned to one model.
Memory is scoped to the person, not the business. OpenAI's business documentation says ChatGPT memories are tied to each individual account and can't be transferred to other users, even inside the same workspace. Microsoft keeps Microsoft Copilot memory in a hidden folder of each user's mailbox, outside Purview retention policies and with no audit log entries.
The portability tools skip the enterprise. Google's memory and chat-history import for Gemini isn't available to Business or Enterprise accounts. And the European guidance on GDPR portability excludes data a provider creates through personalisation or profiling, which describes most of what an assistant remembers.
Treat assistant memory as a company asset: give it an owner, an export path and a home you control, and write all three into the next AI contract you sign. The EU's ban on switching fees lands in January. It won't move your context for you.
// The Hook
Three Hundred Notes Nobody Else Could Read
I run a small fleet of AI agents these days. One handles editing and judgment calls, a couple of others write and review code, and a Mac Mini runs jobs overnight. Over the past few months, one of those assistants built up more than 300 notes about how I work: which newsletter formats are law, which code repos have traps in them, which of my own rules I've contradicted and why. It became the most useful file in my whole operation.
Then I noticed the other agents couldn't read any of it. The notes lived in one vendor's memory folder, in that vendor's layout, loaded by that vendor's tool. Every time a different agent picked up a task, it showed up like a new hire on day one, asking questions the first assistant had answered weeks ago. I was paying for the same education over and over.
So I did what any recovering open source guy would do. I moved the memory somewhere I own. There's now a shared store every agent reads from and writes to, and each vendor's built-in memory is a convenience, not the record. It wasn't glamorous work, and I grumbled the whole way through it.
What bugged me was how familiar it felt. I've spent a couple of decades watching lock-in hide in file formats and proprietary APIs, and I walked straight into a version I didn't recognize at first. I'm one person with a handful of agents. A company with thousands of seats is running the same experiment at scale, mostly without noticing, and that's what the rest of this edition is about.
// The Deep Dive
The Switching Cost You Can't See
Lock-in used to live in the integration layer. It's moving into what your assistants remember, and most of that belongs to the seat, not the company.
The received wisdom on AI vendor risk runs like this: models are commoditizing, the APIs look alike, and a decent abstraction layer lets you swap providers when price or quality shifts. This week's biggest enterprise story fit that frame. Salesforce built its new CRM reasoning model on the premise that it controls the model weights and runs inference entirely inside its own trust boundary, and TechCrunch called it everything the AI labs should fear. Own the model and no vendor can hold you hostage.
That frame treats the model as the thing you'd get locked into, and it misses where the value is piling up. Every major assistant now remembers. ChatGPT draws on past chats, saved memories, custom instructions, files and connected apps such as Gmail. Claude keeps a separate memory for each project. Microsoft Copilot memory, still labeled a preview, is on by default unless a tenant admin turns it off. The more your people use these tools, the more of your company's working knowledge ends up in a store you can't list, can't export in a structured form and, on several business plans, can't see at all. The switching cost has moved from the plumbing to the memory, and it grows every day nobody looks at it.
What Memory Lock-In Is
Memory lock-in is the cost of leaving behind the context an assistant has accumulated (stated preferences, inferred facts, project history, agent run logs) when you change vendors, change plans or lose the employee who built it. Think of it as the difference between a medical chart and a doctor's judgment. Change doctors and you can request your records. What you can't take is everything the old doctor learned by watching you for ten years. AI memory is mostly the second kind, and it shows up at three levels.
Individual. Each employee's assistant memory is a private summary that drifts. OpenAI says memory doesn't retain every detail, ChatGPT decides what's relevant, and what it derives from chat history can change as it updates what's most useful. Anthropic's documented export path for Claude is to ask Claude to write out its memories verbatim, or to view them in settings. That's a prose summary written by the vendor's own model, not a file with a schema.
Team. Memory doesn't pool. ChatGPT Business memories are tied to each individual account and aren't transferable to other users, even within the same workspace. A ten-person account team that has taught its assistants one client's quirks holds ten private copies of that lesson. Project-scoped memory, in Claude and in ChatGPT's project-only memory, keeps contexts apart, which is good for confidentiality and means context is fragmented by design.
Organizational. At company scale, memory sits outside the controls built for everything else. Microsoft documents that Purview retention policies don't apply to Copilot memory, memory actions don't generate audit log entries, and admins can't restrict what information gets added. On ChatGPT Business, turning memory off for the workspace deletes members' existing saved memories, and deleting the workspace deletes them permanently. Admins can switch memory off, and Microsoft's can search it for legal requests. What nobody can do yet is govern what goes in.
How Memory Lock-In Builds
Lock-in doesn't arrive in one contract. It accumulates in layers, and each layer is harder to carry out than the one before.
Preferences. Tone, format, role, pet peeves. Cheap to rebuild, and it's what the consumer switching tools handle: Google's Gemini import works by having you paste a prompt into your old assistant and paste its answer back. A day's loss at most.
Project context. Client history, past decisions, the options already rejected, living in project memories and uploaded files. Weeks to rebuild, and usually rebuilt worse, because nobody remembers what the assistant was told in March.
Workflow and agent tuning. Prompts, guardrails and evaluation work shaped around one model's behavior. Andreessen Horowitz found enterprises now hesitate to switch because changing models can take significant engineering time, with one leader noting all the prompts had been tuned for OpenAI.
Inferred institutional knowledge. What the assistant concluded about your business from thousands of interactions. This is where the law runs out. The EU's Article 29 Working Party guidance on GDPR portability says data created by a personalisation or recommendation process is inferred and not covered by the right to data portability.
Here's how the four assistants most enterprises run stack up on the question that matters: where the memory lives and what stops it from moving.
Assistant | Where memory lives | The portability gap |
|---|---|---|
ChatGPT | Can't be shared or transferred to colleagues; disabling it workspace-wide deletes it | |
Claude | ||
Gemini | Import tools exclude Business and Enterprise accounts, and aren't yet available in the EEA, UK or Switzerland | |
Microsoft Copilot | Outside Purview retention and audit; admins can't restrict what's added |
The honest counterargument is that vendors are racing to make switching easier, not harder. Google built its import tools to win users from rivals, Anthropic ships memory import and export, and the Data Transfer Initiative called early 2026 a turning point for AI portability. That's real progress. Look at who it's for, though: individual consumers switching to a competitor, by copying and pasting a summary one model wrote about them. Those are customer-acquisition tools. Nothing in them moves a project's context, an agent's run history or a team's shared knowledge from one enterprise tenant to another. The same DTI post warns about a reversion to historical patterns of trapping users in online services by their own data, and DTI's latest analysis argues the portability question is shifting from the model to the harness around it, the tools, memory and workspace an agent runs inside.
Regulation is further along on the plumbing than on the memory. The EU Data Act has applied since 12 September 2025, requires software and platform providers to offer open interfaces and export data in a commonly used, machine-readable format, and removes switching charges, including data egress fees, from 12 January 2027. Whether an assistant's inferred memory counts as exportable customer data under those rules hasn't been tested, and the GDPR guidance on inferred data points the other way. Plan as if no regulator is coming to carry your context across for you.
How to Implement Company-Owned Context
The fix isn't to switch memory off. It's to make sure the record of what your company knows lives somewhere the company controls, with vendor memory as a convenience layered on top.
Phase 1: Find the memory you already have (this quarter).
Most companies have never inventoried it. Start there.
List every assistant with memory turned on, by plan and by team. Remember that Copilot's is on by default, so the answer is probably "more than we approved."
Send each vendor four questions in writing: where memory is stored, who in your company can see it, how it exports and in what format, and what happens to it when a seat is removed or the contract ends.
Set a memory policy for regulated work. OpenAI keeps improved memory off by default in regulated workspaces and outside its HIPAA business associate agreement; mirror that caution for anything touching customer or patient data.
Phase 2: Move the record into files you own (next two quarters).
Vendor memory should be a cache. The source of truth should be yours.
Put durable context (client briefs, style rules, product facts, decision logs) in a company-owned repository any assistant can read through project uploads or connectors.
Default to project-scoped memory where it exists, so context attaches to a business unit rather than to one person's account.
Require every agent build to log its prompts, guardrails and run history to your storage, not only the vendor's console.
Phase 3: Contract for the exit (at the next renewal).
Portability you haven't negotiated is portability you don't have.
Write memory export into the contract, in a structured, documented format. The Data Transfer Initiative's principles are a ready-made spec: download and direct transfer in a structured, machine-readable format, with personal data used to customize the service explicitly in scope.
Ask the sovereignty questions alongside it: whose weights, where inference runs, where memory is stored.
Run one switching drill a year. Load a single team's context into a second assistant and write down what broke.
Four factors separate the companies that get this right:
A named owner for AI context, usually knowledge management or the CIO's office, rather than each user.
A company-owned source of truth that every assistant reads from.
Exit terms that name memory and agent histories explicitly.
A switching drill on the calendar, run before anyone needs it.
Common Missteps
Governing memory like a user preference. It lives in personal settings, so it gets treated like a wallpaper choice. At scale, it's institutional knowledge sitting outside your retention and audit controls.
Assuming the data export covers it. A chat-history download is the chart, not the doctor. The inferred layer, what the assistant concluded, is exactly what GDPR portability guidance excludes and what vendor exports compress into prose.
Banning memory outright. Switching it off throws away the productivity it was built to deliver, and on ChatGPT Business it deletes what members already saved. Govern it. Don't just kill it.
Owning the model and calling it sovereignty. Controlling weights and inference answers one question. If your context still lives in someone else's assistant, the lock-in just moved up a layer.
Business Value
ROI considerations:
Every rebuild of project context after a vendor change, reorg or resignation gets paid for twice: once to teach the assistant and once to reteach its replacement. A company-owned context store is paid for once.
Menlo's 11% switch rate tells vendors they can count on renewal. A tested switching drill and a portable context store are what make an exit credible, and a credible exit is what earns a price concession.
Compliance exposure shrinks when memory lives where retention and audit already apply, rather than in a hidden mailbox folder that Purview retention ignores.
Competitive implications: Companies that own their context get to treat models the way the commoditization story promised: interchangeable parts, picked on price and quality every year. Companies that don't will stay with whichever vendor holds their memory, at whatever price that vendor sets at renewal. As model quality converges, context becomes the differentiator. It should be yours.
// Key Takeaways
Govern memory like records, not preferences. Copilot memory sits outside Purview retention and audit, and ChatGPT Business memory is tied to individual accounts. That's institutional knowledge with no institutional owner. Give it one.
Make vendor memory a cache, not the record. Vendor exports are summaries a model wrote, and GDPR portability guidance excludes inferred data. Keep durable context in files your company owns and feed it to whichever assistant you run.
Write memory into every AI contract. Ask where memory lives, how it exports, in what format and what happens at contract end, and use the Data Transfer Initiative's principles as the spec. Ask about weights and inference in the same meeting.
Run a switching drill before you need one. Only 11% of builders switched vendors in Menlo's survey, and vendors know it. Test one team's move every year so leaving is a capability, not a bluff, and do the first one before the EU's switching-fee ban takes effect in January.
// What This Means for Your Planning
In planning terms, your AI vendor decision isn't only a model decision anymore. It's a decision about where your company's working memory will live for the next several years, and most companies are making it by default, one personal settings page at a time.
The boardroom assumption to challenge is that portability is handled because the models are interchangeable. The models may well be. The context isn't, and the switching tools vendors have shipped so far were built to win individual users, not to move an enterprise tenant.
For the next budget cycle, I'd fund three things before the next license expansion. A small line for a company-owned context store and the people who curate it. Memory terms in every AI renewal. And one switching drill before year-end, with counsel reading the Data Act's January switching provisions against your assistant contracts if you operate in Europe.
Here's the question to bring to the next planning meeting: if our three most AI-fluent people resigned tomorrow, how much of what their assistants know about our business would we lose, and would anyone in this room know where to look?
Your AI Sherpa,
Mark R. Hinkle
Founding Publisher, The AIE Network
Follow me on LinkedIn
If you want to get in contact or give me feedback, reply to this email. I read every single one of them.
