What evidence do compliance officers need for AI usage?

A dated chain: what tools were observed, what vendor terms said when checked, what was sanctioned at the time, and how findings were handled. A log-based audit supplies the first three links per period.

Can compliance run AI audits without technical staff?

Yes, apart from requesting one quarterly log export from IT. Uploading and reading the report require no console access or technical background.

How do AI audit findings feed DPIAs?

By prioritization: tools with heavy usage, sensitive department attribution and adverse or silent training verdicts rise to the top of the assessment queue.

What happens when a vendor's training verdict changes?

The change appears as a delta between audit runs, triggering a vendor review and demonstrating that the organization's process detects drift.

How should audit reports be filed for maximum evidence value?

Unedited PDFs with constant quarterly windows, one page of decision minutes per run, retained CSVs for resampling, and an explicit scope note in every filing.

evidence chain

found
terms checked, dated
sanctioned at the time
filed

For compliance officers

AI usage needs an evidence chain, not a narrative

Q: Does an AI usage audit create new privacy risk?

It is minimized by design: hostnames, identities and timestamps only, no prompt content, uploads discarded after each run, and reports deletable within their 90-day window.

Registers, DPIAs, audit responses and insurer forms all ask versions of the same three questions: what was in use, what did the vendor's terms say at the time, and was it approved then. A log-based audit supplies all three links, dated, from the logs your network already keeps.

Build the first link free See the evidence pack

4 linksfound, terms, sanctioned, handled

13,000+vendor training verdicts, dated

$0free preview, no card

90 daysreport retained, deletable earlier

shadow-ai-audit.pdf

Cover of the sample Shadow AI audit report prepared for an example company, with its headline findings.

Sample report View all 11 pages Opens in your browser · PDF

The chain

Four links every AI question resolves to

Whatever the framework, the questions decompose the same way. Each link is a report field, not a reconstruction after the fact.

Link 1

What was found

The tool table: every AI hostname observed in the window, with category, per-user attribution and usage volume. Observed from logs, not remembered from interviews.

Link 2

When terms were checked

Each tool's training verdict carries the date the vendor's terms were reviewed. "As of" statements replace impressions, and changed verdicts between runs are findings in themselves.

Link 3

What was sanctioned

The sanctioned split records your approved list against reality, per run. A year later you can still show what the policy state was in any given quarter, without reconstruction.

Link 4

How it was handled

Your triage minutes plus the next report's deltas close the loop: found, decided, verified. That is the sentence auditors and regulators want to read in your files.

The chain's strength is temporal: every link is dated, so the file answers "what did you know and when" without reconstruction. Reconstruction is where compliance stories fall apart under scrutiny.

The evidence

What the report actually shows

A mock panel from a compliance-relevant audit run. Every bar, every verdict, every date is a field that maps directly to a register or DPIA row.

shadow audit · compliance evidence run

Tool	Category	Trains	Verdict date	Users
chatgpt.com	General chat	Yes, default	2026-09-14	38
deepseek.com	General chat	Yes, default	2026-09-14	12
notion.so	Productivity	No (enterprise)	2026-09-12	24
grammarly.com	Writing assist	Opt-out required	2026-09-10	31
crushon.ai	Companion	Yes, NSFW	2026-09-14	1

Sample data. Dated verdicts from ongoing review of 13,000+ vendor training terms.

Dated verdicts

Every training-on-data verdict carries the date the vendor's terms were last reviewed. Changed verdicts between runs surface as delta findings for your DPIA queue.

Per-user attribution

When the log includes an identity column, the report shows which users or devices reached each tool. This maps directly to the DPIA impact scope and incident triage.

The sanctioned split

Your approved-tool list is reconciled against observed reality. The gap between them is the compliance exposure, quantified per run and comparable across quarters.

Feeding the documents

Where each report field lands in your files

Six standard compliance documents and the report field that supplies each one. No manual extraction, no retyping, and no judgement calls about what to include.

Your document	What it needs	Report field that supplies it
Processing register	Third parties actually touching data	Tool table reconciled against recorded processors
DPIA queue	Which assessments to run first	Training verdicts x user counts x department attribution
Vendor register	Suppliers in use without agreements	Unsanctioned tools with real usage, sorted by volume
Risk register	Quantified, dated exposure entries	High-risk and abusive counts per period, CSV import
Audit responses	"How do you know?" answers	The PDF filed unedited, method statement included
Insurer questionnaires	AI usage controls, evidenced	Cadence plus the last two reports as attachments

The framework-specific detail lives on sibling pages: GDPR, EU AI Act, ISO 27001 and HIPAA, each mapped to the same chain.

The officer's leverage

What changes when the evidence exists

Three working situations, before and after the chain. Evidence does not replace judgement, but it replaces the scramble that usually precedes it.

The vendor question

Before: "we believe staff mostly use approved tools." After: "31 tools observed, 6 sanctioned, here is the migration queue with dates." The second version wins budget and respect from the board.

The incident afternoon

Before: interviews, hope, and calendar holds. After: a targeted window export answers which tools and identities in hours, and the standing baseline shows whether the behavior was new or ongoing.

The regulator letter

Before: a scramble to reconstruct. After: quarterly PDFs on file, each with scope lines and dated verdicts, plus minutes showing decisions. The response drafts itself from the shelf.

The evidence process itself stays defensible: hostnames, identities and timestamps only, no prompt content, uploads discarded after each run, reports deletable within their 90-day window, operated by an EU-based data processor.

Quarterly cadence

The four-quarter evidence rhythm

Each quarter builds on the last. By Q4, your AI file answers auditors from the shelf without any reconstruction.

Q1: Baseline

First run establishes the inventory. Totals, categories, training verdicts, sanctioned share. Everything that follows is measured against this snapshot. File the PDF and the triage minutes.

Q2: Delta

Second run reveals the drift: new tools adopted, verdict changes, sanctioned-share movement. The delta narrative is your first trend statement and the strongest evidence of a working process.

Q3: Remediation trail

Third run proves actions landed. If Q2 flagged unsanctioned tools, Q3 shows whether they were blocked, sanctioned or are still active. Each outcome is a documented decision.

Q4: Annual summary

Four reports, four sets of minutes, a year of dated decisions. The annual review, the insurer renewal and the certification body all get the same file, assembled, not reconstructed.

Same 30-day export window every quarter. Method consistency is half of what makes trend claims defensible under examination.

Working with IT

The two-role loop that keeps evidence flowing

Compliance owns the chain; IT owns the mechanics. The interface is deliberately thin, so neither role blocks the other.

What you ask IT for

One log export per quarter, same window, identity column included when available. Twenty minutes of their time per period. The IT manager page is written to be forwarded as the request and explains the export steps for every supported platform.

What you own after that

The upload or the reading of the report, the triage meeting, the minutes, the filing, and the delta narrative quarter over quarter. No console access needed, no dependency on security headcount, no queue behind other projects.

The one meeting that matters

A 45-minute quarterly read with IT and the affected department heads: verdict changes, new tools, sanctioned-share delta. Minutes from that meeting are link four of the chain and the strongest evidence that governance is active.

Escalation without drama

Abusive-purpose flags and high-sovereignty findings route to legal and HR through the same minutes. The report's risk flags mean escalations arrive pre-triaged and quantified, not as rumors or surprises that trigger over-reactions.

Filing discipline

Six habits that keep the chain admissible

Evidence value is mostly maintained, not created. These are the habits, in the order they most often get skipped under time pressure.

File unedited PDFs

Excerpting reintroduces the assertion problem the report was built to solve. The scope line and evidence statement are written to stand alone in a filed document.

Constant windows

Same 30 days every quarter. Method consistency is half of what makes trend claims defensible when an examiner asks how you know something changed.

Minutes per run

One page: findings acknowledged, decisions made, owners named, deadlines set. The report shows facts; the minutes show governance. Both are needed.

Keep the CSVs

Machine-readable copies let internal audit resample findings without raising new requests, which is the cheapest credibility you will ever buy from the audit function.

Note the scope

"Logged network only; off-network and personal-device use governed by acceptable-use policy" in every filing. Stated boundaries survive scrutiny; silent ones invite questions.

The payoff

Four quarters of this and your AI file answers auditors, insurers and regulators from the shelf. That is the whole point of a chain: once built, it defends itself.

Worked scenario

When an auditor asks about AI controls

A compliance officer at a 400-person company receives a letter from a certification body asking about AI usage monitoring. Here is what the chain produces.

The question received

"Please describe how the organization identifies and monitors AI tools used by employees, including evidence of periodic review and actions taken on findings." A standard ISO 27001 or SOC 2 question.

The answer drafted

"We run a quarterly log-based AI inventory covering all logged network traffic. Each run produces a dated PDF with tool-by-tool training verdicts, a sanctioned/unsanctioned split, and per-user attribution. The attached four reports cover the trailing twelve months."

The attachments

Four unedited quarterly PDFs, each with scope line, tool table, verdict dates and evidence statement. Four sets of one-page minutes documenting the triage decisions. One CSV of the latest run for resampling.

Time to assemble

Under thirty minutes, because the chain was already built and filed. The response letter, the attachments and the evidence trail were already on the shelf waiting for exactly this question.

Accepted log sources

The export you already have is enough

The audit reads the same DNS, proxy and firewall logs your network already produces. No new collection, no agent, no SSL inspection.

DNS filters

Cisco Umbrella, Zscaler, Cloudflare Gateway, DNSFilter, NextDNS, Pi-hole. Each guide walks the export steps from the console you already use.

Firewalls and proxies

Palo Alto, Fortinet, SonicWall, Squid. Syslog key=value lines parse directly, and CSV exports with a header row are accepted from any source.

Any hostname list

A plain text file with one hostname per line is the minimum viable audit. Even a grep of a SIEM export or a spreadsheet column will parse, producing the same evidence chain.

What it costs

Pricing for a standing compliance program

One-time packs for ad hoc needs, monthly plans for quarterly cadence.

Quarterly from packs

Four reports a year from the 5-pack costs $299 total, $60 per run, leaving one spare credit for incident work. Credits never expire.

Monthly plan

The Essentials plan at $99/mo includes one full audit per month, enough for monthly cadence or quarterly runs with incident capacity between cycles.

Multi-site

The MSP plan with client reports ($398/mo) includes 25 full audits per month, suited for organizations with multiple business units, each audited independently under one license.

Full detail on the pricing page.

FAQ

Compliance officer questions

The questions compliance officers ask before committing to the evidence chain.

Can I run this without IT doing it for me?

You need one log export from IT per period; the upload and reading need no technical background. Many officers run the whole loop after the first walkthrough, keeping IT involvement to the export step only.

Are the training verdicts authoritative?

They are dated observations from an ongoing review of 13,000+ vendor terms, built for triage and citation. Final legal weight is always your counsel's call, but the dated "as of" format is designed for filings.

How does this feed DPIAs specifically?

It prioritizes them: tools with heavy use, sensitive departments and adverse training verdicts float to the top of the assessment queue. The GDPR page details the mapping step by step.

What if a verdict changes between quarters?

The delta is itself a finding: it triggers a vendor review and shows your process catches drift, which is exactly what examiners look for when they assess monitoring controls.

Does the audit create new personal-data risk?

Minimal by design: hostnames, identities and timestamps only, no prompt content, uploads discarded after each run, reports deletable within their 90-day window. The description fits one register row.

What does the standing program cost?

Quarterly reports from $60 to $99 each depending on packs, or the $99/mo Essentials plan for monthly cadence. Details on pricing.

Can I share audit reports with external auditors?

Yes. Every report produces a PDF designed to be filed and shared unedited. The scope line, method statement and evidence paragraph are written for an external reader, not just internal use.

Which frameworks does this cover?

The evidence chain maps to GDPR Article 30 and 35, the EU AI Act inventory obligation, ISO 27001 Annex A, HIPAA risk analysis, SOC 2 CC6/CC7, and cyber insurance AI questionnaires. Framework-specific pages detail each mapping.