Naming the AI agents you use.

Most consulting websites mention AI exactly once. The sentence reads something like we leverage AI to deliver enhanced outcomes for our clients, or some near-identical variant. The sentence is, in every case, a refusal to disclose what the AI is actually doing.

After two years of clients asking the same questions about which models I use, what they do, and where their data goes, I stopped writing the sentence and started naming the agents instead. Each one with its model provider, its specific job, and an honest disclosure of what it helps with and what it doesn't.

The naming has turned out to be the most defensible competitive choice I've made.

What "we leverage AI" actually says

Decode the language. Leverage means use. Enhanced outcomes means the same outcomes. For our clients is filler. The sentence describes nothing.

The reason it describes nothing is structural — the firm doesn't want to disclose, because disclosure would invite questions the firm doesn't want to answer:

• Which model is actually being used? Often it's the consumer ChatGPT tier, on a personal account, with no enterprise data agreement.
• What specific work is the AI doing? Often it's the parts the firm is charging expert rates for — research, drafting, analysis — which makes the rate harder to defend if the AI is doing the work.
• Is client data being trained on? If the firm is using a consumer tier, the answer is yes by default, and the firm doesn't want that conversation.
• What's the AI getting wrong? Every model has known limitations. Naming the model means owning those limitations. Handwaving means owning none of them.

The vague sentence is a defense, not a marketing choice. It exists because disclosure would be expensive.

What naming actually means

Every AI use case in your business gets a specific named agent. Each named agent has:

• A specific job — "brief articulation," not "writing"
• A specific model — "Claude Sonnet 4.6," not "AI"
• A specific provider — Anthropic, OpenAI, Google — on record
• An honest helps / doesn't help disclosure
• A canonical name your team and clients can refer to consistently

Side effect: this forces you to actually understand your own AI stack at a level most agencies don't bother with. You cannot name an agent until you've decided what it does, which model it should use, and where its limits are. Naming is, structurally, the act of thinking about what you're using. Most firms skip the thinking and skip straight to the marketing.

The five agents

Here is the full disclosure of the atelier as it runs today.

Five agents, five jobs, four specific models from three providers. Each one with its limits stated next to its capabilities. This is the level of disclosure clients should be able to ask for and most cannot.

Why naming creates trust

Without names, you're asking the buyer to trust an abstraction. They cannot verify the abstraction. They cannot research it. They cannot ask informed questions about it. They have to take the seller's word for what the seller is doing — which is the structural definition of not trusting them.

With names, the buyer can:

• Look up the underlying model on the provider's website
• Check the model's known limitations and biases in the public literature
• Audit the provider's data-handling policy
• Have a specific conversation about which agent is touching which piece of work
• Ask informed follow-ups: does Cartographer use the public version of GPT-4o, or the API tier? where is RAG retrieval running? what's in the index?

Trust is not reduced when you disclose more. It is increased, because disclosure proves you've thought about it. The opposite is also true: agencies that won't tell you what models they use are signaling that they haven't thought about it carefully enough to want the conversation.

Why naming is a moat

In the short term, naming is a competitive cost. Other agencies can claim more without disclosure. They can promise "human-only delivery" while actually using AI. They can say "we use the latest models" and not specify. They will win some deals against you. The cost is real.

In the long term, naming is a moat. Buyers who have seen the level of disclosure once cannot un-see it. Once a buyer knows what they should be getting — five named agents, each with its model, each with its limits — they cannot accept handwaving from the next agency. The bar moves once and doesn't move back.

The market for AI consulting is moving toward disclosure as buyers get more sophisticated. The smart buyers are already there. The mass buyers will be there in two or three years, as the press cycle around hidden AI use continues, as compliance frameworks demand disclosure, and as the cost of not knowing becomes visible.

Naming is a small market today. It's a much larger market within a few years.

The agencies that built their pricing on opacity will have a genuinely difficult migration. The ones that started with disclosure will have nothing to migrate.

The honest disclosure framework

For founders or operators who want to do this for their own practice, here's the framework I use. It is not complicated. It is just a habit most agencies haven't built.

For every AI tool that touches client work, state six things:

1. Name it — give it a canonical name your team and clients can refer to.
2. State its model and provider — Claude Sonnet 4.6, Anthropic. Specific version, specific company.
3. State its job — one sentence, the specific thing it does.
4. State what it helps with — the use cases where it's a clear win.
5. State what it doesn't help with — the honest limits. What it gets wrong. Where you still need a human.
6. State the data posture — is client data sent to it? Under what privacy terms? What's retained, what's logged, what's trained on?

Six items. Most agencies don't do any of them. Doing all six is roughly an afternoon of work per agent. The result is a level of disclosure that is, today, rare in the consulting market — which is why it functions as differentiation rather than as a baseline expectation.

A note on naming style

The agents in this practice are named after navigation: stars and instruments used to find one's way. Cassiopeia, the constellation. Atlas, the figure who held up the sky. Cartographer, the maker of maps. Sextant, the instrument for measuring position. Polaris, the north star.

This was deliberate. The agents are tools that help locate things — locate the brief inside an intake message, locate the bug inside a pull request, locate the relevant context about a client's domain, locate the regression in a model's output, locate the available time inside a calendar. Naming them after navigation makes the metaphor clear: they are not characters, not assistants pretending to be human, not "AI teammates" with personalities.

I avoided human-style names — Sarah, the AI assistant — on purpose. Those names anthropomorphize the tool in a way that distorts the relationship. The tool is not your colleague. The tool is an instrument. Naming the instrument as an instrument keeps the relationship honest.

This is a small thing and a real one. Buyers can feel the difference between an agency saying our AI assistant Sarah will be in your meetings and an operator saying Polaris handles scheduling, here's the model and what it does. The first sounds like a coworker who isn't real. The second sounds like a tool that is.

The closing argument

The sentence we leverage AI is a refusal to think.

Naming the agents is what thinking looks like. It is also, separately, what trust looks like. The two are not coincidentally aligned — disclosure is what allows thinking to be verified, and unverified claims of thinking are not trust, they are marketing.

If you are buying AI consulting from anyone — me or anyone else — ask the test question: name the agents you'll use on my work, the models behind them, and where they fail. Most providers cannot answer. The ones that can are the ones who have done the thinking. The ones who have done the thinking are the ones worth hiring. The same architectural commitment shows up in the portal as first principle — infrastructure built before the contract.

— Mikkel, Bangkok