Avoid AI vendor lock-in by keeping three things outside any single tool: your data (export it on a schedule), your prompts and instructions (store them as plain documents you own), and your automations (build them on connectors and standards more than one vendor supports). Then rehearse an exit once a year so you know it works.
The trap is rarely the AI model itself, which you can swap in an afternoon. It's where the work ends up living. A team that has spent a year inside one assistant has chat history, uploaded files, carefully tuned instructions and a handful of automations wired to it, and some of those can't be exported at all. ChatGPT Business, for example, has no self-serve export of workspace chats, which most owners only discover when they try to leave.
What actually gets trapped when you commit to one AI tool
Lock-in builds up in four layers, and each one needs a different fix. It helps to name them before you look at any export button.
- Conversation history and uploaded files. The drafts, analyses and back-and-forth that sit in chat threads. Most of it is disposable, but some threads hold the only copy of a decision, a pricing calculation or a client summary.
- Instructions and context. Custom instructions, Projects, Gems, custom GPTs, memory and the reference files attached to them. This is the layer people underestimate: a tuned set of instructions can represent weeks of trial and error, and it often lives only inside one vendor's settings screen.
- Automations and integrations. The Zapier or Make steps that call a particular AI, the chatbot on your website, the connector that lets the assistant read your inbox. Changing vendor here means rebuilding and re-testing, not just exporting.
- Admin and habits. Seats, sign-in, sharing rules and the muscle memory of your team. These don't block a move, but they add days to one.
A quick way to see your own exposure: open each AI tool your business pays for and ask, "If this closed in 90 days, what would I lose that exists nowhere else?" If the honest answer includes instructions or files you couldn't recreate from memory, that's your lock-in.
Export routes for ChatGPT, Claude and Gemini in September 2026
Export options differ sharply between personal plans and business workspaces, and in one case the business plan is the less portable one.
| Tool and plan | Self-serve export? | Who runs it | The catch |
|---|---|---|---|
| ChatGPT Free, Plus, Pro | Yes, from data controls in settings | Each user | You get files to read and search, not chats another tool can reopen |
| ChatGPT Business | No workspace chat export | Nobody, self-serve | Share links or copy-paste are the only routes for individual threads |
| ChatGPT Enterprise | No standard export; admins have a compliance route for logs | Admins | Logs are reported to be kept only 30 days unless you archive them |
| Claude Free, Pro, Max | Yes, Settings, then Privacy, then Export data | Each user | The download link expires after 24 hours; not available in the mobile apps |
| Claude Team, Enterprise | Yes, from the organisation's data and privacy settings | Primary Owner only | Members can't export their own chats; deleted items are not included |
| Gemini (personal Google account) | Yes, through Google Takeout | Each user | Chat history and Gems sit under different Takeout entries, so tick both |
Two practical consequences. First, if you are about to move a team from individual ChatGPT Plus accounts onto ChatGPT Business, run each person's export before the move, because afterwards the workspace won't give you one. Second, on Claude Team the export sits with one person, the Primary Owner, so make sure that role belongs to the business owner rather than whoever happened to set the account up. The wider question of who keeps what when someone leaves is covered in who owns AI chat history when an employee leaves.
Keep instructions in a document, not a settings box
The single most useful portability habit costs nothing: write every set of instructions you give an AI as a plain document that lives in your own shared drive, and paste it into the tool from there. The tool's settings screen becomes a copy, not the original.
Here is how that change looks for a café that built a menu-description assistant.
Before: The instructions lived only inside a custom GPT the owner built in 2025. They had been edited a dozen times after bad outputs ("stop saying 'artisanal'", "always mention if a dish is vegan"), and none of those edits were written down anywhere else. OpenAI has since announced that custom GPTs stop running on 11 December 2026, with a migration that turns each GPT into a plugin whose instructions become a skill. The migration helps, but the café still has one copy of its rules, held by one vendor.
After: The same rules sit in a one-page instruction card in the café's shared drive, with a change log at the bottom. The GPT, and whatever replaces it, gets a pasted copy. If the café moves to a Claude Project or a Gemini Gem (becoming a skill from November 2026), setup takes ten minutes. The comparison of GPTs, Claude Projects and Gemini Gems covers which container suits which job; the card works in all three.
A filled-in card for that café, which you can adapt:
INSTRUCTION CARD: Menu descriptions
Owner: café manager Last tested: 2 Sep 2026 (ChatGPT, Claude)
Purpose
Write menu and chalkboard descriptions for new dishes and drinks.
Always
- 20-35 words per item, present tense, no exclamation marks.
- Name the main ingredients in the order a customer would taste them.
- Flag dietary status at the end: (v), (vg), (gf) - only if the
recipe sheet confirms it. If unsure, write [CHECK DIETARY].
- British spelling.
Never
- Words: artisanal, indulgent, elevated, curated, mouth-watering.
- Invent provenance ("locally sourced") unless the recipe sheet says so.
Input you will get
Dish name, recipe sheet text, price.
Output format
Line 1: name. Line 2: description. Line 3: dietary tags.
Test cases (paste these after any tool change)
1. Oat-milk cardamom latte 2. Smoked mackerel on rye 3. Plum frangipane
Change log
14 Jul 2026 - added [CHECK DIETARY] after a gf label went on a dish with rye.
The test cases at the bottom matter as much as the rules. They turn "does the new tool behave the same?" from a feeling into a five-minute check. If you want to go a step further, the instruction layer now has an open format: the Agent Skills specification published in December 2025 describes a skill as a folder with a plain Markdown file of instructions, and it has been adopted by OpenAI's Codex and several other platforms. You don't need to write in that format today, but a Markdown card is already most of the way there. For organising a whole team's collection, see how to build a shared prompt library.
Running the same card in a second tool: a sample comparison
A portable instruction card should produce usable output in any mainstream assistant. Test it by pasting the card and one test case into a second tool you don't normally use. An illustrative result for test case 2, smoked mackerel on rye:
Tool A output
Smoked Mackerel on Rye
Flaky smoked mackerel, sharp pickled cucumber and dill crème fraîche
on dark rye, finished with lemon and cracked pepper.
[CHECK DIETARY]
Tool B output
Smoked Mackerel on Rye
Rich smoked mackerel with crème fraîche, dill and pickled cucumber
on toasted rye bread, brightened with a squeeze of lemon. A proper
lunchtime treat.
(gf)Tool B broke two rules: it added a closing sentence that pushed the length past 35 words, and it labelled a rye dish gluten-free, which is the exact mistake the change log records. The fix is to tighten the card, not to blame the tool: move the dietary rule to the top, add "rye, barley and wheat are never gf", and add a worked example of the correct output. After that change both tools passed all three test cases. This is the whole point of the exercise: you find the ambiguities in your own instructions while the stakes are low.
Automations: make the AI step the easiest part to swap
In an automation, the AI is usually one step among several: a form arrives, the AI classifies or drafts, and the result goes to a spreadsheet, inbox or CRM. Lock-in creeps in when the AI step is built with a vendor-specific feature that has no equivalent elsewhere, or when the prompt text exists only inside that step.
Consider a driving school that routes website enquiries. The zap reads the form, asks an AI to sort the enquiry into "first lesson", "test booking", "intensive course" or "other", then emails the right instructor. Built one way, the prompt sits inside the AI step and nobody has a copy. Built the portable way:
- The classification prompt is an instruction card in the shared drive, with ten real enquiries as test cases and the expected category for each.
- The AI step returns one word from a fixed list, so the steps after it don't care which model produced it.
- The step uses a general connection (the automation platform's own AI step or a standard API connection) rather than a feature only one assistant offers.
Swapping models in that design is a 30-minute job: change the connection, paste the card, run the ten test enquiries, compare. The same thinking applies to connecting assistants to your own systems. The Model Context Protocol, the standard many assistants now use to reach tools and data, was handed to the Agentic AI Foundation under the Linux Foundation in December 2025, so a connector built on it is less tied to one company's roadmap than a bespoke plug-in.
If you call models through an API, keep the request code simple. Many providers, and local runners such as Ollama and LM Studio, accept requests in the same widely copied format, which means changing provider can be as small as a new model name, key and web address. The part that doesn't move automatically is behaviour, so the test cases come with you.
Supplier exits that already happened in 2026
These aren't hypothetical. Each of the following gave customers a deadline, and each rewards a different habit.
- Clockwise, the AI calendar tool, shut down on 27 March 2026 and deleted user data rather than transferring it. Lesson: an export you haven't taken before the closure date may not exist afterwards.
- Sora: OpenAI closed the Sora web and app experiences on 26 April 2026 and the API on 24 September 2026. Lesson: even a product from a large vendor can go, so keep generated assets and the prompts that made them in your own storage.
- HubSpot Breeze agents: five agents were sunset on 23 July 2026. Existing installs keep working but can't be newly installed. Lesson: document what an agent does, step by step, so you can rebuild it as a workflow if it's withdrawn.
- Custom GPTs stop running on 11 December 2026. Lesson: instructions and knowledge files you hold yourself make any migration a copy-and-paste job.
- Kajabi Creator Studio shuts down on 9 November 2026. Lesson: AI features bundled into a platform can be removed while the platform carries on.
For the broader question of what to check before you rely on a smaller supplier, what happens if your AI vendor shuts down has the due-diligence list.
Contract and settings checks that keep the door open
Most portability is decided before you sign. Five checks are worth ten minutes on any AI tool that will hold business content:
- Is there a self-serve export, and who can run it? Test it in the trial. If only an admin role can export, note who holds that role.
- What format comes out? JSON, CSV, HTML or Markdown are fine. A proprietary format, or "contact support", is a warning.
- What happens to your data when you leave? Commercial terms usually say deletion follows termination. Anthropic's data processing addendum, for example, commits to deleting customer data within 30 days of termination, or returning it on request. That is good privacy practice and also a deadline: export before you cancel, not after. The data processing agreement checklist covers the rest of those clauses.
- How much notice do you get when a feature is retired? Look for a stated notice period in the terms or the vendor's past behaviour. The 2026 exits above gave anything from weeks to several months.
- Are you on monthly or annual billing? Monthly costs more but lets you leave within weeks. For a tool you are still testing, the premium buys portability.
A photography studio's first exit drill, costed
To see what this costs in practice, take a three-person photography studio, used here as an illustration, that has run on ChatGPT Business for 18 months at two annual seats plus one monthly seat. The owner wants to know whether a move to another assistant would be painful, without actually moving.
Week 1, inventory (about 2 hours). The owner lists everything the team uses ChatGPT for: client enquiry replies, shoot-day timelines, gallery captions, invoice chasers and a pricing calculator thread. There are 4 Projects, 2 custom GPTs, 23 saved snippets of instructions scattered across the team's notes apps, and 3 zaps that call an AI step.
Week 2, rescue the instructions (about 3 hours). Each GPT and Project gets an instruction card in the shared drive with three test cases. The 23 snippets collapse into 9 cards, because many were near-duplicates. The pricing calculator thread turns out to hold the only record of how package prices were set in 2025, so that reasoning goes into a spreadsheet.
Week 3, the dry run (about 2 hours). The owner pastes each card into a second assistant on an individual trial and runs the test cases. Seven of nine cards pass first time. Two fail: the caption card relies on examples stored as files inside a Project, and the enquiry card assumes the model knows the studio's package names. Both are fixed by adding the examples and the package list into the cards.
Week 4, automations (about 1 hour). The three zaps each have their prompt copied into a card. One zap uses a feature with no direct equivalent elsewhere; the owner notes it as "rebuild needed, about 2 hours" rather than fixing it now.
Total effort: roughly 8 hours of the owner's time, spread over a month. The output is a folder of 9 tested cards, a written list of what can't be exported (workspace chat history) and a two-hour estimate for the one awkward zap. If a price rise or a retired feature arrives next year, the studio can move in a week instead of a quarter. If you'd rather work through this inventory with someone, it's the kind of mapping an AI implementation consultation covers; get in touch if that would help.
Scoring how portable your setup is today
Score each line 0, 1 or 2 and add them up. It takes five minutes and tells you where to spend your first hour.
| Question | 0 points | 1 point | 2 points |
|---|---|---|---|
| Where do instructions live? | Only in tool settings | Some copied elsewhere | All in owned documents |
| Do cards have test cases? | None | A few | Every card |
| Last export of chats and files | Never | Over six months ago | This quarter |
| AI steps in automations | Vendor-specific features | Mixed | Standard connections, fixed outputs |
| Admin and export roles | Unknown or ex-staff | One person, undocumented | Owner holds it, written down |
A score of 8 to 10 means you can switch whenever the numbers favour it. Between 4 and 7 is typical after a year of enthusiastic use; start with instruction cards, because they give the most portability per hour. Below 4, do an export this week before anything else, since that's the one step you can't take retrospectively.
The studio in the example scored 3 before its drill and 9 afterwards. The one point it couldn't win back was chat history on a business workspace, which is exactly why the habit of keeping important conclusions in documents, not threads, matters more than any export button.
Questions about moving between AI tools
Can I move my ChatGPT history into Claude or Gemini?
Not as a working chat history. Exports come out as files you can read and search, not conversations another tool reopens. What does transfer is context: Claude has an experimental memory import that takes a summary you generate in another assistant, and your instruction documents paste straight into Projects or Gems. Treat old chats as an archive and rebuild the useful context in the new tool.
Is using the API instead of the chat app less lock-in?
Usually, yes. Many providers accept the same request format, so an automation built on the API can often switch model by changing a model name, a key and a web address. The work that stays yours is the prompt and the test cases. Budget a few hours to re-test outputs, because a new model follows the same instructions slightly differently.
How often should a small business export its AI data?
Quarterly is enough for most small teams, plus once before any plan change, cancellation or staff departure. Exports are point-in-time, so anything created after the last one is only in the vendor's system. Put the date in the calendar alongside other backups, and store the files somewhere the business controls, not in one person's downloads folder.
Does lock-in matter if I'm happy with my current AI tool?
It matters less day to day and a lot at the worst moment: a price rise, a retired feature or a supplier closing. Several AI products were withdrawn in 2026 with a few months' notice. Keeping instructions in documents and exporting on a schedule costs an hour or two a quarter, which is cheap insurance even if you never switch.
Further reads
- How to Evaluate an AI Software Vendor: A Small Business Scorecard — A scorecard for judging a vendor before you commit, portability included.
- AI Software Contracts: Auto-Renewals, Price Rises, Notice Periods — The renewal and notice terms that decide how easily you can leave.
- Zapier vs Make vs n8n for AI Automation: Which Fits Your Business? — Pick an automation layer that lets you swap the AI model underneath.
- ChatGPT Plus vs ChatGPT Business: Which Plan Does a Team Need? — What changes, including exports, when you move a team onto Business.
- Is Claude Team Worth It for a Small Business? — Weigh Claude Team, which lets the owner export organisation data.
- Open-Source vs Paid AI Models: What Small Businesses Should Know — When an open model you host yourself is the most portable option.
- Questions to Ask an AI Vendor Before You Sign Anything — The questions to put to an AI vendor before, during and after the demo, with what good and walk-away answers sound like and a scoring sheet.
- Per-Seat SaaS Fees vs One Custom Tool: Five-Year Costs Compared — Five-year costs of per-seat software against a custom AI tool, year by year, with a break-even seat count and the upkeep costs quotes leave out.
- How to Prepare Your Business Data for AI, Step by Step — Eight steps from scattered spreadsheets to data an AI tool reads correctly, with a photography studio's 1,400 client records as the example.
- How to Review an AI Tool After 90 Days: Keep, Fix or Cancel — An evidence pack, a ten-point scoring sheet, a worked split verdict for four Copilot seats and a checklist for cancelling without loose ends.
- How to Run a Pre-Mortem Before You Launch an AI Project — A facilitator script, a 75-minute agenda and AI-specific failure prompts for finding what will sink your AI project before it goes live.
- What Is an API? Why It Matters When You Buy Software — A plain-English explanation of APIs for people buying software, with the vendor questions and pricing traps that decide whether tools can connect.
- Restaurant AI Tools: 12 Questions to Ask Before You Sign Up — Twelve written questions for any restaurant AI vendor, with red flags, a scoring sheet and five test calls to make before you commit.
- AI Culling vs Outsourced Editing: Which Saves Photographers More? — AI culling costs $10-$18 a month; outsourced editing costs $0.20-$0.50 an image. Work out which buys back more of your hours per dollar.
- Pet Grooming Software With AI: Features and Prices Compared — What the AI in MoeGo, Teddy, Groomify, DaySmart Pet and Gingr actually does, what each costs a year, and a 14-day trial script for groomers.
- Salon and Spa Software With Built-In AI: What to Compare — Five tests that separate useful salon and spa AI from a feature list, with September 2026 prices for Vagaro, GlossGenius, Boulevard, Mangomint and more.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help pages on exporting ChatGPT data and custom GPT retirement (as reported July 2026); Anthropic Claude help pages on exporting individual and organisation data and on memory import; Anthropic Data Processing Addendum; Linux Foundation announcement of the Agentic AI Foundation; Agent Skills specification; vendor shutdown notices for Clockwise, Sora and HubSpot Breeze agents.