The Blueprint

Artificial intelligence.

Using and building artificial intelligence safely: clear rules for staff, approved tools and suppliers, people checking what matters, AI agents kept within limits, and testing AI systems before they are trusted.

AIM 17 requirements

Foundation

RW-AIM-01Clear rules on what goes into AI

What
Give everyone written rules on what information they may and may not put into AI tools at work.
Why
Most people already use AI tools, often without thinking about where their words end up, so clear rules let them get the benefits without handing confidential or personal information to an outside company.
Protects
Confidentiality

RW-AIM-02Approved AI tools only

What
Keep a list of the AI tools approved for work, and allow staff to use only those.
Why
Free and consumer AI tools may keep what you type, use it to train their models or let their own staff read it, so limiting work to tools you've checked keeps your information in services whose terms and security you know.
Protects
Confidentiality

RW-AIM-03Staff know AI can be wrong

What
Teach staff that AI tools can give answers which sound confident but are wrong, invented or out of date.
Why
People tend to trust a fluent, well-written answer, so knowing that AI can make up facts, sources and figures makes them far more likely to question it before it finds its way into advice, a contract or a decision.
Protects
Integrity

RW-AIM-04A person checks important AI output

What
Make sure a person with the right knowledge checks anything produced by AI before it's used to make an important decision, sent to a customer or published.
Why
AI can be wrong in ways that are hard to spot, so a knowledgeable person checking its work catches errors before they cause harm, and keeps a person, not a tool, accountable for every decision.
Protects
Integrity

RW-AIM-05Faked voices and videos expected

What
Teach staff that voices, videos, photos and messages can now be convincingly faked with AI, so a familiar face or voice on its own doesn't prove who they're dealing with.
Why
Fraudsters use cloned voices and fake video calls to pose as senior people or trusted suppliers, and staff who know this is possible are much more likely to pause and check than to act on an urgent request that seems to come from someone they know.
Protects
Integrity

Established

RW-AIM-06Unapproved AI use found

What
Look regularly for AI tools being used for work that aren't on the approved list, and either approve each one after checking it or stop its use.
Why
People take up new AI tools quickly, and every unapproved one is somewhere your information may be going without anyone having checked it, so finding them brings that use into the open where it can be made safe.
Protects
Confidentiality

RW-AIM-07New AI features checked before use

What
Check each AI feature that a supplier adds to software you already use before allowing it to be switched on.
Why
Suppliers are adding AI assistants to email, documents, meetings and customer systems, often switched on by default, and these can read large amounts of your data or send it on to other companies, so checking first means you know what each feature can see and where your data goes.
Protects
Confidentiality

RW-AIM-08AI suppliers checked first

What
Check how an AI supplier stores, protects and uses your information before you share any of it with them.
Why
An AI service may keep what you send it, store it in another country, pass it to other companies or have weak security of its own, so checking first means you only trust your information to suppliers whose handling you understand and accept.
Protects
Confidentiality

RW-AIM-09Your data kept out of model training

What
Make sure, through contract terms or account settings, that AI suppliers can't use your information to train or improve their models.
Why
Information used for training can become part of a model that other customers use and may later surface in its answers, so ruling this out stops your confidential information, and your customers' data, escaping in a way that can't be undone.
Protects
Confidentiality

RW-AIM-10AI built in only after a risk assessment

What
Assess the risks before building AI into a product, service or business process, and agree how each significant risk will be dealt with before it goes live.
Why
Using AI to make decisions, serve customers or handle data opens up new ways for things to go wrong, from leaked information to unfair or harmful outcomes, and examining those risks before launch is far cheaper than dealing with them after customers have been affected.
Protects
Confidentiality · Integrity · Availability

RW-AIM-11AI agents reach only what they need

What
Give each AI agent access only to the systems and data its task requires.
Why
An agent can be misled by instructions hidden in an email, document or web page it reads, so limiting what it can reach limits the harm it can do if that happens, just as you would limit what a new member of staff can get at.
Protects
Confidentiality · Integrity

RW-AIM-12People approve agents' serious actions

What
Require a person to approve any action by an AI agent that would be costly or hard to undo, such as making a payment, deleting data, changing access or sending information outside the organisation.
Why
However well an agent usually performs, it can misread a task or be manipulated, so a person's approval before serious actions means a single mistake or trick can't turn into a costly incident.
Protects
Integrity

RW-AIM-13Agent actions recorded

What
Record every action an AI agent takes, including what it was asked to do and which systems it used.
Why
A full record lets you see what an agent actually did, notice when it starts behaving strangely and work out what happened if something goes wrong, none of which is possible if its actions leave no trace.
Protects
Integrity

Advanced

RW-AIM-14AI tested against manipulated input

What
Test whether an AI system you have built or deployed can be made to ignore its instructions, or do something it shouldn't, by crafted input such as instructions hidden in the documents, emails or web pages it reads.
Why
AI systems can be talked into misbehaving by text that looks harmless, so testing them the way an attacker would shows up these weaknesses before someone outside finds them.
Protects
Confidentiality · Integrity

RW-AIM-15AI tested for leaking information

What
Test whether an AI system you've built or deployed can be made to reveal information the person using it isn't allowed to see, such as other customers' data or its own hidden instructions.
Why
An AI system that can reach confidential information may hand it to anyone who asks in the right way, so testing for this before release stops it becoming a back door into your data.
Protects
Confidentiality

RW-AIM-16AI tested for harmful replies

What
Test whether an AI system that answers customers or the public can be led into giving harmful, offensive or misleading replies.
Why
An AI system speaking in your name can cause real harm, damage your reputation and lead to legal claims if it can be steered into saying the wrong thing, so finding these weaknesses before release lets you fix them first.
Protects
Integrity

RW-AIM-17AI models from known sources

What
Use AI models, and the data used to train or adjust them, only from sources you have checked, recording where each one came from.
Why
A model or dataset from an unknown source can carry hidden behaviour, harmful code or deliberately poisoned data, so checking where it came from keeps tampered components out, and the record shows you quickly which products are affected if a problem is announced.
Protects
Integrity