Many AI policies apply to tools: which are approved, and what can go into each. Putting a use off limits applies to the use itself, so it covers any tool and anyone, including personal accounts nobody approved. Larger clients and insurers may ask which uses the organisation has put off limits. The positions below count how many of six common limits are set. What each limit costs depends on the organisation's work: a limit on regulated advice costs little where no such advice is given.
Nothing extraLink to this section
People judge each use for themselves, within the law and existing policies on privacy and confidentiality. No value is ruled out in advance, but two people can reach different answers on the same task, both in good faith, and limits specific to AI tend to be added after an incident.
Easy to miss. Existing policies were usually written before these tools, and rarely say whether a particular contract or spreadsheet can go into one.
When to choose it. It makes sense where little sensitive information is stored, or the risk appetite already sets the limit, as Safe first does for client and personal information. Support it with plain examples of how existing policies apply to AI.
One or two thingsLink to this section
One or two limits that people can remember without looking them up and explain to a client. Most uses stay allowed, and the limits cost only the work they affect.
Easy to miss. People tend to read a short list as complete, and treat anything missing as approved.
When to choose it. Choose it where the real exposure is in one or two places. Word each limit precisely: "anything a client has shared, or anything that identifies them" answers more cases than "client information". Say alongside it that judgement still covers everything else.
Several thingsLink to this section
Three or four limits. People know what is allowed on most sensitive work, but will need to check the list. Exposure is low in the areas that concern clients, insurers and regulators, more value is given up, and each limit needs someone to explain and check it.
Easy to miss. Several limits ask a lot of the teams whose work they cover, often HR and client services, who watch colleagues use AI freely while their own work stays manual.
When to choose it. It is best where there are several distinct kinds of sensitive work. Ask the teams each limit affects what it should cover, and help them find the uses that remain.
Most thingsLink to this section
Five or six limits. Much of the work the limits cover stays manual or waits for a person's approval, and people use AI freely only where no limit applies. Exposure is lowest in the areas listed, and the most value is given up: the teams most affected may have little to use AI for while their licences cost the same, and exceptions need someone's time.
Easy to miss. Removing a limit means someone senior accepting the risk it prevented, and few volunteer, so lists tend to stay as written.
When to choose it. It suits organisations whose clients, insurers or regulators expect tight control. Give each limit a review date and a named person who decides whether it stays, and check what a team can still do before buying it licences.
How to chooseLink to this section
Start with what each of the six common limits costs.
- Client and customer information. AI can still structure and draft client work, with the details added by hand, but summarising a client's own documents stays manual.
- Personal information about staff. HR work about individuals stays manual, and staff can be told plainly that their details stay out of AI tools.
- Decisions about people. Nobody has to defend an AI's part in a decision about someone's job. Hiring and reviews run as today, which can mean switching off or replacing recruitment software that already uses AI.
- Advice in regulated areas. Where advice is the main work, this rules out much of what AI could do there.
- Publishing without review. A person, usually the sender, reads AI-written work before it goes out, ruling out automated replies such as a website chatbot.
- System changes without approval. A person approves each change AI suggests, ruling out tools that update records overnight on their own.
The last two require a person's approval rather than keeping AI away, so they cost more as tools that act on their own come into use.
Then answer these together:
- Which types of information does the organisation store, and which would do most harm in the wrong tool?
- Should each limit apply in every tool, or only outside approved ones? The second kind belongs in the policies for each tool.
- What do larger clients, insurers and regulators expect, and do client contracts already mention AI?
- How much will teams decide for themselves? Where teams decide for themselves, this list is the one policy they all share, so word it for them.
Tensions with other decisionsLink to this section
- A high risk appetite with a long list. Open and Willing to take risks rely on a few clear limits and judgement for the rest. Five or six limits send the opposite message, and people tend to follow the written one. Shorten the list, or bring the appetite down.
- Growth as an aim, with client information off limits. Growth means using AI in work clients see, and this limit keeps work on a client's own material manual. Check that help with structure and drafting is enough for the growth planned, or allow client information in approved tools with clear policies.
- Personal judgement with little time to learn. From Balanced upwards, unless most uses are off limits, each person decides much of what goes into a tool. With little working time for AI learning, that judgement depends on whatever they have learnt. Add regular time to practise the policies on real work, or set clearer limits.
When to revisit itLink to this section
Set a review date, and revisit sooner when:
- an incident or near miss shows a limit is missing or in the wrong place
- the same exception keeps being requested
- the organisation takes on new kinds of work or information
- a tool already in use adds AI
- a client, insurer or regulator asks something the list cannot answer
- the risk appetite, or who decides, changes
If the leadership team would like support with this decision, see Getting outside help.
Something missing or out of date here? Tell us.