Stop Prompting. Start Writing Your AI Agent a Dictionary.

Your AI agent is fluent in English. It has never once heard what your team means by a callback.

The Word That Broke The Robot

It is 4:50 p.m. In Manchester. Priya runs a five-person appliance repair business, three vans, one dispatcher, one very tired inbox. Two months ago she wired up an AI agent to answer the phone and triage jobs before a human ever picks up.

A customer calls. "This is a callback on the fridge from Tuesday." The agent hears an ordinary English sentence and books a routine visit two weeks out.

In Priya's trade, a callback is not someone calling back. It is a repeat visit for a fault that should have been fixed the first time, and it is supposed to jump the queue, free, same week, no exceptions. The agent did not know that. It was never told. It booked the job like any other job, the customer waited two weeks for a fridge that was never actually fixed, and Priya spent Saturday morning on the phone apologising.

The agent did not fail because it was a bad agent. It failed because nobody had ever written down what "callback" means at Priya's shop.

Your AI Agent Isn't Broken. It's Uninformed.

Here is how most small businesses deploy AI right now. They turn on a chatbot, or Claude, or ChatGPT, or an agent built on Claude Code or Codex, point it at a task, and expect it to already understand the business.

It doesn't. It can't. A general-purpose AI model has read the entire English-speaking internet and none of your invoices.

It doesn't know that "COD" at your shop means cash collected on delivery, not a video game. It doesn't know a "no-cool" ticket means the air conditioning unit is running but not cooling, which is completely different from "no-power," which means the unit will not turn on at all, which changes which technician gets dispatched and how urgently. It doesn't know that when your team says "will-call," they mean a part sitting on a shelf waiting for a customer to collect it, not a phone call you owe someone.

Every one of those words is obvious to your team. None of them are obvious to the agent. That gap has a name now, among the engineers who build agents for a living: domain language. And on 1 August 2026, one post about it went further, faster, than almost anything else in the hands-on AI-automation world that week.

What Is A Business Glossary For An AI Agent?

A business glossary for an AI agent is a short, plain-language document, usually a single page, that defines the specific words, ticket types, statuses, and shorthand your business uses, so that any AI tool you plug in reads that page before it reads anything else.

Think of it as the induction pack you would hand a sharp new hire on day one. Not the whole employee handbook, just the words: what a callback is, what COD means, what "Tier 2" means in your support queue, what "VIP" means to your team versus what it might mean to a generic script.

Software engineers who build with Claude Code, Codex, and similar coding agents have been quietly solving this exact problem for a year, because they hit it constantly. Developer and YouTuber Matt Pocock, who teaches AI-assisted engineering to roughly 60,000 subscribers through his newsletter, put it in six words on X on 1 August: "Domain language is unbelievably powerful with agents." He was quoting a post from fellow developer-educator Alex Sidorenko, captioned simply "Find the right terms." Pocock's post pulled in 1,035 likes, 678 bookmarks, and more than 78,000 impressions in under a day, one of the highest-engagement posts in the whole agent-building niche that week.

What both of them were pointing at is not a coding trick. It is a business trick that happens to have been discovered first by coders, because coders were the first people paying an AI agent to do real, expensive, judgment-heavy work all day, every day.

Google's own AI Overview, when you search for how to teach an AI agent your business's terminology, now recommends almost exactly the same shape of fix: give the agent a custom glossary of your jargon, feed it in a short document rather than a live chat, and keep it separate from generic instructions. When the engineers who build agents for a living and the search engine that indexes their advice agree on the same fix, it is worth five minutes of your Saturday.

Isn't This Just A Longer Prompt?

No. A prompt is a conversation. It vanishes the moment the chat window closes. A glossary is a document. It sits on a shelf, gets read every time, and never has to be re-explained.

Pocock's own videos on this describe a rule engineers now use for deciding how much of this groundwork to do upfront: shift the checkpoint left for big, foundational work, and shift it right for small, disposable work. Building your business's core glossary is big, foundational work. You do it once, carefully, and it pays out on every automation you build afterwards, your phone agent, your support inbox, your quoting tool, your reporting dashboard, all reading from the same page.

A one-off prompt is a checkpoint you pay for every single time. A glossary is a checkpoint you pay for once.

We have written before about why your AI agent doesn't need more data, it needs context. A glossary is the sharpest, cheapest form of context there is: not your whole company history, just the twenty words that keep tripping the agent up.

Build Your One-Page Business Glossary

You do not need a developer for this. You need forty-five minutes and the same instinct Priya's most experienced technician already has in his head.

1. List the words that would confuse a total outsider. Sit with your longest-serving staff member and ask: "What do we say that a brand-new hire always gets wrong in week one?" Write down every term. For a repair shop: callback, no-cool, no-power, will-call, COD, warranty void. For a salon: no-show fee, patch test, colour correction. For an agency: scope creep, retainer, kill fee.
2. Write a one-line, plain-English definition for each. Not a legal definition. The definition you would give a new hire standing next to you. "Callback: a repeat visit for a fault we already tried to fix. Free. Jump the queue. Book within 48 hours."
3. Add the consequence, not just the meaning. This is the step almost everyone skips, and it is the one that actually changes agent behaviour. Don't just say what a word means, say what should happen because of it: which queue, which price, which urgency, which human gets looped in.
4. Save it as one plain document. A single page. No design, no branding, just term, meaning, consequence. Engineers call theirs a CONTEXT.md file; you can call yours whatever you like, a Word doc, a Google Doc, a text file titled "How We Talk."
5. Paste it into every AI tool you use, every time. Whether that's a custom instructions field in ChatGPT or Claude, a system prompt on your phone-answering agent, or the first message of any automation you build, this glossary goes in first, before the task.
6. Update it the moment an agent gets a word wrong. That failure is not a dead end, it is a missing glossary entry. Add it, and no agent you ever deploy will make that exact mistake again.

What You Have After This

A week from now, you have a one-page document that took less time to write than one bad phone call took to apologise for.

A month from now, every AI tool you plug into your business, your phone agent, your quoting tool, your support inbox, whether it is built on Claude, ChatGPT, or something you have not adopted yet, reads from the same page and makes the same judgment calls your best technician makes without thinking.

Six months from now, you are not writing instructions for AI agents one at a time anymore. You are onboarding them, the same way you onboard a person, with the same document, in minutes instead of months.

Same team. Same tools. A completely different conversation with every piece of software you own.