What your AI did, and what it cost, in records you can check.
Aribot writes every governed agent decision to an audit trail you can check for later edits, and measures AI spend per call by model, feature and agent. Audit, finance and the AI team work from the same records.
Two questions arrive after every AI launch
Once agents and assistants are working inside the business, two questions keep coming back. The first comes from audit and risk: what did this agent do, who allowed it, and can you prove the record has not been tidied up since. The second comes from finance: what is this costing, which team or agent is driving it, and will next month be higher.
In most organisations the answers are spread across places that were never meant to be evidence. Application logs can be edited and are rotated away. Provider dashboards show tokens per account, with no link to the agent or feature that spent them. Invoices arrive weeks after the spending happened. When an auditor asks for a sample, someone exports a spreadsheet, and the spreadsheet becomes the record.
Aribot keeps both answers in the product, next to the agents they describe. The AI audit trail holds every decision the governed gateway makes about an agent, in a form anyone with access can check for later edits. AI cost shows what those agents and your users spend on AI, measured per call, broken down by model, feature and caller, with a budget that warns you early. The two screens link to each other: each paid action is recorded with the decision that allowed it.
Who uses these screens
The same records serve four readers, and each comes to them with a different question. Nobody has to translate one team's numbers into another team's report.
Internal audit and compliance
You sample decisions and ask for proof. The AI audit trail gives you every governed agent decision in order, a check you can run yourself, and an evidence pack with a sign-off block. You no longer have to trust a screenshot someone took on the day.
Finance and FinOps
You need AI spend you can forecast and allocate. AI cost shows measured spend by model, by feature and by caller, a monthly limit with an alert threshold, and a report you can export. Estimated savings sit in their own column, away from measured spend.
AI lead and head of platform
You run the agents and answer for them. When a cost jumps or a block appears, you can see which agent, which model and which task, then fix the cause: a cheaper model for one task, prompt caching, or a narrower permission for one agent.
CISO
You need to know that a blocked action stayed blocked and that nobody rewrote the record afterwards. Blocks, quarantines and reduced access are on the same trail as everything else, and the trail can be checked at any time.
The AI audit trail: every agent decision, in order
Every call an agent makes into your systems through Aribot passes the governed gateway. The gateway decides whether the call goes ahead, and that decision is written to the AI audit trail, whether the answer was yes or no. Decisions people make about agents are written there too, such as a block or a reduction in access from the actions queue.
The trail opens with a count of decisions and the event types they fall into. Each event type is a chip you can click to filter, so a reviewer looking for blocks sees only blocks. Each row shows the event, the agent, when it happened and a record ID, and where a row names an agent there is a link straight to that agent's assessment.
What gets recorded
Fixes and the checks behind them
When Aribot's governed self-healing fixes a threat model, the fix is sealed on the trail with the records that support it: the impact check it passed, the sign-off that approved it and the attack test it survived. The screen counts these separately and lets you filter by kind, so ten fixes with thirty supporting checks read as ten fixes. Threat model fixes are listed here; cloud and code fixes keep their own records in self-healing.
Check the record yourself, at any time
A log is only useful as evidence if a reader can tell whether it has been changed. Each record on the AI audit trail is linked to the record before it. Press Check records and the whole history is checked in one pass. If any entry was edited, removed or put out of order after it was written, the result reads Check failed. If nothing changed, it reads Record intact.
The check is honest about an empty trail as well. A history with no entries has nothing to contradict it, so a naive check would report it as verified. Aribot shows No entries recorded instead, and the evidence pack says the same. An auditor reading the pack can tell a quiet month from a missing record.
We describe the trail as tamper-evident. Detection is the promise you can test: anyone with access can run the check, as often as they like, and see the same result.
Taking the record out of the product
- Export CSV gives the decisions you are looking at as rows: event, agent, time, record ID and sequence.
- Download data gives the full content of the page as a file: the integrity result, the decisions, the sealed fixes and the AI spend report.
- Evidence pack opens a printable report: when it was generated, how many decisions and fixes it covers, whether the integrity check passed, then each section in turn and a sign-off block at the end. Any section that failed to load is marked as such.
AI cost: what you spend, and where
AI cost opens on the spend dashboard. It shows spend for the last 30 days, the number of calls and how many of them have a price, total tokens and how many were served from cache, average latency, a daily trend, and spend by model and by provider. Every figure comes from calls that were metered as they happened.
The usage view breaks the same spend down four ways: by feature, by model, by task and by caller. The highest-cost row in each table is highlighted, which is usually where a review should start. Calls where the provider reported no token usage are flagged as unknown, so a gap in the data never shows up as a zero.
The prices behind the figures are on view. Each model has a price per million input, output and cached tokens in the Pricing tab. A call to a model with no price yet is counted as unpriced and kept apart from measured spend, and an efficiency finding points it out.
The tabs, and what each answers
Findings you can act on
Efficiency findings look for patterns that waste money, such as a large model doing work a smaller one handles well, or prompts sent again and again without caching. Each finding says what it means, why it matters and what to do, with a saving per month and the number of times it occurred. From a finding you can raise a ticket in your tracker, open the usage behind it, or open the AI configuration it refers to. Every saving is labelled measured, modelled or unpriced, and none counts until you act on it.
Budgets that warn you while there is still time
AI spend follows usage, and usage can climb quietly: a new agent, a longer prompt, a loop that calls a model more often than anyone expected. A monthly invoice tells you after the month is over. A budget in Aribot tells you while you can still change something.
Budgets & Alerts holds the monthly AI cost limit that applies to your company, set by your platform administrator, and an alert threshold that a company owner chooses as a percentage of it. The tab shows spend month to date against the limit, coloured as it approaches and passes the threshold, the warnings that have been raised, and any overage for the month. When an alert arrives, the usage and findings views show which agent, model and task drove the increase.
The forecast tab looks ahead. It projects future AI and security spend, costs a proposed change and assesses its business impact, and keeps each of these as an analysis you can return to when the quarter closes.
A quarter at Tellmoor Freight
Tellmoor Freight is a fictional mid-sized logistics company. Every name and number in this story is illustrative. Tellmoor runs four AI agents through Aribot: one reads supplier invoices, one drafts customs paperwork, one writes weekly reports for operations, and one reviews code changes for the platform team. Its people also use Ayurak Intelligence in the product.
Week three: the budget alert
Tellmoor's owner set the alert threshold at 80 percent of the company's monthly AI cost limit. On day nineteen the alert arrives, with ten days of the month left. The finance lead opens Usage & Activity. In the By caller table the invoice-reader agent is the top-cost row. In the By model table, most of that spend is on the largest model available. In the By task table, the work is classifying invoice lines into a handful of categories.
Efficiency Findings already has a finding for the pattern: a large model used for short classification work, with a modelled monthly saving. The AI lead raises a ticket from the finding into the platform team's tracker. The team moves the classification step to a smaller model and turns on prompt caching for the fixed part of the prompt. Over the following days the forecast for the month flattens and stays under the limit. On the spend dashboard, the share of tokens served from cache goes up.
Week seven: the access question
During a routine review the CISO's team notices, in the agent's assessment, that the report-writer agent holds write access to a shared drive it has never used. In the actions queue they reduce its access. The current safe action finishes, then the unused permission is removed, and the decision is written to the AI audit trail with the reason the reviewer gave.
Week twelve: the internal audit
Internal audit samples ten agent decisions from the quarter, including the reduced access and two blocked calls from the customs agent. The auditor presses Check records and gets Record intact. The compliance lead opens the evidence pack, prints it to PDF with the decisions, the fixes and the AI spend for the quarter, and adds the sign-off. The same file is attached as evidence to the matching questions in Tellmoor's detailed assessment on Ayurak Trust.
Nobody rebuilt a spreadsheet for the audit committee. The numbers finance used in week three are the numbers the auditor saw in week twelve.
Why the record and the cost sit side by side
A decision without its cost leaves finance guessing, and a cost without its decision leaves audit guessing. Aribot records each paid action with the decision that allowed it. The AI audit trail shows a summary of AI spend for the period, the evidence pack includes it, and AI cost links each external agent back to its own assessment and scope.
That lets one review cover both. When an agent's spend rises, you can see whether its authority grew at the same time. When an agent is blocked or its access is reduced, you can see what that did to the bill. The review meeting starts from one set of figures.
From the free self-assessment to Trust
Most organisations start with the free AI readiness self-assessment, which shows where policy and process stand against the EU AI Act, the NIST AI RMF and related frameworks. Aribot then governs the AI that is running and keeps the audit trail and cost records described here. A detailed assessment on Ayurak Trust, done with Ayurak's team, asks for evidence behind each answer. Exports from the audit trail and AI cost are the kind of evidence those questions ask for.
The standards this evidence supports
The following public standards and frameworks ask organisations to keep records like these, or to be able to show them on request. Meeting each one still depends on your wider programme.
Article 12 asks that high-risk AI systems allow events to be recorded automatically over their lifetime, and Article 26 asks deployers to keep those logs, where they control them, for at least six months. The AI audit trail is a durable record of what governed agents did and what was refused, and the blocks and quarantines on it show human oversight being used. Whether a system is high-risk, and what the Act requires of you, is a decision for your own legal advice.
Measure asks you to track the metrics that tell you how AI is behaving over time. Manage asks you to put resources where the risk is and to document how risks were treated. Measured spend by model and agent, efficiency findings and the trail of blocks and access changes give both functions something concrete to point to.
An AI management system needs monitoring and measurement, internal audit, and documented information kept as evidence. The audit trail and its evidence pack serve the last two, and AI cost with budgets serves the first for the resources your AI uses.
Agents are identities with access, and auditors test logging and monitoring of access in both schemes. A record of every permission granted, used and withdrawn, which can be checked for later edits, answers the questions they already ask about people.
Assistants and agents that reach Aribot over the Model Context Protocol, signed in with OAuth, pass the same gateway checks as everything else. Their decisions and their model calls appear on the same trail and in the same cost views, by caller.
This page refers to the public text of each standard and regulation. It is general information and is not legal advice.
Questions we are asked
- What ends up on the AI audit trail?
- Every decision the governed gateway makes about an agent: onboarding, access granted, permission used, action blocked, quarantine and access withdrawn, plus governed automatic fixes with their supporting checks. Each entry shows the event, the agent, the time and a record ID. You can filter by event type and export what you are looking at as CSV.
- What does Check records prove?
- That no entry has been changed, removed or reordered since it was written. Each record is linked to the one before it, so the check covers the whole history in one pass and fails if any part of it was altered. A trail with no entries says No entries recorded. It never reports an empty history as verified.
- Can we hand this to an auditor?
- Yes. The evidence pack brings the decisions, the automatic fixes and the AI spend for the period together, states whether the integrity check passed, and ends with a sign-off block. You can print it or save it as PDF. Download data gives the same content as a file for your own tools, and Export CSV gives the decisions as rows.
- Where do the AI cost figures come from?
- From the model calls your agents and users make through Aribot, metered per call, and priced from the per-model token prices held in the Pricing tab. Calls for which the provider reported no token usage are marked as unknown, and calls to a model with no price yet are counted as unpriced until a price is added.
- What is the difference between measured, modelled and unpriced savings?
- A measured saving is calculated from calls that already have a price. A modelled saving is an estimate of what a change would save. An unpriced saving belongs to calls with no price yet, so its value is unknown until a price is added. Savings count only once you act on them, and the screens keep the three apart so an exported figure is one you can defend.
- Can we set a budget and be warned before we hit it?
- Yes. Budgets & Alerts shows the monthly AI cost limit that applies to your company, which your platform administrator sets, and a company owner chooses the alert threshold as a percentage of it. The tab then shows spend month to date against the limit, the warnings that have been raised and any overage.
- Can we see what each agent costs?
- Yes. The By caller view in Usage & Activity separates external agents, which connect as OAuth or MCP clients, from requests made by signed-in people and from background work with no caller. An agent row opens that agent's assessment and scope, so cost and authority are reviewed together.
- Does this cover AI spend outside Aribot?
- The audit trail and AI cost cover the decisions and the model calls that pass through Aribot. AI your teams use elsewhere is found through shadow AI discovery in your repositories and captured as policy through the free AI readiness self-assessment. If you need a wider view of spend, book a call and we will look at your set-up with you.
- Is this legal advice on the EU AI Act?
- No. This page describes evidence the product keeps and refers to the public text of the regulation. Whether a system is high-risk, which role you hold and what you must keep are questions for your own legal advisers. The EU AI Act readiness page shows how Aribot assembles a dossier for each AI system.
Where you will find it once signed in
Demos and related pages
How the AI governance pages fit together.
Fleet, entitlements, scopes and the actions queue.
From finding AI and its access to fixing it, with proof.
A conformity dossier for each AI system.
Find the AI nobody registered.
Watch a call pass the gateway and land on the trail.
Governed fixes, sealed with their checks.
Detailed, evidence-backed framework assessments.
Start with the free self-assessment. Go deeper with evidence.
- Step 1, free
Answer the AI readiness self-assessment online. You get a view of where you stand against the EU AI Act, the NIST AI RMF and related frameworks, in minutes.
- Step 2, with Ayurak's team
A detailed, evidence-backed assessment on Ayurak Trust: each answer supported by evidence, reviewed with our team, and reported in a form you can hand to a board or an auditor.
Start with the records you already have
Connect an agent or an assistant to Aribot and its decisions and spend appear on these screens from the first call. Begin with the free AI readiness self-assessment, or book a call to walk through the audit trail and AI cost with Ayurak's team.
