AI audit
Understand how your existing AI system works across code, data and connected tools. An independent audit of its architecture, permissions and tests, with clear priorities for what to fix before you rely on it.
Is your AI system ready for real users?
A system that works in a demo can still expose data, fail silently or run up model costs once real people use it. The audit shows you where those risks are before your customers find them.
An experienced engineer reads the code, runs the system and traces each request through prompts, model calls, data access and tool permissions.
Code, connections
and behavior
Bring your application or prototype and show the workflows it supports. The review traces requests through the code, model calls and integrations, checking how the system handles context, permissions, failures and handoffs to people.
Your team can use the findings directly, or we can fix them together as a separate integration project.
Deliverables
- Written findingsFindings from the code review and test results, with examples of the problems found.
- Issues in priority orderWhich problems to fix first and why.
- A test planTest cases, expected results and checks to run after changes. Implementing the tests is separate work.
- Recommended fixesWhat to keep, refactor or replace, with dependencies and effort estimates.
- A review meetingDiscuss the findings and answer questions with your technical team.
What audits usually find
Common issues in AI systems built quickly, often under deadline.
Tools with too much access
The agent can read or change more than the person it acts for, and nothing in the application stops it.
Changes tested by hand
Prompts and models change without a test set, so regressions reach users first.
Silent failures
Timeouts, malformed model output and API errors are swallowed or retried in a loop.
Untracked model costs
Token use per request has no limit or monitoring, so the bill grows with traffic.
How the audit runs
Most audits take one to two weeks, from access to the review meeting. We agree the scope and fee before starting.
- 01
Scoping call
30 minutes on what the system does, who uses it and what concerns you. You receive a scope and fee afterwards.
- 02
Review and test runs
With read access to the code and a test environment, we trace real workflows and run targeted checks on permissions, failures and costs.
- 03
Report and review meeting
Written findings in priority order, a test plan and recommended fixes, walked through with your team.
Questions
How long does an audit take?
One to two weeks for most systems, depending on the size of the codebase and the number of integrations. Larger systems can be audited one workflow at a time.
What access do you need?
Read access to the repository, a test environment or demo account, and an hour with someone who knows the system. The audit can run inside your environment.
Our code was written with AI tools. Is that a problem?
AI-generated code is welcome. Many of the systems we review were built with coding assistants or by a single developer, and you do not need an in-house engineering team.
Can you fix what you find?
Yes, as a separate project agreed after the report. The findings are also written so that your own developers or another provider can act on them.
Is our code kept confidential?
Yes. We can work under an NDA, inside your infrastructure and with test data, and agree access before starting.
Discuss a project
Describe the task, the software involved and what you need to change.
Request a 30-minute call