Articles

Our own AI inventory: seven tools, and what each one can touch

2 November 2026

AI-assisted text, reviewed by a person at Heedward.

We did to ourselves what customers' AI questionnaires ask of their suppliers: list every AI tool we use, what each one can touch, and who answers for it.

An AI inventory is a plain list: every AI tool a company uses for its work, what it does, what it can reach, and who answers for it. It is what a customer means when its questionnaire asks which AI tools touch our data? We built ours over two days in October, the way we would for anyone: start from the records that already exist, then ask the people.

What this is not: a security review of the tools, or a verdict on their vendors.

Why should you care? Customer questionnaires now ask suppliers two things about AI: which tools you use, and whether your data, or theirs, is used to train them. If you sell software or services, those questions are coming to you. They are easier to answer from a list you keep than from memory.

Part of the list was easy. Our coding agent signs most of the changes it helps us make, so that record already existed. The hard part was one column: what can each tool touch?

Seven tools, in five kinds

1. The agents that write. Three of the seven rows are the same coding agent, in three places: on our laptops, in the cloud when started from a phone, and as research sub-agents. Most of our written work passes through it. The question: one tool or three? We counted three, because each one can touch different things.
2. The one that acts as us. One agent helps with outreach: it researches companies we might work with and prepares our first contact. It works through a founder's own accounts, so to the other side it is us. It sends only what a founder has approved, and a person approves every reply. The question: what may it do without asking?
3. The one that stays home. Our meeting note-taker transcribes and summarises on a founder's own computer. The audio never leaves it. The question: which step does leave the machine? Here, one does: the summary then goes to the coding agent to become our meeting notes.
4. The one-off. Our logos came from a general chat assistant. No record of ours mentions it. The question: what else did we use once and forget?
5. The reviewers. AI models stress-tested our strategy. Their reviews are labelled as AI output, not expert opinion.

The column everyone fills in too narrowly

"What does it touch?" is usually answered with what we ask the tool to do. That is the narrow answer. An agent asked to edit one file runs on a laptop where it can read every folder that laptop can, including the shared drive. So we wrote down the widest access, not the usual task. The rows got longer, and more honest.

Under the AI Act, using these tools for our work makes us a deployer of each of them. The list is where our own obligations start.

The surprise

Five of our seven tools run on accounts with two vendors, Anthropic and OpenAI, and each founder has their own. Whether a vendor may train on what we type is a setting in each of those accounts, not a firm-wide switch. So the most common questionnaire question, is our data used for training?, depends on each founder's settings.

So we checked. In one founder's accounts, both were on: the coding agent's and the chat assistant's. Switching them off took a minute. It changes what happens from now on; it does not take back what was already shared. The other founder's accounts are next.

We had not looked before. Without the list, we would not have.

Three questions for each tool

Run these on every AI tool your firm uses:

1. What does it do for us?
2. What is the widest thing it can touch?
3. Who answers for it, and what does a person check before its output leaves us?

Our note-taker row answers all three: meeting notes; the audio of our calls, on one computer; one founder answers for it, and names and personal remarks are taken out before the notes go into our shared records.

The same answer, rewritten

Before: "We use AI tools responsibly and in line with our policies."

After: "Meeting notes: an open-source note-taker transcribes and summarises our calls on the computer of the founder who runs it. The audio stays there. The summary then goes to our coding agent to become the notes, and names and personal remarks are removed before the notes go into our shared records."

The second version is longer. Every part of it can be checked.

What shows it

The commit line is a record, not a promise. Each one names the agent that helped and sits next to the change itself and the founder who made it. Nobody wrote a report about it; the system kept it. It has limits too. It shows which changes an agent helped write, not what we pasted into a chat. And a commit without the line does not prove no AI was used. Records like this cover part of an inventory, not all of it. The training settings, for example, live in two vendors' account pages, and no log of ours shows them.

So, what does your list say?

If a customer asked tomorrow which AI tools touch their data, could you send them a table? Ours took two days to build, and it still has gaps. That is fine. A table with honest gaps is easier to trust than a sentence with none.

How we counted

On 24 and 25 October we listed the AI tools used for Heedward's work, starting from the git history and notes in our three repositories, then asked both founders about anything the records did not say. A tool counts when it reads or writes our material, not when one of us tried it once; personal use outside the firm's work is left out. A change counts as AI-assisted when its record carries a co-author line naming an AI agent, or the agent made it itself.