AR7 All articles
Enterprise Technology

Retention Without a Receipt: Auditing What AI Vendors Keep After the Contract Lapses

AR7
Retention Without a Receipt: Auditing What AI Vendors Keep After the Contract Lapses

Photo: enterprise data retention audit dashboard analytics screen, via wallpapertag.com

A renewal negotiation typically settles three things: unit price, seat count, and the service credits that apply when the platform goes dark. What it rarely settles is the less visible question of what the vendor still holds once the contract lapses. For software that functioned mainly as a place to store documents, that omission was tolerable. For the current generation of AI note-takers, coding assistants, and support copilots, it is not. These tools ingest meeting transcripts, repository context, and customer conversations — material frequently more sensitive than anything held in the systems of record they were bought to supplement.

The problem is not that vendors are unusually careless. It is that deletion in AI products behaves like a workflow rather than a switch. A transcript may be condensed into embeddings, replicated into a logging bucket, retained in an evaluation corpus, and cached at an inference provider, each with its own retention clock and its own support path. When a customer requests removal, the answer often arrives as a ticket number instead of a certificate.

What follows is a repeatable audit, executed before renewal rather than after an incident, that treats retention risk as a line item with a cost, a test, and an accountable owner.

Why Retention Rarely Appears on the Renewal Worksheet

Procurement templates are built around obligations that can be measured at signature: encryption standards, access controls, uptime commitments, insurance limits. Deletion usually occupies a single paragraph in the data processing agreement promising removal within a reasonable period or at the customer's instruction. Neither phrase specifies a mechanism, a deadline, or an evidentiary standard.

Meanwhile the footprint expands without announcement. A note-taker attending two hundred weekly meetings accumulates a searchable record of internal deliberation. A coding assistant indexes private repositories. A support copilot reads every customer conversation, including the ones customers believed were confidential. None of this appears on an invoice, which is precisely why it escapes review.

Step One: Map What the Tool Actually Touches

Inventory by data class, not by application. Build a four-column table: data class, ingress point, persistence mode, maximum retention. Data classes worth naming explicitly include meeting audio and transcripts, source code and commit context, customer support conversations, internal documents used as retrieval context, prompts and completions, and derived artifacts such as embeddings, summaries, and evaluation sets.

Two columns matter more than they appear. Persistence mode distinguishes what is stored from what is merely processed in memory, and vendors frequently conflate the two. Maximum retention is the only column that can be compared across vendors, and the only one that produces a defensible number for the renewal file.

Step Two: Classify the Residue

Sort findings into three buckets. The primary store is the material you can see and delete through the interface. Operational residue covers logs, backups, caches, and support archives. Derivative material includes embeddings, fine-tuned weights, and evaluation snapshots. The third bucket is where most audits stall, because vendors describe it in aggregate and customers lack a vocabulary to press further.

Request a written statement identifying which classes feed model training or evaluation, whether that use is opt-in or default, and what the deletion path is for each class. A vendor unable to distinguish derivative material from primary storage has supplied a finding, not a wording inconvenience.

Step Three: Test the Export Before You Need It

Export is the cheaper half of the exit question and the half you can test this week. Run a full export from a sandbox workspace, time it, and verify fidelity. Many AI tools export the transcript but not the summary, or the summary but not the linked action items, or the action items without the source quotations that justified them. If the export format cannot be ingested by your records system, the exit is theoretical regardless of what the contract says.

Record the elapsed time and the format in the audit file. Both become negotiating material.

Step Four: Exercise Deletion as a Workflow

Select one workspace, one user, and one dataset. Submit the deletion request through the ordinary support channel and log three timestamps: acknowledgment, completion, and confirmation. Ask three precise questions in the same ticket. Does the deletion propagate to backups, and over what interval? Does it reach subprocessors and inference partners? Does any copy persist in an evaluation or training corpus?

The useful metric is elapsed days from request to written confirmation, benchmarked against the contractual window. Where the two diverge, you have found the gap between the agreement and the operating procedure.

Step Five: Demand Evidence, Not Assurance

Evidence consists of artifacts. Assurance consists of adjectives. An acceptable package includes a dated deletion confirmation keyed to a specific dataset identifier, a subprocessor register listing data categories and processing regions, an audit report whose period end is recent and whose exceptions are disclosed rather than summarized, and a retention schedule naming each data class with its maximum life.

If those four items cannot be produced within a stated number of business days, that delay is itself a data point for the renewal conversation.

A Short Evidence Request

Five items, one page, sent to every AI vendor in the portfolio:

Turning Findings Into Leverage

Score each vendor on three axes: clarity of the retention schedule, testability of the deletion path, and quality of the evidence supplied. A vendor that scores well on price and poorly on evidence is not inexpensive; it is deferring a cost to whichever team handles the incident or the litigation hold.

Three actions follow naturally. Attach a deletion exhibit to the renewal rather than relying on the standard agreement. Tie a portion of fees to documented completion of a test deletion. Require annual re-attestation of the retention schedule, since products change faster than contracts.

The Clause Worth Adding

Language along these lines is neither novel nor aggressive: the vendor shall delete all customer data, including derivative artifacts, within a stated number of days of termination; shall propagate deletion to subprocessors within a stated interval; and shall furnish written confirmation identifying the datasets removed. Substitution of one clause for three pages of assurance is the entire point.

The audit is not designed to catch anyone. It is designed to make an invisible obligation visible while negotiating room still exists.

All Articles

Related Articles

The Incidents That Never Paged: Treating Near Misses as Your Most Underused Reliability Data

The Incidents That Never Paged: Treating Near Misses as Your Most Underused Reliability Data

Procurement Reads Your Pricing Page Before Your Sales Team Calls: A Field Audit of B2B SaaS Packaging

Procurement Reads Your Pricing Page Before Your Sales Team Calls: A Field Audit of B2B SaaS Packaging

Velocity for the Many, Bottlenecks for the Few: Rethinking Who Actually Benefits From AI Pair Programming

Velocity for the Many, Bottlenecks for the Few: Rethinking Who Actually Benefits From AI Pair Programming