Blog Posts
Most enterprises have deployed AI but skipped the security controls. Polygraf's 25-point AI security checklist maps the controls every enterprise needs to know.
From 2022 to 2025 the SEC fined financial firms more than a billion dollars for not what they said, but for not keeping the record of having said it. That enforcement logic is coming for AI. The EU AI Act makes automatic logging a legal requirement for high risk systems, and a chat transcript is not an audit trail. Here is what a defensible one actually looks like.
There is a pattern in financial regulation that most AI teams have not learned yet. In the last few years, the SEC's biggest and most repetitive enforcement campaign was not about market manipulation or fraud. It was about record-keeping – firms whose employees did business on personal texts and messaging apps and who could not produce the records the law required them to keep. The firms admitted the failures. They paid, in aggregate, well over a billion dollars. The regulator's position was simple and unsentimental: if you cannot produce the record, you have violated the rule, no matter whether the underlying conduct was innocent.
That's what is being done to AI now. The EU AI Act makes automatic event logging a design requirement for high-risk AI systems. The 2026 oversight report of FINRA had a new section on generative-AI and told firms to have governance and to track what autonomous agents do. Sectoral rules that are not even AI-related – HIPAA's audit controls, PCI DSS Requirement 10, SEC and FINRA books-and-records rules – apply to AI systems the instant they touch regulated data, because none of them were written to exclude a technology that did not exist yet. In this guide we cover what regulators actually ask for, what a defensible AI audit record looks like field by field, what the real retention floors are, and where most logging implementations quietly fail.
Regulatory exposure is not a theoretical concept. Polygraf's AI Risk Calculator maps your company's exposure to breach, regulatory, litigation and reputational risk and shows you which record-keeping obligations apply to you, from the EU AI Act to HIPAA, PCI DSS and financial books-and-records rules.
It is worth getting to the details because they set the enforcement stance AI will be left with.
The SEC has charged five broker-dealers, seven dually registered broker-dealers and investment advisers and four affiliated advisers with systematic and long-standing failure to maintain and preserve electronic communications. The firms admitted the facts, admitted that they violated federal securities recordkeeping laws and agreed to a combined civil penalty of more than $81 million. One firm that self-disclosed paid $1.25 million; those that did not paid between $8 million and $16.5 million. This was one wave of a multi-year campaign.
Notice what these firms were not charged with. Not fraud. Not misleading clients. They were charged because business occurred on a channel the firm did not capture and the record was not available when the regulator asked for it. Now replace employee texting on a personal phone with employee prompting an AI tool that logs nothing your compliance team can get to. The structural break is the same and so is the risk.
This is the practical test. When the regulator, auditor or plaintiff's counsel comes, they don't want to see your logging architecture, they want to know about a particular decision on a particular date. Click on each one to see why a transcript fails it and what a real audit record needs.
model_version,
system_prompt_hash, config_hash. An auditor can match the decision to the exact
system state that produced it.retrieved_docs[] with IDs and access basis, plus every
tool_call and its parameters — including the ones that errored and were retried. (For
biometric identification systems specifically, the EU AI Act goes further and mandates recording the exact
reference database checked and the input data that produced a match.)actor, role, and — for consequential actions —
human_reviewer, review_action, and review_ts. The AI Act's
biometric logging provisions explicitly call for identifying the natural persons who verified results.
policy_applied, decision (allow
/ redact / block), matched_rules[], redactions[]. "We have a policy against X"
is only defensible when you can show a history of that policy being evaluated and applied.prev_hash,
record_hash) or WORM-equivalent storage with access controls, written by the security team's
system rather than the application owner's.subject_id and correlation_id across systems,
retained past the applicable floor, and exportable on demand. This is the difference between an afternoon
and a crisis.Here's the concrete schema. Each field group exists because a specific regulator asks for it. Treat this as a floor, not a ceiling — and note that the record separates identity, system state, content, governance, and integrity, because auditors examine those independently.
Note that the record above stores input_ref and output_ref — pointers — rather than
raw prompt and response text. This is on purpose. An audit log that inlines full prompts becomes a new long-retention copy of every piece of sensitive data your AI ever saw and GDPR's storage-limitation principle is a direct push against storing personal data longer than needed. Log the metadata and the governance decision immutably, store the content separately under its own access controls and deletion lifecycle. Otherwise your compliance artifact is a breach in the making.
If your AI governance runs through a third-party cloud gateway, your compliance evidence lives in someone else's infrastructure. The EU AI Act imposes a retention obligation on you as provider or deployer – and "our vendor has it" is a brittle response if that vendor has an outage, changes its retention policy, is acquired or is itself subpoenaed. You can't generate a six-month continuous log record retroactively. Evidence you are legally obliged to keep must be generated and stored where you control the chain end to end.
Retention is where the misinformation is thickest. A lot of published guidance mixes up different obligations – most commonly the EU AI Act's ten-year technical documentation obligation as if it was for operational logs. It is not. This is the real picture.
| Regime | What it requires | Floor |
|---|---|---|
| EU AI Act Art. 12, 19, 26(6) |
High-risk systems must technically allow automatic recording of events over the system's lifetime. Providers and deployers keep those logs to the extent they're under their control. | ≥ 6 months |
| EU AI Act Art. 18 |
A separate duty covering technical documentation and conformity records — not the event logs. Frequently and incorrectly cited as a log-retention period. | 10 years |
| SEC / FINRA Rule 17a-4, FINRA 4511 |
Books-and-records rules for broker-dealers. Records must be preserved in a non-rewriteable form or under a compliant audit-trail alternative, with the most recent portion readily accessible. | 3–6 years by record type |
| HIPAA §164.312(b), §164.316(b) |
Requires audit controls that record and examine activity in systems containing ePHI. The six-year period attaches to required documentation — many organizations extend it to audit logs by policy. | 6 years (documentation) |
| PCI DSS v4.x Req. 10.5.1 |
Audit log history must be retained and available for analysis, with the most recent months immediately available. | ≥ 12 months 3 immediately available |
| GDPR Art. 5(1)(e) |
Sets no retention floor and pushes the opposite way: personal data must not be kept longer than necessary. Your trail needs a defined, defensible lifecycle — not indefinite storage. | minimize |
These obligations are pulling in opposite directions: sectoral rules require retention for years, GDPR requires minimization. The solution is the architecture above – retain the decision metadata and governance record for as long as it is necessary and hold the content under a separate, shorter, deletion-able lifecycle. You do not keep the evidence that a decision was made correctly without keeping a permanent copy of every person's personal data.
If you have read anything about AI logging in the last year, you have probably seen "August 2, 2026" as the deadline for high risk obligations such as Article 12 record keeping. The date has changed and a lot of published material has not.
The European Commission proposed the Digital Omnibus on AI in November 2025 after the implementation of the high-risk regime was clearly lagging behind – no harmonised standards were adopted and notified bodies were not appointed. The political agreement was reached in May 2026. The European Parliament formally endorsed it on 16 June 2026, and the Council gave its final approval on 29 June 2026, with entry into force following publication in the Official Journal. The effect on logging is direct: Article 12's obligations for standalone Annex III high-risk systems now apply from 2 December 2027, not August 2026, and embedded Annex I systems from 2 August 2028.
Three reasons why the extra runway is a trap if you think of it as time off. First, the deferral is narrow – Article 5 prohibitions, the Article 4 AI literacy duty, GPAI obligations and the Article 50 transparency rules were not affected and Article 99's penalty regime has been in force since 2025 August. Second, the legal risk that was at stake did not shift at all: AI-caused harm in 2026 is still subject to GDPR, sectoral financial and health rules, product liability and anti-discrimination law – none of which was paused. Third, the hard part of this job is not the documentation template. It is to find every AI system in your company, classify it and instrument it so that the logs exist. This takes quarters not weeks and it does not get easier by starting later.
There is no US AI logging statute. What there is is more important: technology-neutral rules that already apply and regulators who say they intend to apply them.
FINRA's 2026 Annual Regulatory Oversight Report added a dedicated generative-AI section — new for 2026. Per Sidley's analysis, FINRA expects firms to assess their regulatory obligations before deploying generative AI, to establish governance frameworks supervising its use, to maintain ongoing human monitoring of model outputs, and — most pointedly for audit trails — it notes that autonomous AI agents may require novel oversight, including tracking their actions and restricting system access. That is a logging requirement in all but name, arriving through existing supervision and books-and-records rules rather than a new one.
"Regulators don't fine you for the answer your model gave. They fine you for not being able to show how it got there. An unlogged AI decision isn't a compliant decision that lacks paperwork — for evidentiary purposes, it's a decision that never happened."
— Polygraf AI, on AI record-keepingPolygraf's Behavioral Control Plane is inline at the edge of all AI interaction, both user and agentic, which is exactly where a single audit trail has to be collected.Because it evaluates policy on every input and output, it records the governance layer most implementations miss entirely: which policy fired, what decision was reached (allow, redact, block), which rules matched, what was redacted, and which named, authenticated human was behind the request. It captures the decision chain, not the transcript.
And because Polygraf runs on-premise with zero data egress, the evidence chain never leaves your infrastructure. You are not relying on a third party's retention policy to meet a regulator's obligation on you. The logs are yours, created where the AI is actually running, sub 100ms latency and no GPU needed – that's the audit trail as a side effect of enforcement, not a project you need to fund and staff.
Polygraf records every AI interaction and every policy decision – allow, redact, block – with a named actor and tamper-evident record inline where your AI runs. On-premise, sub-100ms, zero data egress. The audit trail regulators want, automatically generated.
At Polygraf, we envision a future where AI augments human capabilities without compromising safety, privacy, or ethical standards. Trust in our commitment to building this future with you.
© 2026 Polygraf AI. All rights reserved.
Your download will start now.
Please provide information below and we will send you a link to download the white paper.
Data Privacy 
Data Provenance 
Content Generation