An AI writing workflow has serious vendor lock-in when leaving would mean losing source evidence, prompts, review history, team access, automation, or the ability to reproduce accepted output. Check exit paths before adoption and retest them as the process grows. A download button is not enough if the export omits structure or cannot be used elsewhere.
Examine export completeness
Ask for a documented export of prompts, templates, source records, drafts, comments, versions, approvals, tags, user roles, settings, and audit history. Run a small export and inspect the actual files. Check encoding, timestamps, identifiers, relationships, attachments, and whether the format has a published specification.
Google's data export guidance explains that users can create an archive to keep records or transfer data to another service. It also says downloading data does not delete it from Google's servers. Apply both lessons: test portability separately from deletion, and document the provider's process for each.
A healthy exit path should not depend on screenshots or manually copying hundreds of records. If an export is readable only by the original service, require a conversion plan and test it before scale makes the problem expensive.
Separate workflow assets from vendor behavior
Keep controlled copies of prompts, rubrics, style rules, test cases, source policies, disclosure rules, and approval criteria in portable formats. Record proprietary features that may not transfer, such as hidden system instructions, integrated research, provider-specific history, fine-tuned behavior, or a closed template marketplace.
The NIST AI Resource Center supports AI risk-management operations with testing, evaluation, verification, and validation resources. Use a fixed acceptance set to compare the current service with an alternative. Measure source fidelity, unsupported claims, privacy handling, originality, format, review routing, time, and cost.
If acceptable output cannot be reproduced outside one provider even with the documented assets, classify that dependence explicitly. Decide whether the benefit justifies the exit risk rather than hiding it in a renewal decision.
Review ownership and integrations
Check who owns the account, billing, workspace, stored files, source connectors, publishing destination, domain, and automation credentials. Microsoft Learn's cloud-flow sharing guidance distinguishes flow owners from embedded and other connections. That is a useful warning: transferring a workflow owner does not prove every token and destination can transfer.
Look for personal accounts controlling business work, integrations that cannot be reauthorized, service accounts with undocumented owners, and secrets stored inside prompts or scripts. Require a backup owner and a credential-rotation procedure.
Price the exit, not only the subscription
Estimate export time, conversion work, retraining, lost history, parallel operation, quality validation, legal review, storage, and downtime. Review data-retention and deletion terms, rate limits, export limits, account closure, and support for bulk migration. A low monthly price can coexist with a high exit cost.
Watch for contract or technical changes that narrow export, remove access to history, change ownership terms, or increase dependence on proprietary features. Keep acceptance tests and portable assets current so an exit is possible before a crisis.
Run a switching drill
Quarterly or before renewal, export a representative project, import or reconstruct it in a neutral workspace, and run the acceptance set with no live publication. Verify sources, versions, comments, approvals, attachments, and the ability to recover a past decision. Record omissions and time spent.
Warning signs include incomplete exports, undocumented formats, no bulk access, missing source provenance, inaccessible review history, personal-account ownership, irreplaceable integrations, unclear deletion, and no tested alternative. Lock-in is manageable when it is visible, priced, accepted by the owner, and backed by a working exit plan.
Sources
- How to download your Google data — Google
- NIST AI Resource Center — National Institute of Standards and Technology
- Share a cloud flow — Microsoft Learn