Records Retention in the AI Era
Your retention schedule lists board minutes, 990s, personnel files, and grant agreements. It almost certainly says nothing about the chat thread where a program director worked through a client situation, the transcript of a board meeting captured by a notetaker nobody invited, or the eighteen months of prompts sitting in a staff member's personal account. Those are records your organization created, and right now most nonprofits could not say where they are, who can read them, or when they disappear.
Most nonprofit retention schedules were written between 2004 and 2010, in the wave of governance tidying that followed Sarbanes-Oxley and the redesigned Form 990. They are competent documents about paper and email. They assign a number of years to each category of file, they name a custodian, and they get reviewed roughly never. The categories they list are the categories that existed when they were drafted.
In the past three years your organization started generating entirely new classes of material. Conversations with assistants that contain program details, donor information, and draft decisions. Meeting transcripts and automated summaries. Prompt histories that show what staff asked and what the tool told them. Generated drafts that became final documents. Logs held by vendors rather than by you. None of it appears on the schedule, which means none of it has a defined lifespan, an owner, or a rule for what happens when a lawyer sends a preservation letter.
This matters in two directions at once, and the directions pull against each other. Keeping too little is a risk when a record turns out to be needed for an investigation, an audit, a grievance, or a funder inquiry. Keeping too much is a risk because everything retained is discoverable, breachable, and subject to whatever promises you made to the people described in it. A retention schedule is the document where an organization resolves that tension deliberately instead of letting a vendor's default setting resolve it by accident.
What follows covers what your existing schedule is actually for, the new categories AI creates, a practical test for deciding whether something is a record, how to write the new rows, why legal holds are the part most organizations get wrong, and the vendor settings that quietly determine whether your policy is real or aspirational.
What a Retention Schedule Is Actually For
It is worth restating the purpose, because organizations that treat retention as a filing exercise write schedules that do not survive contact with a real problem. A retention schedule does four jobs.
It satisfies specific legal obligations to keep certain things. Some are permanent, including articles of incorporation, IRS determination letters, board minutes, deeds, and filed 990s. Some run for defined periods tied to employment law, tax law, contract terms, and funder requirements. Organizations with federal awards generally face a records retention requirement running several years from the submission of the final expenditure report, and grant agreements frequently add their own terms.
It gives you permission to destroy the rest. This is the underrated half. Routine, consistent, documented destruction under a written policy is a defensible practice. Ad hoc deletion by whoever needs disk space is not, and the difference becomes visible at exactly the wrong moment. Sections 802 and 1102 of the Sarbanes-Oxley Act apply to nonprofits and make it a crime to destroy or alter records with intent to obstruct a federal investigation or proceeding, and the practical protection against an accusation like that is being able to show that a document went away because the schedule said so, on the date the schedule said, before anyone knew there was a dispute.
It answers the Form 990 question honestly. The 990 asks whether the organization has a written document retention and destruction policy. Federal law does not require one, but the question is there because the IRS regards it as good governance, and answering yes while operating without a functioning schedule is the kind of gap that gets noticed during an examination or by a funder reading your return.
And it supports the board's duty of care. Directors are responsible for the organization's information as much as its money. A schedule that has not been reviewed since before generative tools existed is evidence that this particular oversight duty has been dormant, which is a reasonable thing for a board to want to fix in a single meeting. Our guide to board-level AI risk registers covers where this sits among the other governance items competing for that agenda slot.
The Categories Your Schedule Is Missing
Before you can assign retention periods you have to know what exists. Most organizations are surprised by the length of this list, and by how much of it lives outside systems anyone is administering.
Assistant conversation histories. The threads staff have with general-purpose tools. These frequently contain the substance of real work: draft communications, analysis of program data, thinking about a personnel matter, or details of a client situation pasted in for help writing a case note. Where staff use personal accounts, this material is not in your control at all, which is the shadow AI problem wearing a records management hat.
Meeting transcripts and automated summaries. Notetakers produce a verbatim record of conversations that previously left only a set of minutes. That is a substantial change to what exists about your board and staff meetings, and it is often created by a tool one participant enabled without a decision by anyone. We have written separately about AI notetakers in board meetings and why the transcript question deserves an explicit answer.
Prompts and system instructions. The instructions embedded in a workflow, the templates staff reuse, and the configuration of any custom assistant. These are closer to procedure documentation than to correspondence, and they matter when you need to explain later how a particular output was produced.
Outputs that entered the record. A generated grant narrative that was edited and submitted, an AI-drafted policy the board adopted, a summary that became the basis for a decision. Once an output is used, it is part of the trail behind an organizational act, and the interesting question is whether you retained any evidence of what was generated versus what a human changed.
Retrieval indexes and embeddings. If you have built anything that searches your own documents, there is now a derived copy of that content in a vector store. Deleting the source document does not delete the derived copy unless something explicitly does that, and organizations regularly discover that content they believed was purged is still answerable through the search layer.
Vendor-side logs. Usage logs, audit logs, and API request records held by the providers you use. You may have contractual rights to them, obligations about them, or no idea they exist. They are still records concerning your organization, and they will be sought if there is ever a dispute about what a tool did.
Governance artifacts about AI itself. Tool approvals, vendor assessments, staff acknowledgments of your acceptable use policy, incident reports, and any bias or accuracy testing you performed. These belong in the permanent or long-retention part of the schedule, because their whole purpose is to demonstrate later that a decision was made carefully. The material described in documenting AI workflows falls squarely in this group.
An inventory prompt for your next finance and operations meeting
Answer these before writing a single retention period
- Which AI tools are in use, including ones nobody formally approved
- Which are on organizational accounts and which on personal ones
- What each tool retains by default, and for how long
- Whether an administrator can search, export, or delete that content
- Where transcripts and summaries are stored after a meeting ends
- Whether any client, patient, or student data has been entered anywhere
- What happens to all of it when a staff member leaves
Is a Chat Thread a Record?
This is where organizations stall, because the honest answer is that it depends, and people want a rule. A workable rule exists, and it is the same one that has always applied to email. Not every email is a record. An email that documents a decision, a transaction, an obligation, or an interaction with a client is. The medium never determined the answer. Content and function did.
Apply that to conversations with an assistant and most threads turn out to be working material, equivalent to a scratch pad or a conversation at a desk. Nobody suggests retaining every idea a staff member had while drafting something. But a meaningful minority of threads are records by function. A conversation where a manager worked through a disciplinary matter documents an interaction with an employee. A thread analyzing program outcomes that informed a report to a funder is part of the support for that report. A conversation containing a client's information is a client record living in the wrong system entirely.
The practical consequence is that you cannot resolve this purely with retention periods, because it depends on what staff put into the tools. That makes the acceptable use policy and the retention schedule two halves of one control. If your policy says client identifiers never go into a general-purpose assistant, the retention question for those threads becomes far simpler. If your policy is silent, you have client records scattered through conversation histories with no schedule, no access control, and no deletion path. This is one of several reasons an acceptable use policy pays for itself well beyond the risks it was written for.
A second consequence is that anything used to support an organizational assertion needs to be traceable. If a generated analysis informed a decision the board made, or a drafted narrative went into a grant report, the organization should be able to describe how that output was produced and what a person verified. That does not mean retaining every intermediate step. It means the final artifact carries enough provenance to answer the question, an idea we developed in our discussion of audit trails for AI-assisted work.
Treat as a record
Function, not format, decides
- Anything documenting a decision the organization acted on
- Content describing an identifiable client, donor, or employee
- Analysis that supports a figure or claim reported externally
- Meeting transcripts where the meeting itself is minuted
- Tool approvals, assessments, and incident reports
Treat as working material
Short retention, routine disposal
- Drafting help on documents that exist in final form elsewhere
- General questions with no organizational information in them
- Formatting, summarizing, and rewriting of already-retained content
- Experiments and tool evaluation that produced no decision
- Routine scheduling and administrative back-and-forth
Writing the New Rows
A retention schedule row needs four things: what the category covers, how long it is kept, who owns it, and where it lives. The fourth is the one that trips up AI-era categories, because the honest answer is often a vendor's servers under settings nobody has looked at.
For general assistant conversations, a short default with an exception path works well. Something in the range of ninety days to a year, automatically purged, with a documented process for moving anything that qualifies as a record into the system where that kind of record belongs. The default should be short precisely because you do not want a growing archive of unclassified material containing who-knows-what. The exception path matters because without it, staff will keep threads indefinitely to avoid losing something.
For meeting transcripts, decide by meeting type rather than setting one global rule. Board and committee meetings have minutes as the official record, and a verbatim transcript alongside approved minutes creates a second, longer, unapproved account of the same meeting. Many boards conclude that transcripts should be deleted once minutes are approved. Staff meetings, supervision sessions, and client-facing conversations each deserve their own answer, and the answers will differ.
For retrieval indexes, the rule should be that deletion of a source document propagates to any derived index within a defined period, and someone should have actually tested that this happens. This is where policy and reality most commonly diverge, and it is a straightforward thing to verify: delete a test document, then ask the system about its contents.
For AI governance artifacts, retain them at least as long as the systems they describe, plus a margin. If you evaluated a vendor in 2024 and used it until 2028, the evaluation is the evidence of your diligence and should outlive the tool. Permanent retention for policy versions and approvals is a defensible default, since these are small documents and their entire value is historical.
For anything containing client or program participant information, the answer is that it inherits the retention rule of the underlying client record, whatever that is under your funding agreements, professional obligations, and applicable state law. This is the strongest argument for keeping that content out of general-purpose tools in the first place: once it is there, the tool has inherited a retention obligation it was never configured for.
Starting points for the new rows
Adjust to your obligations, then write them down
- General assistant conversations: short default, auto-purge, documented exception path
- Board meeting transcripts: delete once minutes are approved
- Staff meeting summaries: retain with the project or personnel file they concern
- System prompts and configurations: retain as procedure documentation
- Retrieval indexes: deletion of source propagates within a stated period
- Vendor assessments and approvals: retain beyond the life of the tool
- Anything with client information: inherits the client record retention rule
The Legal Hold Problem Nobody Has Solved Yet
Here is the part that should concern your board most, because it is the point where an automated retention setting becomes a liability rather than a control.
When litigation is reasonably anticipated, an organization has a duty to preserve relevant information, and routine destruction has to stop for anything within scope. Every organization with a retention policy is supposed to have a hold process that suspends normal deletion. Those processes were built around email systems and file servers, where an administrator can flip a switch that stops deletion for named custodians.
Now consider a wrongful termination claim where the relevant conversation happened in a manager's assistant thread, on a personal account, with a thirty-day auto-delete setting. There is no administrative console. There is no custodian list that includes that account. The deletion is automatic, it continues while everyone is discussing the claim, and by the time anyone thinks about it the material is gone. Whether that ends up characterized as routine operation of a policy or as a failure to preserve depends on facts and on how well you can document what you did and when.
The mitigations are unglamorous and effective. First, know which tools staff use, which is only possible if you have a real inventory rather than an assumption. Second, prefer organizational accounts with administrative visibility over personal accounts for anything work-related, which is a licensing cost worth paying for exactly this reason. Third, write the hold procedure to name AI tools explicitly, listing for each one who can suspend deletion and how. Fourth, when a hold is issued, send the notice to custodians with specific instructions about their assistant histories and transcripts, because a generic instruction to preserve documents will not prompt anyone to think about a chat thread.
Fifth, and most easily overlooked, check whether your vendors can honor a hold at all. Some enterprise offerings support retention configuration and legal hold. Consumer tiers frequently do not. If a tool cannot suspend deletion on request, that is a real limitation to record in your assessment of it, and it may be a reason to prefer a different tool for work touching sensitive matters.
Update your hold procedure to answer these
Before you need it, not after
- Which AI tools hold organizational content, named individually
- Who has the administrative rights to suspend deletion in each
- What the hold notice tells custodians about chats and transcripts
- How personal-account use is handled when it holds relevant material
- Which tools cannot support a hold, and what you do about that
- How you document the date the hold was issued and to whom
Your Policy Is Whatever the Vendor Settings Say
A retention schedule that your systems do not implement is a description of intentions. For traditional records this gap was manageable because the systems were yours. For AI tools the retention behavior is determined by a vendor's configuration, a plan tier, and defaults that change without much announcement.
Four settings determine most of your reality. How long conversation content is retained, and whether that is adjustable. Whether your content is used to improve the provider's models, which is a distinct question from retention and matters independently. Whether an administrator can search and export across the organization rather than only within a single user's account. And whether deletion is genuine removal or a hidden flag, including from backups and from any derived index.
Ask these questions during procurement rather than after adoption, and record the answers in writing with a date, because they change. A tool assessed in early 2025 may have entirely different defaults today, and the assessment on file is what demonstrates you looked. The evaluation approach in our guide to updating your data governance policy for the AI era covers how to slot these questions into a process you can actually sustain.
Then verify rather than assume. Set a retention period, wait, and confirm the content actually disappeared. Delete a document from a retrieval index and confirm the system can no longer answer questions from it. Ask an administrator to export a colleague's history and see whether that is possible. These tests take an afternoon and they are the difference between a policy and a hope.
Finally, assign the review. A retention schedule with no owner and no cadence returns to the state yours is in now, which is out of date in ways nobody has noticed. An annual review, timed alongside the 990 preparation when governance policies are already under discussion, is the lowest-friction way to keep it current. Add a trigger so that adopting any new tool that stores content requires a schedule row before it goes live, which stops the backlog from rebuilding.
Questions to put to every vendor in writing
Record the answers with a date, and re-ask annually
- What is retained by default, for how long, and is it configurable
- Is our content used for model training, and can we opt out
- Can an administrator search, export, and delete across all users
- Does deletion remove content from backups and derived indexes
- Can routine deletion be suspended for a legal hold
- What happens to our content if we cancel the account
Conclusion
Records retention is a policy area where nonprofits have been quietly running on documents written for a different technological era. The schedule is not wrong about board minutes and 990s. It is simply silent about the largest new category of organizational content in a decade, and silence in a retention schedule means the vendor's default wins.
The work required is smaller than it looks. Inventory what exists, apply the same content-and-function test you already use for email, write rows for the new categories with short defaults and clear exception paths, name AI tools explicitly in your legal hold procedure, and verify that your vendors behave the way your policy assumes. That is a project measured in weeks, not quarters, and most of it is decision-making rather than implementation.
Do it before you need it. The value of a retention schedule is entirely in having had it in place beforehand, because a policy adopted after a preservation letter arrives explains nothing about the material that went away before it did. This is one of the few governance tasks where a single afternoon of decisions now genuinely changes the position you will be in later.
Bring Your Retention Schedule Up to Date
We help nonprofits inventory what their AI tools are storing, write the policy rows that cover it, and verify the settings actually match.