Working together
Three fixed-scope starting points, the four ways larger engagements are usually shaped, and the practical questions procurement asks, so you can establish early whether this will work. Cairn Cyber Advisory Ltd contracts on a business-to-business basis.
Starting points
Named, fixed-scope engagements for the questions that come up most. Fixed fee, agreed in writing after a short scoping call.
AI tool approval pack
Typically 2 to 3 weeks
What would it take for this tool to be approved, and what does the evidence look like?
- For
- A business lead whose team has chosen or bought an AI tool that must pass security and risk review.
- You provide
- Access to the vendor documentation and contract, the intended use and data, the configuration, and a named contact in security or risk.
- You receive
- A threat model of the tool as you will deploy it
- A control design
- An evidence pack mapped to your review process
- An executive decision brief with residual risk and a recommendation
The decision stays with your reviewers. I won't also act as the independent reviewer of a system I prepared for approval.
Example, fictional
Larchwick Mutual's claims function wanted to pilot a vendor-hosted assistant that summarises claim files and drafts letters to members. The pack gave its Technology Risk Committee six rated risks, a control for each, and a recommendation.
Recommendation: Approve with conditions, for a 90-day pilot on motor and property claims. Injury claims stay out until UK-region processing is evidenced.
See a sample executive decision brief →
Fixed fee, agreed in writing after a short scoping call.
AI vendor assurance pack
Typically 2 to 3 weeks
What will a bank's review ask of our product, and how strong are our answers?
- For
- AI vendors and fintechs selling into banks, asset managers and custodians, whose deals stall at the buyer's security and architecture review.
- You provide
- Your architecture documentation, the security features as they exist today rather than on the roadmap, and working sessions with your CTO and whoever owns security to settle the answers.
- You receive
- A threat model of the product as a bank would deploy it
- An answer sheet to the questions a major bank's security and architecture review will ask, worked through with you so the answers are yours and hold up
- A control mapping to the frameworks those buyers work to, including ISO 42001, the NIST AI RMF and DORA's third-party expectations
- A gap list in priority order: where your answers are weak today, and what would make them strong
I have written AI security standards and run approvals from the buyer side at a global custodian bank and a global asset manager, so I know what reviewers ask for and what evidence satisfies them. The pack is yours to present. It is not an independent assurance opinion, and I won't review the same product for a buyer.
Example, fictional
A Series A fintech sold an AI assistant that summarises client onboarding files for relationship managers. Two bank deals had stalled at design review, on the same questions.
Gap 1: Two controls a bank's reviewers will ask about are on the roadmap, not in the product. Say so in your answers, with the delivery dates and the interim controls, so the reviewer hears it from you rather than finds it.
Fixed fee, agreed in writing after a short scoping call.
Microsoft 365 Copilot readiness review
Typically 1 to 2 weeks
What will people be able to find on day one that they shouldn't, and what has to be contained first?
- For
- Organisations about to switch on Copilot or a similar assistant across Microsoft 365.
- You provide
- Read access to the Microsoft 365 admin and Purview reporting, or someone who can run the reports with me; any sensitivity labelling and retention policies; and a named contact in IT.
- You receive
- An oversharing and permissions exposure assessment
- A containment plan in priority order
- Go-live conditions
Example, fictional
Before switching on Copilot for 300 head-office staff, Larchwick wanted to know what it would surface. The review found board papers and HR case files on SharePoint sites shared with every employee.
Go-live condition: Copilot is not enabled for anyone until the 14 sites shared with all staff are restricted to their members, starting with the board and HR sites.
Fixed fee, agreed in writing after a short scoping call. Background: Before you switch on Copilot.
Agent tool-access review
Typically 1 to 2 weeks
What can this agent actually reach and change, whose authority does it act on, and what would stop it?
- For
- Teams putting an AI agent into production, or switching on agent features in a product they have bought, where the agent can call tools, MCP servers or connectors into your systems.
- You provide
- The agent's design and tool list, the MCP server and connector configuration, the identities and permissions it runs with, its logs, and a named contact in the team that built or bought it.
- You receive
- An inventory of every tool, MCP server and connector the agent can call, and what each can read and change
- An authority map: whether each call runs as the agent or as the user, with what scope, and who consented
- A blast-radius assessment for prompt injection and tool misuse, in priority order
- A control design: scope reductions, approval steps for actions that change data, and the invocation logging a reviewer will ask for
The review covers the agent's access as configured. It is not a penetration test of the tools themselves.
Example, fictional
Before the reserve-proposal agent went live, Larchwick asked what else it could reach. It called four tools through the vendor's MCP connector, and one of them could read every claim file in the system.
Priority 1: The connector acts as the signed-in handler, with that handler's claim permissions, not as one service account that can open every claim. Until the vendor supports that, the agent stays switched off.
Fixed fee, agreed in writing after a short scoping call. Background: Agentic AI breaks your identity model.
AI security baseline for startups and smaller firms
Typically 1 week
What AI is in use across the firm, what protects our data in it, and what should we fix first?
By the end of the week you have a register of the AI in use, each tool mapped to the controls that protect your data and where they fall short, and a prioritised list of next steps.
- For
- Seed and Series A fintechs, and smaller firms including law firms, using AI tools without a security function.
- You provide
- An hour with the owner or managing partner, a list of the software and subscriptions in use, and access to whoever runs your IT, in-house or outsourced.
- You receive
- An AI register: each tool, who uses it, and what data goes into it
- A control mapping: for each tool, the settings, contract terms and access controls that protect your data, and the gaps
- A prioritised list of next steps, with an owner for each
- A one-page summary for partners or the board
The register covers the AI your people declare and your systems show; finding undeclared use through network monitoring is outside a one-week baseline. Controls are mapped against practical settings and contract terms, not a full framework such as ISO 42001.
Example, fictional
Tolland Rae, a 20-person Edinburgh law firm, knew its people used AI but not which tools, or with what client data. By the end of the week it had a register of nine tools, three of them used on personal accounts.
Next step 1: Move everyone using a personal AI chatbot account onto the firm's business licence, so client data is no longer used to train the vendor's models. The practice manager owns this and has two weeks to complete it.
Fixed fee, agreed in writing after a short scoping call.
How an engagement runs
Every engagement, fixed-scope or not, follows the same four steps.
Agree the question and the decision criteria
Before any analysis, you and I agree what is being decided and what a good answer looks like.
Output: A written scope, and the criteria the decision will be judged on.
Gather the evidence
Designs, configuration, contracts, and short interviews with the people who own them.
Output: An evidence index.
Analyse
A threat model, findings rated by risk, and the controls that close them.
Output: Findings and control design.
Hand over
The result is written for the people who decide and for the people who maintain it.
Output: A decision brief for executives, and working artefacts your team can maintain.
Larger engagements
For work that does not fit a starting point. Each is scoped in writing before anything starts.
Discovery
Where do we actually stand, and what should we do first?
- Typical scope
- One programme, platform or area of AI adoption, reviewed against the frameworks you are held to.
- What you provide
- Existing designs, policies and risk registers; short interviews with the people who own them (typically four to eight, 45 minutes each); a sponsor who can unblock access.
- What you receive
- Written assessment, prioritised risk picture, and a costed recommendation on what to do next.
- Done when
- The assessment and recommendation have been walked through with your sponsor.
- Indicative duration
- 1 to 2 weeks, depending on scope and access to evidence
- Who decides
- Your organisation. I provide a recommendation.
Example, fictional
After the claims pilot, Larchwick's board asked where the mutual stood on AI across the business. Discovery reviewed its policies, risk register and the AI tools in use, with six interviews.
First priority: Give every AI tool an owner. Eleven of the 26 tools in use have none, including two that process members' personal data.
Architecture review
Should this design, platform or third-party system go ahead, and on what conditions?
- Typical scope
- One defined system at a decision point, with its integrations, data flows and third-party dependencies.
- What you provide
- Design documents and data flows; configuration or infrastructure-as-code where it exists; vendor documentation and contracts; time with the architects and engineers who built it.
- What you receive
- Threat model, findings rated by risk, and an approve / approve-with-conditions / reject recommendation.
- Done when
- The recommendation and any conditions are delivered in writing and presented to the forum making the decision.
- Indicative duration
- 2 to 4 weeks, depending on scope and access to evidence
- Who decides
- Your organisation. I provide a recommendation.
Example, fictional
The claims assistant's vendor offered an agent that could update claim records and change reserves, which the original approval had ruled out. Larchwick asked whether to allow it.
Recommendation: Approve with conditions. The agent may propose a reserve change but not make one: a handler with authority for the amount approves each change, and every proposal is logged against the claim.
Embedded advisory
How do we get security architecture right while the programme is still making its decisions?
- Typical scope
- A named workstream, such as the security architecture for an AI platform, a cloud target state, or a set of standards.
- What you provide
- A seat in the programme's design authority and governance meetings; access to designs and repositories; a named lead I report progress to.
- What you receive
- Target-state security architecture, design decisions recorded with their rationale, security requirements for delivery teams, and a handover pack your team can maintain.
- Done when
- The agreed deliverables are complete and handed to a named owner in your team.
- Indicative duration
- 3 to 12 months, part or full time, depending on scope and access to evidence
- Who decides
- Your organisation. I provide a recommendation.
Example, fictional
Larchwick built an internal AI platform, so teams could use models without each buying a tool. The engagement ran for six months, part time, in the platform's design authority, and ended with a handover to the platform team.
Design decision: Every model call from the platform goes through the gateway, which enforces UK-region processing and logs each prompt and response against the calling service. One place to enforce the data residency policy, and one place to evidence it.
Retained advisory
Who do we call when a design question or escalation needs a senior security architect?
- Typical scope
- Reviews, escalations and design input, up to the agreed days each month. Response times are agreed in the retainer.
- What you provide
- A named point of contact, and notice of what is coming up in the month ahead.
- What you receive
- Agreed availability for reviews, escalations and design input, with a standing point of contact.
- Done when
- Each month closes with a short note of what was reviewed and advised. The retainer runs until either side ends it on the agreed notice.
- Indicative duration
- Ongoing, defined days per month, depending on scope and access to evidence
- Who decides
- Your organisation. I provide a recommendation.
Example, fictional
Once the platform was live, Larchwick kept two days a month for reviews and escalations.
Monthly note: Reviewed the vendor's new sub-processor notice: no change to where data is processed, no action. Escalated one item: the pricing team's proof of concept calls a model directly, bypassing the gateway.
Fees are fixed and agreed in writing after scoping. Embedded and retained engagements are a fixed monthly fee.
Contracting and IR35
Engagements are contracted business-to-business through Cairn Cyber Advisory Ltd and are intended to operate outside IR35. I work to defined deliverables rather than a role, direct my own working methods, and am not integrated into your organisational structure.
I expect to review the Status Determination Statement before contracts are signed, and I am happy to discuss how the engagement should be structured so the determination reflects how the work will actually be performed.
This describes how I contract. It is not tax advice, and the status determination remains the client's responsibility.
Capacity and continuity
I take on a limited number of concurrent clients, so the time agreed at scoping is the time you get. Planned absence is agreed at scoping.
Working material is kept current on your systems wherever your policy allows, so the work is never held only in my head. Every engagement ends with a handover your team can maintain.
Insurance
- Professional indemnity
- Arranged per engagement, to £5m
- Public liability
- Arranged per engagement, to £5m
Cover is placed for each engagement, and the certificate is provided at contracting. Higher limits can be arranged where a framework requires it.
Onboarding
- Initial enquiry. Use the contact form with an outline of the problem and timescale.
- Call. I will reply with a link to book a 30-minute call if it looks like a fit. No charge, no obligation.
- Scope and proposal. Written scope, deliverables, timescale and price.
- Contracting. Your paper or mine. NDA first where you would prefer.
- Vetting. I hold a clean record and can complete standard financial services screening, including BPSS, credit and criminal record checks and employment referencing.
Supplier due diligence
Onboarding a supplier usually means a security questionnaire, and a security practice that answers one badly is answering the wrong question about itself. Most of what is asked is already published here: the capability statement carries the company, insurance and registration detail in a form you can attach to a procurement record, and the privacy notice states exactly what personal data the practice processes and for how long.
On the technical side: no client data is stored on this website, which has no database, no cookies and no analytics scripts. The one third-party script is Cloudflare Turnstile, which the contact form uses for spam protection; nothing else loads from a third party.
The enquiry form posts to a single function that verifies the submission, relays it by email and persists nothing. Client material is handled under the terms of the engagement contract, on the client's systems wherever the client's policy allows it.
A vulnerability disclosure contact is published at /.well-known/security.txt, which is a fair thing to check before hiring anyone to review your architecture.
Common questions
- Do you work outside IR35?
- Yes. Engagements are contracted business-to-business, worked to defined deliverables rather than a role, and are intended to operate outside IR35. I review the Status Determination Statement before contracts are signed.
- Will you work on our paper or yours?
- Either. I am happy to work on your standard consultancy terms, and I can provide my own where you would prefer. An NDA can be signed before any detail is discussed.
- Can you complete our vetting?
- Yes. I hold a clean record and can complete standard financial services screening, including BPSS, credit and criminal record checks, and employment referencing.
- Do you work remotely, on site, or both?
- Both. The practice is based in Edinburgh and works remotely across the UK, with on-site time where the work genuinely needs it. Workshops, design authority sessions and stakeholder work usually do.
- Do you subcontract or use associates?
- No. The work is led and delivered by me, so you get senior advice directly rather than through an account team or a junior bench. Where a piece of work needs a specialism I do not hold, I will say so and help you find it rather than cover it thinly.
- Will you be named as our supplier or reference?
- Only with your written permission. No client is named on this website, and case studies are anonymised by sector and scale. The same protection applies to you.
Start a conversation
An outline of the problem, the constraints and the timescale is enough to establish whether this is a fit. I normally reply within one working day and will tell you directly whether I think I can help.