The Raisy Blog
AI Agent Audit Trail: Every Action Leaves a Receipt
Every bot call and process run writes an immutable, dated log row. Here is how Raisy turns AI audit trails into something you can point to.
An AI that works for your company but does not log its work is not an employee. It is a risk. When a machine acts on your behalf, the only thing that makes that safe is the ability to reconstruct what it did, when, and why. Raisy treats that trail as a product requirement, not a compliance chore: every bot call, every process run, every approval, and every promotion writes an immutable, dated log row you can point to.
What we set out to change
Two weaknesses come up again and again with AI assistants. The first is opacity: a model produces an answer, and nobody can say which rule, which input, or which run produced it. The second is drift: you have no way to tell whether a bot you approved three weeks ago is still doing what you approved, because nothing records the boundary it crossed.
The standard answer in the market is to show a chat transcript. A transcript is not an audit trail. It shows what a person typed and what a model replied; it does not show the process that ran, the gate that was passed, or the version of the rules that applied. (This is the same standard we hold the bot lifecycle to, and the reason every run in a process can be reconstructed. Browse the blog.)
The idea / how Raisy fixes it
Raisy logs the system, not the chat. Every execution is an event with a type, a timestamp, an actor, and the version of the rules that governed it:
- A bot call records which bot ran, which kernels were loaded, what it prepared, and what mode it ran in.
- A process run records which graph of steps executed, in what order, and where it paused.
- An approval records who approved, edited, or rejected a prepared action.
- A promotion records when a bot moved from MANUAL to SEMI_AUTO to FULL_AUTO, and why.
Each row is appended, not overwritten. A kernel that changes creates a new version; the old version stays in the trail. Nothing a machine did for you silently disappears.
What the user actually sees
Open a bot's record and you see its life as a dated event stream: hired, called, gated, approved, edited, promoted. Each row carries a timestamp, the actor (bot or human), the action taken, its outcome, and, where a gate was involved: the approver. Each event expands to its detail: the input, the prepared output, the gate that passed. You can replay what happened without trusting a summary.
Figure 1. Documented AI incidents nearly tripled in two years (123 → 362). Every incident on this curve is an event where someone needed a receipt. Source: Stanford HAI AI Index 2024–2026.
AI Incident Database backfills historical years; per-year totals can be revised upward (AI Index 2026 note).
The chart above is not a product screenshot; it is the market context. Documented AI incidents nearly tripled in two years: 123 in 2023, 233 in 2024, 362 in 2025, per the AI Incident Database figures reported in Stanford HAI's AI Index Reports. Every one of those incidents is an event where someone needed a receipt and either had one or did not. Regulation is moving the same way: the EU AI Act (Regulation (EU) 2024/1689) requires high-risk AI systems to automatically record event logs over the system's lifetime (Article 12), with the regulation applying in full from 2 August 2026 and the Article 6(1) obligations following on 2 August 2027. Automatic logging is becoming law, not preference.
Why this matters beyond convenience
The promise is auditability by design, not as an afterthought, the same standard an ISO 9001 process would demand. When a machine acts on your behalf, the trail is what turns it from a tool you gamble on into a workforce you can stand behind. It is also what makes the human gate meaningful: a reviewer with no way to reconstruct a run has no basis for approving it.
FAQ
Does this apply to my strategy or my department?
Yes. The log is per entity: your company owns its rows, its bots, its process runs.
Can a bot act without a receipt?
No. An external action always passes a gate, and every gate writes a row. Autonomy changes how often you review, never whether the action is recorded.
Is the log editable?
Rows are append-only. A correction is a new row that references the original; the original stays in the trail.
What happens when I promote a bot?
The promotion is logged with its mode and timestamp, so you can always tell when and why a bot’s latitude changed.
Sources
Stanford HAI AI Index Report 2024, chapter 3 (https://hai.stanford.edu/assets/files/hai_ai-index-report-2024_chapter3.pdf); Stanford HAI AI Index Report 2025, chapter 3 (https://hai.stanford.edu/assets/files/hai_ai_index_report_2025.pdf); Stanford HAI AI Index Report 2026, chapter 3 (https://hai.stanford.edu/assets/files/ai_index_report_2026.pdf): AI Incident Database counts of documented AI incidents per year (123 in 2023, 233 in 2024, 362 in 2025). Regulation (EU) 2024/1689 (EU AI Act), Articles 12 and 113 (https://eur-lex.europa.eu/eli/reg/2024/1689/oj). Raisy System Manifest (MR0: review gate before external effects; MR0a: every action, output, and state change is logged). Execution rows are reproducible from Raisy’s immutable audit log of bot and process calls once the public execution-log API ships: pending and logged as a setup task before publication.
Keep reading
Multi-Agent Supervisor Architecture, Compared
J.P. Morgan runs one supervisor over three sub-agents. Raisy runs a Chief of Staff over process rooms that own state. Where the two designs diverge.
An OpenAI Dots Alternative for Companies
Dots gives one person their own computer. A company needs one owner for a function, a gate before anything leaves, and a number somebody watches.