Approved tools and the vendor checklist

The guardrails go in before day 0, before anyone is asked to change how they work. Their job is simple to say and hard to do: make the safe way the easy way, so the controls live in the tools rather than in a policy nobody reads. There are six of them in this module. The first is a list.

A policy nobody reads isn't a guardrail

Most organisations already have something written about AI. It usually says what people mustn't do, sits on the intranet, and changes very little. People don't go around the rules because they're reckless. They go around them because the approved route is slower, vaguer or missing, and the personal account on their phone works in thirty seconds.

The cost of that gap is now measured.

A guardrail is different from a policy. It's built into the way work already happens, so following it takes less effort than going around it. The approved tools list comes first because every other guardrail hangs off it: the data map says which tools each kind of data can go into, and the spending limits set a cap on each tool. Without the list, there's nothing to attach them to.

One table everyone can see

The list is a single table, published where people already work. It has five columns.

• Tool. The name people actually use for it.
• Plan. The same product can be safe on a business plan and unsafe on a personal one. The plan decides whether your data trains the vendor's models, how long it's kept and whether you control who has access.
• Approved for. What people may use it for, in words they'd use themselves.
• Not approved for. Just as important. “Not for regulated data” or “not for anything sent to a client without a person checking it” saves a hundred questions.
• Who to ask. A named person, so a question takes one message, not a meeting. Usually the person who runs the loop.

Publish it on the intranet, pin it in the team chat, and put it in onboarding. If people have to hunt for the list, they'll use whatever is nearest.

Include the AI you didn't buy

A growing share of the AI in your organisation arrived without anyone choosing it. Vendors switch AI features on inside software you already pay for: the email client drafts replies, the CRM summarises accounts, the video-call tool takes notes and sends them round. In 2024, Gartner forecast that by 2026 more than 80% of independent software vendors would have generative AI built into their enterprise applications, up from less than 1% in 2023. Each of those features handles your data in a new way, often under terms that changed in an update nobody read.

These belong on the list too, with a plan, an “approved for” and a person to ask, just like a tool you bought on purpose.

A sandbox, so trying something isn't going around the list

People will always find new tools before the list does. That's a good thing: your keenest users are scouting for you. The question is where they try them. A sandbox is a pre-approved place to experiment with public or made-up data, kept apart from live systems. It turns curiosity into something you can see, and it gives the front door a steady supply of tools worth checking properly.

Eight questions before anything goes on the list

Every new tool answers the same eight questions before it's approved. They come from the pilot kit's vendor checklist, and they fall into four groups.

• Your data (questions 1 to 3). Does the vendor train on it, can you switch that off in writing, where is it kept and for how long, and will they sign the agreements the law requires?
• Access and control (4 and 5). Can each person's access be removed the day they leave, and can your admins control the AI features, including new ones the vendor switches on later?
• What's underneath (6 and 7). Which other companies and models sit behind it, how you get your data out, and what happens when the model changes.
• Money (8). What it costs at full seat count, and whether there's a hard cap you can set. That's what makes the spending limits in lesson 4 enforceable.

Build the list for your own organisation below. Start with the tools people already use, then add the AI features your software has switched on, then the sandbox.

Not banned in secret

The list has one more rule, and it matters as much as the table: tools that aren't on the list aren't banned in secret. Raise them at the front door. Anyone can put a tool forward. The person who runs the loop gives a first answer within 10 working days, and the tool either passes the checklist and joins the list, waits in the sandbox, or gets a clear reason why not.

This is the promise from module 1 doing its work. If the only response to an unlisted tool is a ban, people stop mentioning them, and you lose sight of exactly the use you most need to see. Module 3 brings everything already running into the open, and this list is where most of it lands.