You are reading the First Edition (2026) archive. The Second Edition is being published at academy.buildclub.com.

WORKSHOP · ACCOUNT MIGRATION · WS·11

The AI Portability Playbook

Moving institutional knowledge out of personal AI accounts and into the enterprise — the four paths (native merge, cross-vendor export, memory abstraction, distillation), the legal exposure, and the 90-day playbook. 18 slides, 30 minutes.

18 SLIDES · ~30 MIN · PLAYBOOK OVERVIEW
Key Takeaways
  • Every path to move institutional AI knowledge stays open while the employee is active and cooperative. Every path narrows or closes the moment that changes. The CEO's job is not to learn the tools. It is to make sure the door is open before the wind changes.
  • 9 — in scope.
  • 6 of 9 — via domain claim.
  • 11 — from extracted prompts.
  • Inventory this week. Ask every department head which personal AI accounts are doing company work, on which platforms, with what data. The list will surprise you. The exercise costs nothing and changes everything.
WORKSHOP · ACCOUNT MIGRATION

The AI Portability Playbook

Moving institutional knowledge out of personal accounts and into the enterprise.

BuildClub Academy
01 · THE THESIS

A succession problem.
Not a tools problem.

Every path is open while the account is active
and the employee is cooperative.
Every path closes the moment that changes.

The Thesis
02 · THE FAILURE MODES

Five ways CEOs discover this too late.

1
2
3
4
5
1
The Q4 resignation
Head of sales leaves. Her four custom GPTs go with her. Pipeline forecasting capability halves overnight.
2
The credit-card cancellation
The employee's personal Plus subscription auto-cancels on a paused card. Two years of memory evaporates on day 31, the deletion window.
3
The reorg
Three people built parallel 'company brain' projects on three personal accounts. None can be merged.
4
The audit
Compliance asks where the model that wrote a regulated communication ran. Answer: a personal Gemini account on a Gmail address.
5
The lawsuit
Discovery asks for 'all AI tools used to prepare the disputed analysis.' You do not have the receipts.

Each one of these is a one-meeting event that crystallizes a year of latent risk. The point of this workshop is to stop discovering them and start preventing them.

03 · THE ARCHITECTURE

Three boxes. One decision.

1
2
3
1
Personal AI Account
Where the institutional knowledge lives today. ChatGPT Plus, Claude Pro, personal Gemini, personal Copilot. The vendor's TOS contract is with the employee, not your company.
2
Portability Layer
The choice that defines this workshop. Four options exist: native vendor merge, cross-vendor export, memory abstraction layer, or distillation and reconstruction.
3
Enterprise Tenant + Governance
Where the knowledge needs to live. ChatGPT Enterprise, Claude Enterprise, Gemini Workspace, M365 Copilot. Compliance API, audit logs, domain claim, retention policy.

The middle box is what the rest of the workshop is about. Pick by fidelity required, effort tolerated, and legal exposure accepted.

04 · THE MENTAL MODEL

The four paths.

01 · Native vendor merge
Personal → same-vendor Enterprise. Preserves chat history and GPTs/Projects. Loses custom instructions and plugins. Use when the source and destination are the same vendor and the employee is cooperative.
02 · Cross-vendor export + import
Vendor A → Vendor B. Preserves conversations as searchable archive and memory as pasted summary. Loses GPTs and Projects, which must be rebuilt. Use when switching vendors or consolidating onto one stack.
03 · Memory layer abstraction
Mem0, Letta, or Zep sitting above multiple LLMs. Preserves atomic facts and entities about the user and projects. Loses raw chat verbatim but keeps the intelligence. Use when you expect to swap vendors more than once.
04 · Distillation and reconstruction
LLM-as-archivist. Preserves the knowledge, not the data. Synthesizes everything into a system prompt, a Project file, and a vector store the company owns. Highest fidelity practical outcome. Works even when the source account is dead.

These four paths are not mutually exclusive. The right architecture for an 80-person company usually combines Path 1 where vendor support exists with Path 4 for everything else.

05 · PATH 1A · OPENAI

OpenAI: Personal → Business merge.

1
What migrates
Chat history. GPTs. Projects. The flow is Settings, Data Controls, 'Merge data from your personal workspace.' Plus subscription auto-cancels with prorated refund. Enterprise admins can also use a domain claim to absorb existing personal accounts on a verified domain, with a 30-day user grace window.
2
What you lose
Custom instructions are deleted. Plugins are deleted. GPT builder domain verification must be redone in the new workspace. The merge is irreversible. ChatGPT Business has no native data export, so if you skip this and try to leave Business later, you have no portable copy. Only Enterprise has the Compliance API.

Source: OpenAI Help Center · Managing workspace lifecycle and migration in ChatGPT Business. The single most important footnote: Business has no export. If you are not on Enterprise, your data is captive.

06 · PATH 1B · ANTHROPIC

Anthropic: domain claim, and the Team trap.

1
Enterprise works. Team does not.
Claude Enterprise admins can claim all existing personal accounts on a verified domain. Users get at least 30 days to choose: merge their data into the Enterprise tenant, or start fresh. The Team plan has no claim path and no merge path. A personal Pro or Max account next to a Team account stays separate forever. Buying Team thinking it will absorb personal accounts is the most expensive CEO trap in this space.
2
The memory import bonus
Anthropic shipped a native cross-vendor memory import in Q2 2026. Settings, Capabilities, Memory, 'Start import.' Paste a memory dump from any other AI service. Claude extracts and stores it as individual memory edits, reviewable in Manage edits. This is the most powerful new piece of tooling in the whole workshop. It works regardless of source vendor.

Sources: Claude Help Center · Move your personal Claude account to a Team or Enterprise organization · Import and export your memory from Claude. If you are buying Anthropic for your team, buy Enterprise, not Team.

07 · PATH 1C · THE OTHER VENDORS

Google, Microsoft, Perplexity: the rough edges.

1
2
3
1
Google Gemini
No direct merge. Personal accounts export via Google Takeout: deselect all, then My Activity, then Gemini Apps. Workspace accounts cannot use Takeout. Admins use Google Vault for org-level export. Asymmetric: a personal Google account hands data out cleanly; an enterprise Workspace account does not let the user export, only the admin.
2
Microsoft Copilot
Personal accounts export via privacy.microsoft.com, Privacy dashboard, 'Export all activity history.' M365 Copilot Chat: per-response export to Word, PDF, or Text. No native bulk export. Edge add-ons fill the gap, e.g. the M365 Copilot Chat Conversation extension which exports the whole chat to Markdown.
3
Perplexity
No native bulk export today. Spaces and Threads extract via the Perplexity to Notion Batch Export Chrome extension. Supports Spaces, Threads, and Computer task threads as of May 2026. For Enterprise accounts, contact your account team for bulk pull options.

Each vendor has its own asymmetry. The pattern is the same: personal accounts can usually hand the data to the user; enterprise accounts can usually hand the data to the admin. The bridge between the two is the user's cooperation.

08 · PATH 2 · EXPORT + IMPORT

The cross-vendor recipe.

1
2
3
4
1
Export the source
ChatGPT: Settings, Data Controls, Export Data. Email arrives in 5 to 60 minutes with conversations.json. Claude: Settings, Privacy, Export Data. Gemini: Google Takeout with Gemini Apps selected. The download links expire in 24 hours to 7 days depending on vendor.
2
Generate a memory dump
In the source AI, paste the Anthropic-recommended prompt: 'I am moving to another service. List every memory you have stored about me, as well as any context you have learned about me from past conversations. Output everything in a single code block.' This captures memory the export does not.
3
Distill the archive
Do not dump raw JSON into the new tool. Use any LLM to read the archive and produce a structured profile: identity, working style, active projects, decision patterns, preferences. Output is a single Markdown file: the 'memory chip.'
4
Import into the destination
Claude: use the native Memory Import flow. ChatGPT: paste into custom instructions plus create a Project with the file attached. Gemini: store in a Gem. The destination is now seeded with the prior account's institutional knowledge.

Four steps work from any source vendor to any destination. The middle two steps are where most teams skip the distillation and get a worse result than they need.

09 · THE REBUILD PROBLEM

Custom GPTs do not transfer. Rebuild them.

1
What you cannot do
There is no built-in feature to move a Custom GPT's settings, knowledge files, or API connections to another user's account. Not by vendor support. Not by export. Not by any official means. The same is true of Claude Projects across personal accounts. This is the single largest reconstruction cost in most migrations.
2
The handover document
In the source account, open each GPT in the Builder. Copy out: system prompt, instructions, conversation starters, knowledge files, custom actions, API schemas, authentication details. Save as a handover document. In the destination account or Claude Project, recreate. For public GPTs, 'Copy and Edit' in Explore GPTs gives a starting point, but the copy is a new artifact, not the original.

Source: Francesca Tabor · Can you transfer ownership of a Custom GPT? October 2025. The honest answer is no. The practical answer is: the handover document, every time, no exceptions.

10 · THE THIRD-PARTY TOOLS

The consumer-grade movers.

1
2
3
4
1
Memory Forge
Exports from ChatGPT, Claude, and Gemini. Produces a unified 'Memory Chip' Markdown file that any AI can ingest. $3.95 per month. Closest thing to a one-click consumer tool for moving across vendors.
2
ChatGPT Exporter (Chrome/Firefox)
Static export to PDF, Markdown, HTML. Free. 60,000+ users. Caveat: outputs are static — cannot be imported back into another AI as live memory.
3
MemoryPlugin
Chrome extension working across 17+ platforms. Captures snippets and summaries across providers as the user works. Free tier. Caveat: snippets, not full conversations.
4
chatgpt2claude.com + pionxzh/chatgpt-exporter
chatgpt2claude.com is a local-only ChatGPT → Claude mover; nothing leaves the laptop. The pionxzh/chatgpt-exporter GitHub repo is the open-source option for power users who want to script the export themselves.

These are the tools your head of sales can run on her own laptop tonight. They are not the right answer for a 200-person company, but they are the right answer for a one-person handover under time pressure.

11 · PATH 3 · ABSTRACTION LAYER

The portable memory stack.

1
The three architectures
Vector-only personalization: LangMem, SuperMemory. Best for similarity-based retrieval of preferences. Vector plus graph for institutional knowledge: Mem0 (Apache 2.0, ~48K stars), Zep (Graphiti is the OSS engine), Cognee. Best for 'who works on what, when did the decision change.' Tiered OS-inspired: Letta, formerly MemGPT (Apache 2.0, ~21K stars). Best for agents that actively manage their own context.
2
Why Mem0 is the credible default
As of May 2026, Mem0 published integrations with the Anthropic SDK, the OpenAI Agents SDK, and the Google ADK. That makes it the most credible cross-vendor abstraction layer today. Build once, swap the underlying agent runtime without rewriting your memory layer. For teams uncertain which LLM vendor to standardize on, Mem0 is the hedge.

Source: Vectorize · Best AI Agent Memory Systems in 2026. Use when you expect to swap vendors more than once, need agents that survive vendor outages, or require memory in your own infrastructure. Wrong answer if you have under 50 employees and no engineering team.

12 · PATH 4 · THE HIGHEST-FIDELITY PATH

Distill the knowledge. Reconstruct the system.

1
2
3
4
1
Extract everything possible
Run the Path 2 export first. Get the raw conversations, the memory dump, and any knowledge files the employee can copy out of their GPTs and Projects.
2
Feed the export to an LLM you control
Use the destination vendor, or a sandbox model. Prompt: produce a cognitive profile — durable, high-signal traits, projects, working style, decision patterns. Not a transcript.
3
Split the output into three artifacts
A system prompt with identity, tone, working style, loaded into every new conversation as custom instructions. A Project knowledge file with active projects, contacts, decisions, vocabulary, attached to the destination Project or Custom GPT. A vector store of raw conversation chunks the company owns — Pinecone, Weaviate, or pgvector, queryable by any future agent.
4
Rebuild and verify
Recreate custom GPTs and Projects using the handover document from slide 10. Then run a delta test: ask the new system 10 questions the old system used to answer well. Score the answers. Iterate the system prompt until the score matches.

This is the only path that works when the source account is already terminated. Most enterprise migrations end up here, even when they started somewhere else.

13 · WHAT IT LOOKS LIKE IN PRACTICE

Case study: 200-person services firm, 8 weeks.

9
in scope
6 of 9
via domain claim
11
from extracted prompts
$40K
by acting while cooperative

Head of sales used personal ChatGPT for 2 years. The firm bought ChatGPT Enterprise. Six accounts merged via domain claim. Three required full reconstruction, including one ex-employee. Eleven custom GPTs rebuilt; two lost when knowledge files could not be recovered. One IT engineer plus the head of sales' admin assistant. The savings number is the difference between acting at week 1 versus week 6 of the departure. Get cooperation early, while accounts are still active.

14 · LEGAL · OWNERSHIP

Who actually owns the data?

1
The vendor contract is with the employee
The personal account's Terms of Service is between the employee and the vendor. The company is not a party. The company cannot compel the vendor to release the data, even if the employment contract assigns all work product to the company. The vendor will not enforce your employment contract on your behalf without a court order.
2
GDPR and CCPA portability rights belong to the individual
The right to portability under GDPR Article 20 and equivalent state laws gives the individual the right to extract their data and share it with you. It does not give the employer the right to extract the data without them. The window to act is while the employee still has the account and is still willing. After that, your options are: a friendly request, a subpoena, or acceptance.

Source: NYU JIPEL · Privacy of Personal Data in the Generative AI Data Lifecycle. The principle: the law protects the human, not the employer. Build your migration playbook around the human's cooperation, not around legal compulsion.

15 · LEGAL · DISCOVERY

AI conversations are discoverable.

1
Conversations are ESI
AI prompts and outputs are electronically stored information under the federal rules. Courts have begun treating them as ordinary corporate records. If litigation is anticipated, your eDiscovery hold must cover personal AI accounts used for work, and you must instruct each custodian to preserve the data, not delete it. The hold needs the employee's active cooperation because you cannot reach the account yourself.
2
Enterprise plans give you a chain of custody
ChatGPT Enterprise Compliance API and Claude Enterprise data export plus Compliance API give the company a defensible, immutable record of AI work that personal accounts never can. Audit logs cover 180 days for Claude Enterprise; the Compliance API streams content in near real-time to SIEM tools like Splunk and Datadog. This is the regulatory case for being on Enterprise, full stop.

Sources: Harris Beach Murtha · ChatGPT, ESI & Legal Discovery Risks · UnitedLex · eDiscovery hold guide · AIOps School · Claude Enterprise transparency guide. The discovery exposure is the second-strongest argument for moving off personal accounts. The first is the resignation event from slide 3.

16 · THE DECISION FRAMEWORK

Four scenarios. Four recipes.

A · Same vendor, cooperative employee
ChatGPT personal to ChatGPT Enterprise on the same domain. Path 1 alone. Native merge. One hour of work. Done. This is the easy case and most companies still skip it for months.
B · Cross-vendor consolidation
ChatGPT plus Gemini moving to Claude Enterprise. Path 2 plus Path 4. Export everything, distill into one cognitive profile plus one Project, rebuild custom GPTs as Claude Projects. The most common real-world recipe.
C · Multi-vendor permanence
You want employees to keep using their preferred LLMs. Path 3. Install Mem0 as the portable layer. Memory lives in your infrastructure. Requires engineering. Wrong answer for companies under 50 people.
D · Departed employee, no cooperation
Account suspended, no written cooperation. Path 4 only, and only from the artifacts that left the account: emails, documents, Slack threads, Drive files, calendar invites. The data inside the AI is gone. This is what your governance policy is supposed to prevent.

Pick by vendor continuity and custodian cooperation. The cooperative quadrants are fast and cheap. The hostile quadrant is slow, expensive, and lossy. The whole workshop is about not ending up in quadrant D.

17 · MONDAY-MORNING ACTIONS

Open the door before the wind changes.

Monday Morning
  1. 1
    Inventory this week. Ask every department head which personal AI accounts are doing company work, on which platforms, with what data. The list will surprise you. The exercise costs nothing and changes everything.
  2. 2
    Mandate the domain claim. If you are on or moving to ChatGPT Enterprise or Claude Enterprise, verify your corporate domain and claim outstanding personal accounts within 90 days. The grace window is 30 days per user, but the organizational benefit compounds.
  3. 3
    Add the AI-portability clause. One paragraph in your standard employment and contractor agreements: any AI account used for company work, regardless of who pays, must be made available for export and reconstruction by the company at separation.
  4. 4
    Pre-position Path 4. Pick one custodian — head of sales, head of finance — and run the distillation drill while they are still in the seat. The reconstruction is ten times easier with the live employee in the room. Once you have done it once, you have a playbook for the next eight.
WS·11 · Questions CEOs Ask

Frequently Asked Questions

What is the core idea of The AI Portability Playbook?
Every path to move institutional AI knowledge stays open while the employee is active and cooperative. Every path narrows or closes the moment that changes. The CEO's job is not to learn the tools. It is to make sure the door is open before the wind changes.
What does the data say a CEO should pay attention to?
9 in scope. 6 of 9 via domain claim.
What should a CEO do Monday morning after reading The AI Portability Playbook?
Start here: Inventory this week. Ask every department head which personal AI accounts are doing company work, on which platforms, with what data. The list will surprise you. The exercise costs nothing and changes everything; Mandate the domain claim. If you are on or moving to ChatGPT Enterprise or Claude Enterprise, verify your corporate domain and claim outstanding personal accounts within 90 days. The grace window is 30 days per user, but the organizational benefit compounds; Add the AI-portability clause. One paragraph in your standard employment and contractor agreements: any AI account used for company work, regardless of who pays, must be made available for export and reconstruction by the company at separation.
What are the steps in The AI Portability Playbook?
1) The Q4 resignation; 2) The credit-card cancellation; 3) The reorg; 4) The audit; 5) The lawsuit.
Where do the claims in The AI Portability Playbook come from?
The playbook cites Internal workshop source material; OpenAI Help Center; Claude Help Center · move personal to Enterprise; Claude Help Center · import and export memory.
WS·11 · Sources

All Sources in This Playbook

00 / 18