AI Workflow
Safety Review
Map the workflow
Test how it can fail
Fix the weak points
You know what could go wrong and what to fix first.
Questions worth answering
Have you implemented an AI tool, agent, chatbot or copilot that processes confidential data?
Does it have an evaluation suite?
Who has access to its logs?
Can you shut it off quickly?
Can it be manipulated by prompt injection, including indirect prompt injection?
Does it have agency: can it send email, fill in invoices or edit CRM records?
If so, can it do anything else, such as make refunds?
Does it require approval for consequential actions?
Can you be confident that it is retrieving documents from a database of documents you're happy to expose to AI?
Are there guardrails in place to make sure it's behaving appropriately?
What happens when it confidently provides an incorrect response to a question where incorrect advice may cause serious issues: does it refer to a human?
What the review covers
We perform an independent review of your AI agents, chatbots or other AI-enabled workflows that are in use by your business to assess their suitability and identify potential improvements. This includes assessing risks under the EU AI Act (generally applicable from 2 August 2026 with some high-risk obligations coming into force on subsequent dates) and other legislation where AI may be regulated, such as data protection legislation, equality legislation, consumer protection legislation or sector-specific legislation. If you're using AI anywhere in the EU or it produces any outputs into the EU, then you may well need to comply with the EU AI Act. We'll be able to advise.
What it tells your owners
This is NOT ISO 42001 certification but uses elements of ISO 42001 in order to structure the review. It will give your legal and other owners an idea of your exposure to AI-related statutory, contractual, reputational or other issues. It will tell you if you're happy for your AI-enabled workflows to see information from employees/customers/etc. It will identify places where the model should not be allowed to take action without a human stepping in and give your developers clear instructions on how to make it so. It will recommend suitable replacements for any shadow-AI that shouldn't be in use.
How we test
The review is based around the OWASP LLM Top 10 for 2025, extended to cover common issues we've seen with agentic systems in the wild. We will review your system prompts, RAG architecture and implementation, tool calling, logs, guardrails, LLM provider implementation and more. We will test your implementation, attempting prompt injections, both direct and indirect, jailbreaks, tenant escape, malicious use of tools, markdown image data exfiltration and more. Think of it as an extension of pen-testing to include prompts and outputs, including how your business handles them. We'll be accessing everything read-only and will provide a single report outlining findings ordered by importance, recommendations, suggested implementation, including code snippets for your developers, and effort required to implement them. We'll also confirm areas that were fine. This will be accompanied by a list of areas we did not cover and questions for you to consider. We'll happily answer questions afterwards, provide implementation support (additional cost) or connect you with partners who can help with anything else.
What it costs
It's not cheap and not easy. If you've deployed a company-wide Copilot but don't know who's using their own Copilot accounts, dodgy extensions or sketchy agents and automations, we need to talk. It will be priced on a fixed-fee basis and the fee will be agreed before we start. This will be based on the number of different workflows in-scope. Implementation support or other assistance with AI-related issues can be discussed at the same time so you're not stuck without support afterwards. We'll be brutally honest. We'd rather not waste your time or money.
Get in touch
Send us a list of your AI-enabled workflows with links and the technology/risk owners for each, a description of the tools they use and a couple of example prompts and we'll come back to you with next steps.