What to Do Before Building an AI Agent for Your Business
Direct answer
Before you build an AI agent for your business, build the company brain: private retrieval with citations across your real records. Here is why it works.
- Before you build an AI agent for your business, build the company brain: private retrieval with citations across your real records. Here is why it works.
- The strongest AI work starts with one operational bottleneck, one owner, and one result the team can inspect.
- Use the article as the diagnosis layer, then move into a scoped build, proof path, or commercial workflow page.
If you are about to build an AI agent for your business, the most expensive mistake has nothing to do with the model.
I run an AI consultancy, and companies usually call us after a pilot has died. The post mortem is nearly always the same. It usually started with someone asking how to train ChatGPT on their company data. They picked a capable model, wired it to one tool, gave it a job title, and waited for magic. The agent then answered questions like a smart stranger sitting in reception. It knew everything about the world and nothing about the business. It could not say which customers were angry last quarter, what the refund policy is in practice, or why the Tuesday delivery always slips. That knowledge existed, but it lived in call recordings, in seven thousand old tickets, in a shared drive nobody opens, and in the head of one operations manager who was on leave. Real businesses are messy like this. Every one of them. The pilots that survive deal with the mess before they deal with the agent.
What actually kills an agent pilot?
Most agent pilots fail because the agent has no context about the organization, not because the model is weak. The vendor demo worked because the demo company is clean. Tidy data, one system, no history. Your company has nine years of history spread across tools that do not talk to each other. When an agent without that context starts answering customers or drafting reports, it gets the safe things right and the important things wrong, and your team learns to distrust it inside a week. Distrust is fatal. Once a manager has caught the agent inventing a policy, they start checking every output by hand, which costs more time than the agent saves. The failure gets blamed on the technology. The actual cause was sending a new worker onto the floor with no induction, no handbook, and no key to the filing cabinet.
What is a company brain?
People ask whether they can train ChatGPT on their business data. Almost nothing gets trained. What you build is a company brain, a private retrieval layer across everything your business has already recorded. Vendors would call it an AI knowledge base. Call transcripts, support tickets, documents, contracts, meeting notes, customer records. You ask it questions in plain language and it answers in plain language, and the part that matters most is that every answer carries a citation pointing back to the exact source it came from. Not a confident summary floating free of evidence. A claim with a receipt. Building one is not glamorous work. It means untangling permissions, deciding what counts as the truth when two systems disagree, and accepting that some knowledge only exists in someone's head and needs to be written down for the first time. That is precisely why it is valuable. The mess is the moat. Anyone can rent the same model you can. Nobody can rent your decade of operational history.
What does this look like in production?
We run one of these at TwoHundred for a client right now. The chief executive asks the business direct questions and gets sourced answers in seconds. What did we promise this customer in March. Which suppliers slipped this quarter. Before, those questions went down a chain of managers and came back days later as a summary shaped by whoever wrote it. Now the answer arrives with citations, and anyone can click through to the original call or document and check it. The second effect is the one I find more interesting. Every agent we deploy afterwards sits inside that same brain. The agent drafting customer replies reads the same history a fifteen year employee would. It stops behaving like a stranger in reception. It behaves like someone who has worked there for years, because in the only sense that matters, it has.
Why do citations matter so much?
Citations are the difference between an answer you trust and an answer you have to verify. People forgive a colleague who says they are not sure, and they forgive a system that shows its sources. What they never forgive is confident invention. An agent that cites the ticket, the call, the clause in the contract, earns wider scope over time because checking it costs thirty seconds and a click. That cheap verification is what lets a cautious team hand over bigger jobs. First it drafts replies, then it handles refunds under a threshold, then it briefs the board pack, and at each step somebody clicked the source a few times, found it solid, and relaxed. An agent that cannot show its working never earns that ladder. It gets quietly demoted to drafting birthday messages. I have watched both happen inside the same company in the same quarter.
What happens to the management layer?
Here is the consequence almost nobody prices in. A large share of middle management work is moving information around. Collecting status from below, translating context for the people above, and packaging it all into something a meeting can digest. It is honest work, and it is also exactly what an agent sitting inside a company brain does in seconds, with sources attached. When that routing gets automated, spans of control widen. A leader who could manage six people can suddenly see and direct twenty, because the brain answers the questions that used to require a meeting. I do not think this means firing managers. The good ones move up the stack, from routing information to making judgment calls on it. But the org chart that assumes every eight people need a full time human information relay is going to look dated within a few years, and the companies that notice first will run leaner than their competitors.
Where would I start?
Start with where your knowledge actually lives, not with the agent. For most businesses that means calls, tickets, documents, and customer records, in that order of neglect. Get them into one retrieval layer, insist on citations from day one, and put it in front of one impatient executive with real questions. Not test questions. The questions they would otherwise fire at a department head on a Sunday night. You will find gaps immediately. Some answers will be wrong because the underlying record is wrong, and fixing those records is half the value of the exercise. Only then attach agents to it. This is unfashionable advice in a market selling autonomous everything, and it is the difference I keep seeing between the pilots that die in week three and the systems still running a year later, doing more than anyone planned for them.
Agents are easy to buy. Context is not. Build the brain first.
Related reading
Related implementation paths
AI implementation services
Turn the article into a scoped first system with clear ownership, data, and measurement.
AI workflow automation
Automate one operational workflow inside the tools the team already uses.
AI agent development company
Design agents around jobs, tools, approval points, and measurable business outcomes.
Questions this article answers
What actually kills an agent pilot?
Most agent pilots fail because the agent has no context about the organization, not because the model is weak. The vendor demo worked because the demo company is clean. Tidy data, one system, no history. Your company has nine years of history spread across tools that do not talk to each other. When an agent without that context starts answering customers or drafting reports, it gets the safe things right and the important things wrong, and your team learns to distrust it inside a week. Distrust is fatal. Once a manager has caught the agent inventing a policy, they start checking every output by hand, which costs more time than the agent saves. The failure gets blamed on the technology. The actual cause was sending a new worker onto the floor with no induction, no handbook, and no key to the filing cabinet.
What is a company brain?
People ask whether they can train ChatGPT on their business data. Almost nothing gets trained. What you build is a company brain, a private retrieval layer across everything your business has already recorded. Vendors would call it an AI knowledge base. Call transcripts, support tickets, documents, contracts, meeting notes, customer records. You ask it questions in plain language and it answers in plain language, and the part that matters most is that every answer carries a citation pointing back to the exact source it came from. Not a confident summary floating free of evidence. A claim with a receipt. Building one is not glamorous work. It means untangling permissions, deciding what counts as the truth when two systems disagree, and accepting that some knowledge only exists in someone's head and needs to be written down for the first time. That is precisely why it is valuable. The mess is the moat. Anyone can rent the same model you can. Nobody can rent your decade of operational history.
What does this look like in production?
We run one of these at TwoHundred for a client right now. The chief executive asks the business direct questions and gets sourced answers in seconds. What did we promise this customer in March. Which suppliers slipped this quarter. Before, those questions went down a chain of managers and came back days later as a summary shaped by whoever wrote it. Now the answer arrives with citations, and anyone can click through to the original call or document and check it. The second effect is the one I find more interesting. Every agent we deploy afterwards sits inside that same brain. The agent drafting customer replies reads the same history a fifteen year employee would. It stops behaving like a stranger in reception. It behaves like someone who has worked there for years, because in the only sense that matters, it has.
Why do citations matter so much?
Citations are the difference between an answer you trust and an answer you have to verify. People forgive a colleague who says they are not sure, and they forgive a system that shows its sources. What they never forgive is confident invention. An agent that cites the ticket, the call, the clause in the contract, earns wider scope over time because checking it costs thirty seconds and a click. That cheap verification is what lets a cautious team hand over bigger jobs. First it drafts replies, then it handles refunds under a threshold, then it briefs the board pack, and at each step somebody clicked the source a few times, found it solid, and relaxed. An agent that cannot show its working never earns that ladder. It gets quietly demoted to drafting birthday messages. I have watched both happen inside the same company in the same quarter.
What happens to the management layer?
Here is the consequence almost nobody prices in. A large share of middle management work is moving information around. Collecting status from below, translating context for the people above, and packaging it all into something a meeting can digest. It is honest work, and it is also exactly what an agent sitting inside a company brain does in seconds, with sources attached. When that routing gets automated, spans of control widen. A leader who could manage six people can suddenly see and direct twenty, because the brain answers the questions that used to require a meeting. I do not think this means firing managers. The good ones move up the stack, from routing information to making judgment calls on it. But the org chart that assumes every eight people need a full time human information relay is going to look dated within a few years, and the companies that notice first will run leaner than their competitors.
Where would I start?
Start with where your knowledge actually lives, not with the agent. For most businesses that means calls, tickets, documents, and customer records, in that order of neglect. Get them into one retrieval layer, insist on citations from day one, and put it in front of one impatient executive with real questions. Not test questions. The questions they would otherwise fire at a department head on a Sunday night. You will find gaps immediately. Some answers will be wrong because the underlying record is wrong, and fixing those records is half the value of the exercise. Only then attach agents to it. This is unfashionable advice in a market selling autonomous everything, and it is the difference I keep seeing between the pilots that die in week three and the systems still running a year later, doing more than anyone planned for them. Agents are easy to buy. Context is not. Build the brain first.