The AI Trading Audit Trail: What to Keep When a Model Touches Your Plan

When an AI model only answered questions, the record of a trade was simple: the plan, the order, the fill and the journal. When the model drafts the plan, checks the risk, reviews the order and, in an agentic setup, sends the order, the record becomes a chain. If any link is missing, the review is guesswork.

Regulators are making the same point. FINRA’s 2026 oversight report lists auditability and transparency as a core risk of AI agents: complicated, multi-step reasoning can make outcomes difficult to trace or explain. The CFTC’s staff advisory in December 2024 reminded registered entities that existing recordkeeping and supervision obligations apply when AI is used. For an individual trader, the lesson is not to imitate a compliance department. It is to keep enough evidence that you can answer a simple question a week later: what did I know, what did the model say, what did I check, what did I decide, and what happened?

This guide defines an audit trail for AI-assisted trading. It is part of the AI in Trading framework, and it is educational material, not legal advice. If you are a registered firm, your recordkeeping obligations are set by your regulator and your compliance team; this is a process aid, not a substitute.

If you cannot reconstruct the decision, you cannot review it. An audit trail is the difference between a lesson and a story.

What an Audit Trail Is Not

The phrase can sound bureaucratic, so it is worth being precise about what you do not need.

It is not a screenshot dump. A folder of charts with no context is not evidence; it is clutter.
It is not a P&L record. The result tells you what happened, not why you decided it.
It is not a full transcript of every chat. Most of a conversation is irrelevant. The record should be the decision, not the noise.
It is not a substitute for the journal. The journal captures the human experience; the audit trail captures the inputs, the model output and the approvals that led to the order.
It is not a compliance guarantee. It is a learning and accountability tool. If you are regulated, check the specific retention and recordkeeping rules that apply to you.

The Six Records

A complete trail has six records. Each answers a different question.

#	Record	Question it answers	Minimum fields
1	Input	What did the model see?	Instrument, contract, timeframe, timestamp, source, screenshot or data file
2	Prompt and model	What was asked, and by which system?	Prompt text or version, model name/version, settings, date and time
3	Output	What did the model produce?	Brief, plan draft, risk check, order list or critique
4	Verification	What did the human check?	Levels confirmed, figures recalculated, claims rejected, source links
5	Decision	What was approved?	Bias, scenario, order, size, stop, target, approver, time
6	Outcome	What happened, and what did you learn?	Fill, exit, R multiple, compliance score, emotional state, lesson

The six records are deliberately small. You do not need to store every token. You need enough to reconstruct the decision without relying on memory.

Record 1: The Input

The input is the chart, the calendar, the journal extract or the order list the model was given. It should be timestamped and traceable.

Instrument, contract month, timeframe and session.
Date and time of capture, in a stated time zone.
The source: platform, data feed or screenshot file.
Any redaction applied, and what was removed.
The file name or hash, if you want to prove the exact image later.

A chart without a timestamp is an illustration, not evidence. A chart with an account number in the header is a privacy incident. Redact first, then store. The privacy guide for AI memory and trading journals covers the redaction workflow.

Record 2: The Prompt and Model

If you change the prompt and the output changes, you need to know which version produced the decision. This is the equivalent of version control for your thinking.

Save the prompt text or a named prompt version (for example, “Brief v3 — observations only”).
Record the model name and version, and any settings that affect the output.
Record the date and time of the session.
Note any attached files, connectors or tools the model used.
Note whether memory was on or off.

You do not need to preserve every experimental prompt. You need the one that produced the plan you traded. A simple naming convention is enough: date, task, version. The AI Trading Prompt Pack is organised by task, which makes versioning easier.

Record 3: The Output

Store the output exactly as it was produced, not a cleaned-up summary written after the trade. The original shows what the model actually said, including the errors.

The brief, plan draft, risk check or order list.
The model’s stated uncertainties, if any.
Any levels or figures it produced that you later rejected.
The time the output was received.

Keeping the rejected output is useful. It shows whether the model has a recurring failure mode, such as inventing a level, confusing a contract or misreading a timeframe. Over a month, those errors become a pattern you can fix in the prompt or the workflow.

Record 4: The Verification

This is the most important record, and the one most traders skip. Verification is the human work that turns a model output into evidence.

Which levels did you confirm on the chart?
Which figures did you recalculate independently?
Which claims did you reject, and why?
Which source did you use for the calendar or the contract specification?
What did the model miss?

A useful format is a short verification line in the decision log: “Confirmed 29,505 and 29,635 on the 5-minute; recalculated size at 40 points × $2 × 3 = $240; rejected the 29,480 level as not visible; calendar includes 10:00 release.” That line is worth more than a page of narrative. The chart analysis pillar explains the evidence grades that make the check concrete.

Record 5: The Decision

The decision record is the bridge between the plan and the broker. It should be specific enough that a stranger could see what was authorised.

The scenario name and the branch (primary, secondary or no-trade).
The bias and the line in the sand.
The order type, size, entry, stop and target.
The risk in currency and as a percentage of equity.
Who approved the order, and when.
Whether the order was manual, approved by a human or sent by an agent.

If an agent places the order, the decision record should also include the agent’s name or identifier, the permissions it used and the limit that allowed the order. The agentic trading oversight guide sets out the control model; the audit trail is how you prove the controls were followed.

Record 6: The Outcome

The outcome record closes the loop. It should connect the result to the decision, not just report the P&L.

Fill price and time; any slippage.
Exit price and time; realized R multiple.
Plan compliance score (1–10).
Emotional state before, during and after.
What would have made the decision better?
What would have made it worse?

This is the material the weekly review needs. A profitable trade that broke the rules is a warning, not a success. The audit trail lets you see the difference. The journal fields guide explains why compliance is the field that matters most.

The One-Page Decision Log

You do not need a database. A single table, one row per decision, is enough for most discretionary traders.

Field	Example
Date / time	2026-10-22 08:42 ET
Instrument / contract	MNQ example contract
Input	60m and 5m captures; calendar; plan v4
Prompt / model	Brief v3; model name and version recorded
Output	Six-section brief; three open questions
Verification	Levels confirmed; size recalculated; one level rejected
Decision	Scenario A; 3 contracts; stop 40 pts; 2R target
Approval	Human approved at 08:51 ET
Outcome	Filled 08:54; stopped 09:20; −1R; compliance 9/10
Lesson	Entry was valid; stop was correct; no change

The log is not a trade diary. It is the evidence layer beneath the diary. Keep it short enough that you will actually fill it in.

Versioning Prompts and Models

Models change. Prompts change. If you do not record the version, a review can become misleading: you may attribute a change in output to the market when it was caused by a model update or a prompt edit.

Name prompts by task and version. “Brief v3 — observations only” is easier to audit than “the prompt I used last month.”
Record model updates. Note the date when the provider changes the model or the default settings.
Keep a change note. One line on why the prompt changed and what you expected to improve.
Re-test after a change. Run the new prompt on a known chart set and compare the output with the old version. The Backtest Checklist discipline applies to workflows too: change one thing, observe, record.

What Not to Keep

An audit trail is not a licence to hoard sensitive data. Keep the decision, not the exposure.

Broker credentials, API keys and passwords.
Full account numbers, card numbers and identity documents.
Raw statements with balances, addresses or tax identifiers.
Other people’s personal data or client information.
Live order IDs that could be used to access an account.
Screenshots that show account selectors or order panels you do not need.

Redact before you store. If the trail is compromised, the damage should be limited to your process, not your account.

Retention and Access

Define a retention schedule before you need it. A simple starting point:

Record	Suggested retention	Reason
Raw screenshots	30–90 days	Enough to review the recent period; large and sensitive
Decision log	12 months or longer	Comparable across regimes; useful for annual review
Prompt and model versions	As long as the decision log	Needed to interpret the output
Outcome and journal	12 months or longer	The learning record
Credentials	Never in the trail	Security

Store the trail somewhere you control, with access limited to you or your team. If you use a cloud drive, check the sharing settings and enable two-factor authentication. If you are subject to specific recordkeeping rules, follow those instead of this suggestion.

The Ten-Minute Weekly Audit

Once a week, before the review, check the trail.

[ ] Every traded decision has an input, a prompt version, an output and a verification line.
[ ] Every order has a stop, a size and a risk figure.
[ ] The decision log matches the broker statement.
[ ] Any agent activity has an approval or a limit that authorised it.
[ ] Rejected model claims are recorded, not deleted.
[ ] No credentials, account numbers or unnecessary personal data are stored.
[ ] The prompt and model versions are current.
[ ] The week’s records are backed up.

If a record is missing, do not reconstruct it from memory. Mark it missing. The gap is itself evidence: it shows where the process is weak.

Frequently Asked Questions

What is an AI trading audit trail?

A record of the inputs, prompts, model versions, outputs, human checks, decisions and outcomes behind an AI-assisted trade. Its purpose is to make the decision reviewable and the process improvable.

Do I need an audit trail if I trade my own account?

You do not need one to satisfy a regulator unless specific rules apply to you. You do need one to learn from your decisions and to detect drift. Without it, the weekly review relies on memory, and memory is biased by the outcome.

How long should I keep the records?

It depends on your situation. A practical starting point is 30–90 days for raw screenshots and 12 months or more for the decision log, prompts and journal. If you are regulated, follow the applicable retention rules.

Should I store the full chat transcript?

Not necessarily. Store the prompt, the output and the verification. A full transcript is noisy and may contain sensitive data you do not need. If the reasoning matters, keep the relevant excerpt.

How do I audit an agent that acts on its own?

Require an activity log and a daily reconciliation. The log should show the input, the action, the permission used and the result. The reconciliation compares the log to the broker statement and to the limits you set. If the agent cannot provide that, reduce its authority.

Does an audit trail improve performance?

Not directly. It improves the quality of the review, which is the mechanism by which process changes. A better review does not guarantee a better return; it makes the reason for the result easier to identify.

Next Step

Start with one row per decision. Use the Trading Journal Template for the outcome and the one-page decision log structure above for the inputs and approvals. If you use an agent, read the oversight guide first and make the log a condition of access.

Related Guides

Sources

FINRA, 2026 Annual Regulatory Oversight Report, “GenAI: Continuing and Emerging Trends.” Auditability and transparency are identified as AI agent risks; multi-step reasoning can be difficult to trace or explain. finra.org
FINRA, “Emerging Trend in GenAI: Observations on AI Agents” (27 January 2026). finra.org
US Commodity Futures Trading Commission, Staff Advisory No. 24-17, “Use of Artificial Intelligence in CFTC-Regulated Markets” (5 December 2024). Existing recordkeeping and supervision obligations apply to AI use. cftc.gov
US Commodity Futures Trading Commission, Customer Advisory: AI Won’t Turn Trading Bots into Money Machines (25 January 2024). AI cannot predict the future; guaranteed returns are a fraud red flag.
European Commission, “Timeline for the Implementation of the EU AI Act.” Transparency rules (Article 50) apply from 2 August 2026; high-risk obligations follow on a later timeline. europa.eu
NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), for the govern-map-measure-manage structure and the human-review principle. nist.gov

This article describes a process aid, not a regulatory recordkeeping programme. If you are a registered entity, consult your compliance team and the applicable rules. No statistic in this article is invented.

Trading involves substantial risk of loss and is not suitable for every person. This article is educational information, not financial, investment, legal or tax advice, and it is not a personal recommendation or solicitation. AI output can be inaccurate or incomplete; every level, figure, order and record must be verified by the trader. Past performance and backtests do not guarantee future results. Never trade with money you cannot afford to lose.