What makes software AI-ready? Eight things to check

Software is AI-ready when an AI assistant can use it safely on your behalf. That means it can find things, do things, and be stopped from doing the wrong things. A chatbot in the corner of the screen doesn't count.

Here's what to check. I've put them in the order we met them when we rebuilt our own product.

1. Can everything be done without the screen?

If an action only exists as a button, an AI can't do it. Every action needs a door at the back as well: a way to do the same thing by instruction.

Ask the supplier what share of the product can be driven that way. If the answer is that you can look up contacts, it's a demo.

2. Is each action described in plain words?

An AI chooses what to do by reading a description of each action. Vague descriptions give you an AI that picks the wrong one.

We ended up writing them the way I'd brief a new starter. What this does. When to use it. When not to.

3. Does it use the person's own permissions?

The AI should act as whoever is logged in and see only what they can see. No master key.

4. Does it ask before anything risky?

Send, spend, approve, delete. Each should stop, show what is about to happen, and wait for a yes.

5. Are there limits?

A cap on messages and spend per hour means a mistake stays small.

6. Is there a log?

Every action recorded: what, when, for whom. If you can't look it up, you can't trust it.

7. Does it tell the truth when something fails?

This one surprised me. An AI is only as good as the errors it gets back. If the software just says something went wrong, the AI guesses, and you get a confident wrong answer.

If it says the message was not sent because you've hit the hourly limit, the AI can tell you that and you can act on it.

8. Is it safe to try twice?

Connections drop. An AI will sometimes repeat a step. Sending the same invoice twice because of a retry is not acceptable, so the software has to recognise a repeat and ignore it.

We'd already solved this in our mobile apps, for bad phone signal. Same problem, new reason.

How do you keep it AI-ready?

This is the part people forget. A product changes every week. New buttons get added. If nobody adds the matching action for the AI, the product drifts back to being click-only.

We have an automatic check for it on Plan@Job. If someone adds a feature without the action an AI can use, or a written reason why it shouldn't have one, the check fails.

That one rule is why the whole product can be run from a chat, and not just the easy tenth of it.

How do you test a product in ten minutes?

1. 01Connect it to Claude or ChatGPT. If you can't, you have your answer.
2. 02Ask it to find something specific. A customer, a job, an unpaid invoice.
3. 03Ask it to change something small.
4. 04Ask it to send something. See whether it stops and asks you first.
5. 05Ask it for something you shouldn't have access to. It should refuse.

Five questions, and you'll know more than the sales call told you.

If you're having software built, put these eight in the brief from day one. Adding them later costs a lot more. It's how we build, and the reasons are in the future is not software. Tell me what you're planning.