Step 1: Scope Your Use

Before assessing risks or choosing tools, describe precisely what you want to do and what you would share with the AI tool. Being specific here makes the rest of the assessment straightforward; a vague scope cannot be assessed.

This step has three parts: define the use, assess the level of autonomy, and understand what you are sharing.

Define what you want to do

Answer these three questions:

  1. What activity will AI assist with? Be specific. "Using AI for coding" is too broad. "Using an AI coding assistant to generate unit tests for the payments service" is better.

  2. What data will be involved? What will you send to or share with the AI tool? Source code, user research transcripts, database schemas, support tickets, design briefs?

  3. Who is affected? Just you and your team? End users of the service? Members of the public whose data is being processed?

Categorise your use

The categories below describe tasks people do while delivering a service, not the services themselves. There is one exception, product features, which covers AI you build into the service.

Most AI use falls into one of eleven categories. Some uses span more than one. If so, assess against each and apply the more restrictive mitigations.

These categories are a thinking aid rather than an exhaustive list. They exist to get you to the right risks quickly by pointing at the closest well-understood shape of work. If your use does not fit neatly, pick the nearest one or two, note why the fit is imperfect, and carry on. What matters is that you assess the right risks, not that you chose the right label. If you are doing something the categories genuinely do not describe, work through Step 3 on its merits and feed the gap back so the framework can be updated.

Category The key characteristic
Software development AI generates or modifies code that may end up in the product
Code analysis Large volumes of existing code are sent to the AI; the output is analytical, not production code
Synthetic data generation AI produces artificial data that will be handled as though it were safe
Product feature AI directly interacts with or affects end users of the service
User-facing support AI processes support data to help the team respond to requests from people
Live service operations AI works against production systems: alerts, incidents, logs, infrastructure
User research AI processes what real people told you, to represent them to the team
Design AI shapes how the service works and how people move through it
Content AI produces text that reaches the public in the department's name
Business analysis AI produces requirements or decision artefacts that others treat as settled reasoning
General productivity AI assists with routine work tasks that do not shape a decision

For the full description of each category, covering the tools typically used, what is shared, and the typical risk fingerprint, see Reference: Use-Type Profiles. (That reference also explains the underlying types of AI: GenAI, LLMs, ML, and NLP.)

Assess the level of autonomy

Two teams can use the same tool on the same data in the same category and face very different risks, because of how much the AI is allowed to do on its own. Autonomy cuts across every category, so record it separately.

Pick the level that matches what the AI is actually permitted to do:

Level What it means Example
Suggests AI proposes; a person decides and does the work themselves Inline code completion; a list of possible root causes
Drafts for review AI produces a complete artefact that a person reviews and edits before it takes effect A generated pull request; a drafted support response awaiting agent approval
Acts with approval AI performs actions in a real system, but each action needs explicit human approval first An agent that runs commands or edits files with per-action confirmation
Acts autonomously AI performs actions in a real system with no human in the loop for each action Automated ticket closure; auto-remediation of an alert

Two things to be careful about. First, the level is what the AI is permitted to do, not what you intend to let it do. If the approval prompt can be turned off, or a "yes to all" option exists and gets used under pressure, assess at the higher level. Second, autonomy tends to creep: a tool introduced at drafts for review acquires an auto-apply setting, or a team that reviewed every suggestion in week one stops by week six. Note the level you assessed, and reassess if the way the tool is used changes.

You will use this in Step 3, where higher autonomy raises specific risks and adds extra mitigations, and in Step 4, where fully autonomous action requires SRO approval regardless of the overall risk level.

Understand what you are sharing

What you share with the AI tool is often the most important factor in determining what mitigations are needed. Most uses involve code, data, or both. Assess each, because they have different risk profiles.

If you are sharing code

🛑 Secrets must never be shared with any AI tool. API keys, credentials, tokens, database connection strings, and service endpoints must be removed before any code is shared. Run a secrets scanning tool before sharing code with any AI service. If you find secrets, remove them and re-evaluate what you should share.

Beyond secrets, consider:

If you are sharing data

What is the data classification? This is a hard gate. Under the Government Security Classifications Policy, government information carries one of three classifications: OFFICIAL, SECRET, or TOP SECRET. ‑SENSITIVE is not a fourth tier. It is a handling caveat applied to the minority of OFFICIAL information that needs extra care. It still changes what you must do.

Classification AI implications
OFFICIAL Tools on your project's register are generally appropriate, subject to the risk assessment in Step 3.
OFFICIAL with a ‑SENSITIVE marking Only tools cleared for it may be used. Check the classification recorded in the tool's profile (Step 2). Do not make that call yourself: if the tool is cleared only to OFFICIAL, escalate to the SRO. Client approval is likely needed. The AI Playbook is explicit that public or unassured generative AI must not process OFFICIAL information carrying additional markings.
SECRET / TOP SECRET 🛑 Stop. See below.

If your data is SECRET or above

🛑 Do not use external AI services. This is a hard stop, not a high risk rating: do not continue to Step 2.

The Government Security Classifications Policy requires SECRET and TOP SECRET information to be handled on dedicated, accredited systems, by security-cleared people, on a strict need-to-know basis. No commercially available AI service meets that bar, so this assessment cannot end in "proceed".

AI use at these tiers is not impossible, but it has to happen inside an accredited environment. Take specialist security advice, and do not treat anything here as authority to proceed. It may also be possible to use local, non-internet-connected LLMs on this data, but only with specialist security advice.

Other factors beyond classification

Classification is not the only factor. OFFICIAL data can still carry significant risk. Also ask:

When both code and data are involved

Many uses involve both, for example code analysis on a codebase that contains sample records, or a product feature whose defining prompts are shared alongside user data at runtime. Assess each independently and apply the more restrictive set of mitigations. See Reference: Use-Type Profiles for how code and data typically combine in each use type.

Record your scope. Note what the AI will be used for, its category, its autonomy level, what code and data will be shared, the data classification, whether PII is involved, and any consent or contractual constraints. You will use all of this in Step 2 and Step 3.


Next: Step 2, Check the Tool Is Eligible >