Govern the AI risk you already have.

The reason this is urgent

Most mid-market companies do not have a single, named AI strategy. What they have is dozens of unsupervised AI footprints. A marketing team using ChatGPT to summarize customer interviews. A sales engineer pasting prospect data into Claude. A developer accepting suggestions from an AI coding assistant that was trained on someone else's code. Finance running Copilot over a workbook full of customer financials. Every one of those is a data-handling decision, made by an individual contributor, without a written rule.

At the same time, your customers are sending you AI security questionnaires. Your auditors are asking what governance you have over your AI use. Your contracts now contain AI clauses that someone signed without reading. The risk surface is real, and it is already in the environment.

What we cover

• NIST AI Risk Management Framework (AI RMF) 1.0. Map, Measure, Manage, Govern. Applied to the actual AI use cases in your environment, not a generic template.
• OWASP LLM Top 10. Prompt injection, data leakage, insecure plugin design, model denial of service, foundation-model supply chain risk, and the rest of the current OWASP list.
• Model governance. Inventory of models in use (internal, vendor-embedded, agentic), lifecycle policy, model card requirements, retirement criteria, and ownership.
• Data pipeline review. Where training data comes from, how it is sanitized, how PII is filtered, how prompts and outputs are logged and retained, and how that maps to your privacy obligations.
• AI vendor risk. Reviewing the AI clauses in your SaaS contracts: data residency, training opt-out, breach notification, model versioning, sub-processor disclosure, deletion rights.
• ISO/IEC 42001 readiness. The new AI management system standard. For organizations that need a certifiable governance program, this is the structure that gets you there.
• Acceptable use policy. Practical written guidance for what employees can and cannot put into LLMs, with named exceptions and a path for new use cases.
• AI incident playbook. What to do when a model leaks, hallucinates a damaging output, or fails an audit. Roles, communications, regulatory notification.

NIST AI Risk Management Framework

The AI RMF is the closest thing the United States has to a shared language for AI governance, and it is the right baseline for a mid-market program. It has four functions, and a readiness engagement walks each one against the AI footprint in your environment.

• Govern. Roles, accountability, AI policy, board reporting, training. Who owns AI risk and how is it raised.
• Map. Context, AI use case inventory, impact assessment. What AI exists in the environment and what it touches.
• Measure. Risk identification, metrics, monitoring. What does "good" look like and how do we know.
• Manage. Risk treatment, response, communication. What we do when the measurement says there is a problem.

Deliverable is a written AI RMF profile sized to your environment: governance charter, use case inventory, risk register, measurement plan, and response runbook.

OWASP LLM Top 10

For organizations building with LLMs (chatbots, internal copilots, retrieval-augmented assistants, agentic workflows), the OWASP Top 10 for Large Language Models is the security checklist that should be on every architecture review. We map the application architecture against each risk and write the trade-offs down.

• Prompt injection (direct and indirect)
• Sensitive information disclosure
• Supply chain (model, weights, data, fine-tune source)
• Data and model poisoning
• Improper output handling
• Excessive agency and over-permissioning
• System prompt leakage
• Vector and embedding weaknesses
• Misinformation and hallucination handling
• Unbounded consumption (cost, rate, denial of wallet)

AI vendor risk

The most leveraged AI work in a mid-market environment is not building a model. It is contracting around someone else's. Every renewal in the next twelve months will arrive with AI clauses bolted in. Most are written to favor the vendor.

We read the contracts with you, name the questions to ask before renewal, and draft the redlines that protect your data and your customers' data. The deliverable is a vendor matrix plus a contract appendix template you can reuse.

ISO/IEC 42001 readiness

ISO/IEC 42001 is the first management system standard for AI. If you sell into regulated industries, into governments, or into large enterprise that already has an ISO 27001 program, expect 42001 to land on a procurement questionnaire within the next eighteen months.

A readiness engagement maps your existing governance against the 42001 controls, names the gaps, and authors the documentation needed to be audit-ready. It is structured the same way our SOC 2 and ISO 27001 work is structured: gap analysis, policy authoring, evidence runbook, audit support.

Common engagement models

What this is not

This is not red-team prompt-injection-of-the-week theater, and it is not a research practice. It is governance, risk, and compliance for AI, designed so that when a regulator, an auditor, or an enterprise customer asks how you manage AI risk, you have a written answer with a paper trail behind it.

When the work needs deep ML research expertise (formal model evaluation, mechanistic interpretability, model red-teaming at the weights level), we will tell you, and we will point you to a specialist who does that for a living.

Common questions

Start the conversation

Two questions to think about before the first call. What AI footprints already exist in your environment, and who is being asked about them. Once we have those, the right shape of engagement falls out quickly.