Launched:self-serve migrations intoSuperhuman Docs (Coda)
Try it now
01Agent-first
Runs where you already work
Plug it into Claude, ChatGPT or Cursor. Describe the move in plain English; the agent runs it.
02Engineer-led
Our production engine, unlocked
The pipeline our engineers use on managed enterprise migrations — the same code, now something you can drive yourself.
03Pricing
Try 10 pages free, then $1 a page
Credit-based, pay-as-you-go. No scoping call, no quote — sample it on your own docs before you spend anything.
04Sources
NotionSlabConfluenceSoonGoogle DocsSoon
Skip to content

AI Agent Deleted Your Data — Who Is Responsible?

If an agent-run migration deletes or corrupts records, the AI vendors' terms leave the clean-up with you. What those terms say, and what an accountable migration looks like.

Roopendra Talekar Roopendra Talekar · · 15 min read
AI Agent Deleted Your Data — Who Is Responsible?
TALK TO AN ENGINEER

Planning a migration?

Get a free 30-min call with our engineers. We'll review your setup and map out a custom migration plan — no obligation.

Schedule a free call
  • 1,500+ migrations completed
  • Zero downtime guaranteed
  • Transparent, fixed pricing
  • Project success responsibility
  • Post-migration support included

AI Agent Deleted Your Data — Who Is Responsible?

If an AI agent deletes or corrupts records during a migration, the customer owns the outcome. The AI vendor does not, and neither does the operator of any MCP server in the path. Anthropic's and OpenAI's business terms say so plainly. Those terms are reasonable for a general-purpose tool. They were not written for a job that moves every record from one production system to another.

For founders and operations leaders weighing a DIY AI-agent migration against a migration partner, this chain of accountability matters. This article covers what the AI vendors' terms say, how migration data gets lost or changed without anyone misbehaving, what a bad load costs once users work in the destination, and what it looks like when a named person owns the result.

What do AI service terms say about data loss?

Anthropic's and OpenAI's business terms put responsibility for outputs on the customer, disclaim warranties, and cap liability at 12 months of fees. These are not hidden clauses. They are the core terms of a general-purpose platform.

Anthropic's Commercial Terms

  • D.3. "It is Customer's responsibility to evaluate whether Outputs are appropriate for Customer's use case". The same section asks the customer to consider human review where appropriate before using or sharing outputs.
  • L.2. The services and outputs are provided "AS IS" and "AS AVAILABLE", without warranty of any kind.
  • L.3. Neither party is liable for consequential or indirect damages, and the excluded losses include data. Each party's liability is capped at the fees the customer paid for the services in the previous 12 months. (anthropic.com)

OpenAI's Services Agreement

  • 4.3. "Customer is solely responsible for all use of the Outputs". The customer is also responsible for evaluating whether outputs are accurate and appropriate for its use case.
  • 12.2. The services are provided "AS IS".
  • 14.2. Each party's total liability is capped at the amount the customer paid OpenAI in the twelve months before the event that gave rise to it. (openai.com)

Both caps have exceptions, such as indemnification obligations. Neither exception turns the vendor into the owner of a migration result.

MCP servers add parties, not accountability

Anthropic's Claude Code documentation says Anthropic does not security-audit or manage third-party MCP servers (code.claude.com). OpenAI's documentation says remote MCP servers have not been verified by OpenAI (developers.openai.com). Both statements are fair. A model vendor cannot vouch for a server that someone else runs. The result is that no single vendor owns the end-to-end behaviour of an agent-run migration.

Why these terms are reasonable

These terms are reasonable for a general-purpose tool. Anthropic and OpenAI sell model access and tooling, not migration services. They cannot see your source schema, your field mapping or your destination's automations. It would make no sense for them to warrant the outcome of a job they did not scope and cannot inspect.

The consequence is concrete. When you use a general-purpose tool to run a data migration and the result is wrong, there is no second party to share the cost. The clean-up is yours.

Warning

General-purpose AI terms are not migration outcome terms. If a migration corrupts production data, the model vendor's business terms do not make that vendor your migration backstop.

How migration data gets lost without anyone misbehaving

Most data loss in a migration does not need a rogue agent. It comes from documented platform behaviour. Vendor MCP servers are designed for conversational work on a few records at a time. Each vendor keeps a separate bulk or import path for large volumes, and the MCP tools do not call it. The generic tools do what their documentation says, and that documentation was written for a different job.

Documented behaviour What it does to a migration
Zendesk's list tickets endpoint: "Archived tickets are not included" (developer.zendesk.com) Archived history stays behind unless the run uses the Incremental Ticket Export
A Zendesk ticket created through the normal create endpoint is stamped with the current time (developer.zendesk.com) Years of history are re-dated to the migration day
Tickets from Create Many Tickets "may be affected by your business rules", which can include emails to end users (developer.zendesk.com) Customers can be emailed about years-old issues
HubSpot's CRM search is capped at 10,000 total results for any query (developers.hubspot.com) An export built on search stops at 10,000 records, with no error
Salesforce hosted MCP: deleting a master record also deletes all its detail records, and "no undelete tool is available through MCP" (developer.salesforce.com) A clean-up delete can remove more than intended, and a restore needs another path
Atlassian Rovo MCP: delete tools are off by default, and deletions are permanent (support.atlassian.com) Turning deletes on for a clean-up makes each one final
Zendesk's Idempotency-Key: "Keys expire after two hours." (developer.zendesk.com) A retry sent two hours or more later creates a duplicate ticket

None of these is a defect in an AI model or an MCP server. Each tool behaves as documented. The gap is that a migration needs the import path, the full export, a restore plan and safe retry rules. A general-purpose agent working from public documentation has to find each of those, on every run. When one is missed, the terms above decide who pays, and it is the customer.

Why is a migration a hard job for a general-purpose agent?

A data migration combines broad read access, mass write access and irreversible business side effects. It pushes against the safety advice the AI vendors themselves publish.

Five ways a migration goes wrong silently

Five failure types matter most here. They differ in how easy they are to detect and to reverse.

Failure type What happens Detectable without deep audit? Reversible after go-live?
Silent truncation Records dropped at API pagination ceilings or context limits No. Record counts look plausible Only before new data mixes in
Timestamp overwrite Created and updated dates stamped with the migration run date No. The fields are populated No. The historical sequence is lost
Referential orphaning Child records (notes, attachments, line items) created without valid parent links Partly. Broken links surface over time Partly. Orphans can be relinked if parents exist
Automation double-fire Standard create endpoints fire workflows, SLA timers and outbound emails on migrated records No. Emails are sent and timers run No. Emails cannot be recalled
Soft-delete omission Archived or deleted records left out of default list responses No. Absent records look as if they never existed Only if the source is still accessible

Any process can produce these without raising an error, whether it is an agent, a script or a person with a CSV file. Each one is invisible to a process that treats "no exception thrown" as "migration successful".

Safety guidance meets migration requirements

OpenAI says to keep tool approvals on and to stop untrusted data from directly driving agent behaviour (developers.openai.com). Anthropic says autonomous agents carry the potential for compounding errors and recommends extensive testing in sandboxed environments (anthropic.com). OWASP recommends human-in-the-loop controls for privileged operations (genai.owasp.org). Microsoft gives the same advice: test in a safe environment, keep a human in the loop for high-stakes tasks, limit permissions and keep detailed logs (learn.microsoft.com).

A migration pushes against each of those. Nobody can approve 50,000 or 500,000 writes one at a time, so approvals drift toward "always allow". Least privilege is awkward because the job needs read access to the whole source and create or update access across the destination. The MCP security best practices warn against exactly that kind of omnibus scope (modelcontextprotocol.io). And the source is full of outsider-written text, such as ticket bodies, email threads, CRM notes and form submissions. The model must read that text as data and not as instructions.

Context windows and search ceilings

An agent works by reading data into a context window and issuing tool calls. A migration involves hundreds of thousands of records. Claude Code, for example, warns when MCP tool output exceeds 10,000 tokens and limits it to 25,000 tokens by default (code.claude.com). That is a sensible default for conversational work.

Vendor MCP servers are built for conversational work, not bulk export. HubSpot's CRM search, which the remote MCP server's search tool relies on, caps any query at 10,000 total results, 200 objects per page and five search requests per second per account (developers.hubspot.com). An export of every contact built on search stops at 10,000 records. No error is raised unless someone checks the total against the source.

That is silent truncation. It is the most dangerous kind, because nothing in the run reports that records are missing.

Prompt injection: the lethal trifecta

Security researcher Simon Willison describes the lethal trifecta for AI agents as access to private data, exposure to untrusted content, and a way to communicate externally. A migration agent has all three by design. (simonwillison.net)

The agent holds read access to everything in the source and write access to the destination. It reads untrusted content constantly: ticket bodies, emails, CRM notes and web form submissions written by outsiders. If someone once submitted a support ticket containing prompt-injection instructions, an agent reading that ticket during a migration could act on them in the destination system.

Researchers have shown the mechanism. Invariant Labs showed in May 2025 that a malicious GitHub issue could lead an agent connected to the GitHub MCP server to copy private repository data into a public pull request (invariantlabs.ai). General Analysis showed in July 2025 that a support ticket containing instructions could lead an AI coding assistant, connected to the Supabase MCP server with a privileged key, to expose a secrets table (generalanalysis.com). These were research demonstrations, not customer breaches. Supabase's advice afterwards was blunt: "Never connect AI agents directly to production data." (supabase.com)

System fields and soft deletes

Default list calls leave records out. Zendesk's list tickets endpoint does not include archived tickets, and Zendesk recommends the Incremental Ticket Export endpoints to get all tickets (developer.zendesk.com). A process that uses the default list call migrates the visible records and reports success. Archived history stays behind in the source, with no warning.

Keeping history needs import-specific paths. Zendesk's Ticket Import API can set created_at, updated_at and solved_at on tickets and created_at on comments, and only admins can use it (developer.zendesk.com). Salesforce lets audit fields such as CreatedDate be set only on insert, and only when the org enables "Set Audit Fields upon Record Creation" and the loading user has that permission (developer.salesforce.com). A run that uses the generic create tools stamps each record with the migration date, and authorship collapses onto the connected user.

Anthropic's engineering team recommends that agents process large data in code outside the model. It notes that models "may be more likely to make mistakes when copying data between tool calls". (anthropic.com) For production moves, a deterministic loader and human sign-off follow the same advice.

What does a bad load cost after go-live?

The price of the AI tooling is not the cost that matters. The real cost arrives after the migration has run, users have started working in the destination, and someone finds the data is wrong.

You cannot simply drop and re-run. Once users work in the destination, new live records mix with migrated ones. Support agents have replied to tickets. Sales reps have logged calls against migrated companies. A bulk delete of migrated records would destroy live work too.

IDs have been issued and linked. The destination assigned new IDs to migrated records. Other records, automations and integrations now reference those IDs, and so do bookmarks, webhook payloads and spreadsheets. Re-importing creates duplicates with new IDs and breaks every link. Fixing it needs a map from old IDs to new ones, built after the fact.

Automations have already fired. If the migration used normal create endpoints instead of import-specific APIs, the destination treated each migrated record as new activity. Triggers fired, SLA timers started, and customers may have been emailed about issues resolved years ago. An email cannot be recalled by deleting the ticket.

Engineering time goes to diagnosis, not product. Engineers read API logs, write one-off scripts to find bad records, and reconcile source and destination by hand. Each hour spent on this is an hour not spent on the product.

Trust in the new system erodes. Users who find wrong data early start to doubt the platform. Side spreadsheets appear, activities go unlogged, and adoption slows.

A second attempt by the same route carries the same risk. Without a frozen mapping, a run ledger and an independent check, a second run cannot show that it is right either. Anthropic's documentation says model results are not fully deterministic, even at temperature 0 (platform.claude.com). A re-run that routes records through a model may not reproduce the first run.

No credible statistic exists for the cost of a migration redo, so none is given here. It depends on data volume, the destination platform and how long users worked before the problem surfaced. Most of the cost is engineering time, user disruption and lost trust, not tooling.

Who is accountable when nobody is accountable?

In a traditional service engagement, accountability is contractual. A vendor scopes the work, agrees a price, delivers a result and owns the rework if the result does not match the scope. With a general-purpose AI tool, that chain does not exist. The model provider sold you model access. The MCP server operator published a connector. Nobody sold you a migration, so nobody owns the outcome of one.

Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls (gartner.com). In a migration, the risk control that matters most is a named owner of the result.

What does an accountable migration look like?

Accountability in a migration is a set of engineering controls and commercial commitments. They exist before the first record moves.

What makes a migration pipeline deterministic

"Deterministic pipeline" is the core technical difference between a managed migration and an agent-run one. It has four properties:

  1. Idempotency. Running the same job twice produces the same destination state. This needs keyed matching: source record IDs are matched to destination records before writing, and a match is updated rather than created again. Any process that calls a create endpoint twice creates two records.

  2. Record-level logging. Every source record is logged with its source ID, its destination ID after the write, the time of the write and the response code. If the pipeline stops, it resumes from the last confirmed write. A chat session has a conversation history, not a per-record ledger.

  3. Deterministic field transformation. The same source value produces the same destination value on every run. That means the transformation is code with a defined input and output, not a model call.

  4. Independent validation. A separate process reads the destination after the migration and compares counts, field-level samples and referential links with the source. Because it is not the process that ran the migration, it does not inherit the same blind spots.

None of these needs the absence of AI. AI is useful for schema discovery, first-draft field mappings, transformation code, tests and spot checks. The dlt team's HubSpot-to-Attio migration used AI to draft the data model, transforms and tests while people reviewed every change and ran a staged rollout (dlthub.com). ClonePartner uses AI the same way, as a tool inside its engineering process. The data movement itself runs in code that has the four properties above.

A dedicated engineer owns the project

Every managed migration has a dedicated migration engineer who owns it from scoping to cutover. The same engineer builds the field mapping, runs the samples and runs the cutover. Only that engineer can view the customer's data.

This matters because migration problems are diagnosis problems. When something is wrong, you need a person who knows the source schema, the destination schema, the mapping decisions and the edge cases. A chat session does not carry the reasoning behind a mapping choice made three days earlier. The engineer who made it does.

One fixed price, starting at $500

ClonePartner charges one fixed price per project, starting at $500. There is no hourly billing, no hidden fees, no change orders and no per-record metering. A free 30-minute scoping call produces the fixed-price proposal.

Because there are no change orders, extra sample runs or messier-than-expected data do not change the price. The cost of getting it right sits with the team doing the work, not with the customer.

Unlimited sample migrations, and re-migration at no additional cost

Before the full migration runs, the customer reviews a sample migration on real data. If anything is wrong, such as a field mapping, a format or missing records, the mapping is adjusted and another sample runs. Unlimited sample migrations are included in the price.

The field mapping is frozen only after the customer signs off the sample. The same mapping runs the full migration and the final delta, so the production run matches the run the customer signed off. If a run has to be redone, re-migration is at no additional cost.

Multi-stage validation

A migration is not done because a chat transcript says so. ClonePartner validates in stages:

  • Record-count parity: the number of records per object in the source matches the destination.
  • Field-level validation: specific values mapped correctly.
  • Referential integrity checks: child records such as notes and attachments link to the right parents, such as companies and tickets.
  • Checksums: a mathematical check that data arrived intact.

The output is a reconciliation report, not a chat message. "No errors" and "correct" are different things. A missing field, a mismatched ID or 500 records skipped at a pagination limit do not throw errors.

A Monday go-live after a 30-minute delta

The bulk historical load runs in the days before go-live while the team keeps working in the old system. Before the working day starts on Monday, the old system is frozen and a final delta migration moves only the records that changed. That takes about 30 minutes. The team logs in to the new system on Monday morning.

Most migrations complete in 1 to 5 business days, depending on data volume, complexity and API rate limits.

Info

Zero downtime depends on the source system's API and change tracking. Legacy systems without either may need workarounds that introduce brief sync delays. For more on how this works in practice, see How We Run Migrations at ClonePartner.

Security and compliance

ClonePartner holds SOC 2 Type II and ISO 27001, independently audited, with GDPR, CCPA and HIPAA safeguards. Data is encrypted with AES-256 at rest and TLS 1.2 or higher in transit. NDAs and DPAs are available on request, and a BAA is signed before any PHI is handled. Staging databases are purged no later than 30 days after project close. Fields are copied from source to destination without a language model reading customer records. See trust.clonepartner.com for details.

How should you evaluate your migration approach?

Factor General-purpose AI agent run Managed migration
Who owns the outcome You, under the AI vendor's terms A dedicated migration engineer owns the project
Field mapping Decided during the session Frozen after sample sign-off, reused for the full run and the delta
Timestamp preservation Depends on whether the run finds each platform's import path Import-specific APIs and permissions, set up per platform
Archived and soft-deleted records Missed by default list calls unless the run asks for them Part of what the migration engine already handles
Validation The run reports on itself Record-count parity, field-level validation, referential integrity checks, checksums
Audit trail Chat transcript Run ledger and reconciliation report
Cost of a bad run Your engineering time and business disruption Re-migration at no additional cost, within one fixed price

If your migration is small and low-risk, and you have the engineering time to validate every record, an AI-assisted approach with strong human review can work. If you are moving a production CRM, help desk or knowledge base, where users will be working in the destination on Monday, the accountability gap is the risk you are pricing.

The real decision is not "AI or no AI"

AI service terms are not the problem. They are honest about what a general-purpose tool can promise. The issue is the gap between what the tool's terms cover and what a migration needs: a correct result, with evidence, owned by someone other than you.

If you are weighing a DIY agent migration against a partner, ask one question before anything touches production: who, by contract and by process, is on the hook when the first run is wrong? If the honest answer is "still us", you are not delegating migration risk. You are only automating the attempt.

ClonePartner has completed 1,500+ successful migrations across CRM, help desk and knowledge base platforms. If you want a fixed-price proposal before anything touches production, start with a free 30-minute scoping call.

Frequently Asked Questions

Who is responsible when an AI agent deletes or corrupts data during a migration?
The customer. Anthropic's Commercial Terms make it the customer's responsibility to evaluate whether outputs suit its use case, provide the services and outputs "as is", exclude lost data from liability and cap liability at the fees paid in the previous 12 months. OpenAI's Services Agreement makes the customer solely responsible for all use of the outputs, provides the services "as is" and caps liability at the amount paid in the prior twelve months.
Can you recover from a bad migration load after go-live?
Not easily. Once users work in the destination, new live records mix with migrated ones, destination IDs have been issued and linked, and automations have fired. You cannot drop and re-run without destroying live work. Recovery means finding the bad records, patching them without touching live data, re-importing and re-linking. Most of that is manual engineering work.
Why are vendor MCP servers not built for bulk data migrations?
Vendor MCP servers are designed for conversational work on a few records at a time. Each vendor keeps a separate bulk or import path for large volumes, such as Zendesk's Ticket Import API, and the MCP tools do not call it. A migration also needs broad write access and reads untrusted text, which makes the approvals and least privilege that AI vendors recommend hard to apply.
What does a failed data migration cost?
No credible statistic exists for the cost of a migration redo, because it depends on data volume, platform and how long users worked before the problem surfaced. Most of the cost is engineering time, user disruption and lost trust in the new system.
How is a managed migration different from a general-purpose AI agent run?
In a managed migration, a dedicated migration engineer owns the project. It runs at one fixed price, starting at $500, with unlimited sample migrations and re-migration at no additional cost. The mapping is frozen after the customer signs off a sample, and the result is checked with record-count parity, field-level validation, referential integrity checks and checksums. Go-live is on a Monday, after a final delta of about 30 minutes. In a general-purpose agent run, the customer is both the operator and the accountable party.

More from our Blog

AI Agent Data Migration via MCP: The 2026 Guide
CRM/Help Desk/Migration Guide

AI Agent Data Migration via MCP: The 2026 Guide

AI agents are useful for drafting migration mappings and code. Vendor MCP servers are built for a few records at a time, so the bulk load needs a migration engine and someone accountable for the result.

Nachi Raman Nachi Raman · · 12 min read