To move a company from ChatGPT to Claude, assess the workflows and access rules before transferring content. Inventory what employees depend on, establish what can be recovered, and test its proposed replacement. A successful chat or memory import isn’t evidence that the company’s integrations, permissions or API applications still work.
This guide is for IT and engineering leaders planning an organizational move, not an individual switching personal assistants. If you’re still choosing a platform, start with our ChatGPT Enterprise vs Claude Enterprise comparison.
Vendor documentation checked October 16, 2026. The assessment method below is our recommendation, not a completed customer migration or a claim of automatic conversion.
Separate three migration scopes
The word “ChatGPT” can hide several purchases and technical systems. Keep these scopes separate in the inventory and statement of work.
| Scope | What to inventory | What the pilot must establish |
|---|---|---|
| Employee workspace | Users, projects, custom GPTs, instructions, knowledge files, sharing and connected systems | The intended employees can complete the workflow without gaining access to restricted information |
| API application | Application code, model calls, tools, service accounts, stored resources, usage and billing | The application still meets its behavior, reliability and cost requirements with the proposed provider |
| Coding tools | Repositories, instruction files, skills, hooks, MCP configuration and device policies | The proposed setup preserves required development and security behavior |
OpenAI documents API projects as separately managed resources. They have their own members, service accounts, keys and usage controls. ChatGPT and API billing are separate, too. Changing employee subscriptions doesn’t migrate an application backend.
For the coding-tool branch, use the existing Cursor-to-Claude Code mapping and managed settings guide. Don’t treat a workspace pilot as proof those controls have been tested.
Build an inventory with an owner for every workflow
Start with one department that has recurring, measurable work. Ask its users to identify the tasks they would lose if their current workspace disappeared tomorrow.
Record these fields for each task:
- Business purpose: the input, expected output, frequency and person who accepts the result.
- Source assets: approved documents, instructions, templates and their authoritative locations.
- Dependencies: connected applications, tool actions, scheduled work and API services.
- Access: who can use, edit and share the workflow; which account reaches downstream data.
- Recovery evidence: what was actually collected, when, by whom and with which permission.
- Target and gap: the proposed destination, required rebuild, unresolved limitation and accountable owner.
Record unknowns explicitly. A list of GPT names isn’t an inventory of their behavior. An administrator’s list of connected apps doesn’t explain which employee workflows depend on them.
For example, a sales-proposal assistant may depend on approved pricing files, a CRM connection and a human sign-off. Moving its writing instructions preserves only one part of the task.
What can you export from ChatGPT Enterprise?
Don’t plan an Enterprise migration around the personal-account Export data button. OpenAI’s export guidance says self-service exports aren’t available in Business or Enterprise workspaces. Ask the workspace owner to establish an authorized recovery route before promising an archive or cutover date.
For eligible Enterprise workspaces, the Compliance Platform exposes logs and metadata through permissioned access. Its documentation specifies a 30-day retention window for the Compliance Logs Platform. That isn’t a promise that every workspace asset is recoverable for 30 days. Don’t treat compliance collection as a complete backup or a Claude import format.
Use the following table as an assessment worksheet, not as a guarantee that every item can be exported.
| Asset | Recovery evidence to collect | Proposed treatment |
|---|---|---|
| Instructions and templates | An owner-approved copy and a working example | Adapt and test against the same acceptance criteria |
| Knowledge files | The authoritative original, version, owner and permitted audience | Rebuild approved target knowledge from source files; verify retrieval and access |
| GPT actions and connectors | Endpoint or integration specification, authentication owner and allowed operations | Rebuild or replace the connection; test both permitted and forbidden actions |
| Chats and compliance records | Actual coverage, timestamps, authorized collection method and retention requirements | Keep an approved archive where needed; don’t promise executable history in the new workspace |
| Roles and sharing | Current membership and effective access, including downstream accounts | Configure target permissions and verify them independently |
| API resources and credentials | Application-owned configuration, resource inventory and secret-management references | Scope a separate application change; provision target credentials without copying secrets into the assessment |
Mark each asset as collected, needs reconstruction, archive only, or unavailable/unverified. These labels describe evidence, not completion. A collected instruction still needs testing in its target environment.
If access or retention prevents recovery, log the gap and ask the owner whether the task can be rebuilt from approved originals. Don’t bypass workspace restrictions or delete the source to force a switch.
Map custom GPTs to a workflow, not a product label
A custom GPT can combine instructions, files and external actions. Those pieces need separate decisions. Don’t assume “GPT becomes Project” is a complete mapping.
Claude’s project documentation describes project instructions, uploaded knowledge and sharing permissions. These can be candidate destinations for part of a workflow. They don’t, by themselves, establish equivalent action behavior or access to a connected business system.
Likewise, the documented memory import can import selected context. Anthropic calls the import experimental, so verify what was actually retained. It’s not evidence of a complete company-workspace migration. Check availability in the actual plan and organization configuration.
For each mapping, record the old behavior, proposed replacement, evidence and remaining gap. If a feature has no verified equivalent, keep that limitation visible in the buying decision.
Prove one workflow before expanding
Choose a task with a named owner, approved source data and a result someone can reliably judge. Avoid starting with an irreversible financial or operational action.
An illustrative first pilot is a proposal draft built from a test CRM record and an approved pricing document. A reviewer checks the output before anything reaches a customer.
| Test | Evidence to keep |
|---|---|
| Same task, same approved inputs | Source versions, configuration, outputs and reviewer decisions from both systems |
| Allowed and denied users | Results for an employee who should have access and one who shouldn’t |
| Removed access | Behavior after changing workspace and downstream permissions |
| Missing or outdated source | Whether the assistant flags the problem instead of inventing a value |
| External action | Which identity executes it, required approval and the resulting system record |
| Operating effort | Setup, correction, review and administration time, plus measured usage cost |
Set pass conditions before running the comparison. Don’t infer success from a fluent answer or a cheaper model call. Count rework and failed attempts, and keep the same business acceptance standard.
A small pilot screens for obvious gaps; it doesn’t certify security or predict every future workflow. Expand only after the owner accepts the result and unresolved risks have a documented treatment.
Choose an assessment provider by its deliverables
Ask a provider to show the shape of its output before agreeing to a wider rollout:
- An inventory tied to actual assets, owners and collection evidence.
- A mapping that separates demonstrated equivalents from proposed rebuilds and unsupported features.
- A data-flow and access review covering model providers and connected systems.
- A reproducible pilot with acceptance criteria, failures and remediation owners.
- A rollout estimate that includes administration, training, overlapping subscriptions and rollback arrangements.
Ask which steps are automated, manual or unavailable. A scanner can help discover configuration; it doesn’t prove that the target workflow behaves correctly.
We build Jarvis at ASCENDING. If you’re evaluating Jarvis Crosswalk for this work, confirm the current collection coverage, assessment outputs and delivery scope with our team. Don’t assume automatic migration or a tested write path into Claude. Discuss your assessment scope with ASCENDING using the inventory and unresolved gaps above.
Keep the source environment available under the agreed retention and access policy until the pilot is accepted. Decide who can stop the rollout, restore the old workflow and approve retirement.
References
- OpenAI: Managing projects in the API platform
- OpenAI: Managing billing for ChatGPT and the API platform
- OpenAI: Exporting your ChatGPT history and data
- OpenAI: Compliance Platform for Enterprise and Edu customers
- Anthropic: Create and manage projects
- Anthropic: Import and export memory
Questions about a company ChatGPT-to-Claude migration
How should a company start moving from ChatGPT to Claude?
Inventory the workflows, owners, source files, instructions, integrations and permissions you need to preserve. Separate employee workspaces from API applications and coding tools. Test one representative workflow before committing to a company-wide move.
Can employees export a ChatGPT Enterprise workspace from Settings?
OpenAI says self-service data export is unavailable in Business and Enterprise workspaces. Ask the workspace owner about authorized access. Eligible Enterprise compliance access is not a complete workspace backup or a ready-to-import Claude package.
Can custom GPTs move directly into Claude?
Don't assume a one-click transfer. Recover approved instructions and source assets where available, then map the workflow to the appropriate Claude configuration. Treat actions, authentication, sharing and tool behavior as rebuild-and-test items unless a supported import has been demonstrated.
Does importing memory move the whole workspace?
No. A memory import can carry selected context, but you must verify what was retained. It doesn't establish that company projects, integrations, permissions or API applications have been transferred. Check the exact feature and plan before using it.
What should an AI migration assessment deliver?
Ask for an asset inventory, a source-to-target mapping, an access and data-flow review, a tested pilot, and a costed rollout plan. Each unresolved gap should have an owner and an acceptance condition. These are recommended assessment outputs, not a promise that a particular tool automates them.
