AI in your business: don't rent a chatbot, build it on your own account
I run an AI assistant at home and I'd never let that version loose in a business. (There's a separate post on why: OpenClaw at home.) The obvious follow-up question is: so what does the business version look like?
This is that post. It's written for the owner of a trades or service business with twenty to two hundred staff, because that's who I build for. But the shape is the same for anyone.
Access is not the same as a system
Most businesses are "doing AI" the same way. Someone signs the company up for ChatGPT or Claude. Staff get logins. A memo goes round about not pasting customer data in. And then each person is left to work out, on their own, what to ask, what to paste, what to trust and what to send.
To be fair to those products: the business plans do have real controls. Admin consoles, privacy settings, access management, and they keep getting better. If that's what you've got, configure them properly, because most businesses haven't.
But controls on a chat window still leave the hard part undone. The chat window doesn't know what job the person is trying to finish, which records are current, what that person is allowed to see, or what a good answer looks like in your business. Every staff member has to supply all of that, every time, from memory. Some will be brilliant at it. Most won't, and they shouldn't have to be.
So the real question isn't "which AI should we buy?" It's "what work do we want done, and how do we make that easy to do well?"
What the research says
I went looking for evidence rather than opinions. Four findings shaped how I build.
AI embedded in a specific job helps, and helps the less experienced most. A study of more than 5,000 customer support agents found an AI assistant built into their workflow lifted issues resolved per hour by roughly 15% on average, with the biggest gains for newer and lower-skilled staff. The expertise of the best people became available to everyone. Note what that was: not a chat login, an assistant inside the job.
The same tool can help or hurt depending on the task. In a Harvard and BCG experiment with 758 consultants, people using AI on tasks that suited it finished about 25% faster and got through about 12% more work. On a task just outside what the model could do, the AI users were less likely to get the right answer than people working without it. They trusted it past the edge. Choosing where to use it matters as much as having it.
The businesses getting value redesigned the work. McKinsey's State of AI survey found the organisations reporting the most value were nearly three times as likely to have fundamentally redesigned workflows. That's a survey, so it shows association rather than proof, but it matches everything I've seen: buying access and redesigning work are different activities.
Careless AI output makes work for everyone else. Researchers at BetterUp and Stanford gave it a name: workslop. Output that looks finished but pushes the real thinking, checking and fixing onto whoever receives it. In their survey of 1,150 US desk workers, 40% said they'd received some in the past month, and reported losing about two hours to each one. That's self-reported, not measured, but you'll recognise it: the three-page AI email that should have been one clear paragraph.
And one caution against believing your own feelings. In METR's 2025 experiment, experienced software developers took 19% longer with AI tools while believing they were faster. Tools have improved since, and I wouldn't use that to say AI slows people down. I use it to say: measure the actual result. How it feels is not data.
Start with the task, not the technology
Take a customer ringing for an update on their job.
In the chat-window version, the staff member opens the job system, copies the notes, opens the AI, pastes them in, types a prompt, reads the answer, wonders whether the completion date is right, edits it, and pastes it into an email. Six tools' worth of judgement, and it only works if they're good at all of it.
In the version I build, they open the job and press Prepare customer update. The system pulls the records that person is permitted to see, notices what's missing ("no parts ETA recorded"), writes a short draft in the company's format, and shows the evidence behind the completion date it's proposing. The staff member checks it and sends it through the normal route.
They didn't choose a connector, assemble context, remember a prompt or interpret a vague score. Those decisions were designed into the task once, by someone who understood both the business and the AI.
That's the pattern for everything:
- A clear job to complete. Recognisable actions: prepare a quote, summarise a job, find the approved procedure. Not a blank box.
- The right information. Current, relevant records, retrieved under that person's own permissions.
- Controlled actions. Read, draft, approve and send are separate permissions, enforced by the system.
- Useful output. The required format, length and pricing rules applied, and missing information flagged.
- Meaningful review. The reviewer sees the evidence, the proposed changes and the exceptions, not just a block of confident text.
- Measured value. Time to complete, time to check, corrections, cost, and whether the outcome was right.
And don't reach for AI where a rule will do. A weekly sales report should be calculated by the same queries every Friday. Use the model for the part that needs judgement, like explaining why a number moved, not to reconstruct the whole report from a fresh conversation each week. Anthropic's own guidance on building agents says the same thing: start with the simplest thing that works, and add autonomy only where it earns its extra cost and risk.
How I build it
Custom isn't the only way to get there, and the studies above don't prove it's always the best one. For a business that wants its data, its rules and its options kept in its own hands, this is the shape I use.
1. It runs in your own account
Not mine. Not a vendor's. An AWS account the business owns, in the region closest to you (for us, Auckland). The data, the logs and the configuration sit inside it. You hold the admin access. If you ever fall out with whoever built it, it's still yours.
Serverless and pay-per-use: Lambda for the logic, DynamoDB for the records, S3 for documents. No servers to patch, no idle cost at 3am, and a bill that follows actual use.
2. The model is a service you call, not a place you go
AWS Bedrock gives you Claude and other models as an API under AWS's terms. The model provider doesn't get access to your prompts or answers, and they aren't used to train the models. Your system sends the question plus only the snippets that person is allowed to see, and gets an answer back.
Two things I want to be precise about, because they're often oversold.
Where it's processed. An Auckland account does not by itself mean New Zealand-only processing. Bedrock routes requests across regions depending on the model and the profile you choose. A geography-bound profile keeps processing inside that geography (the Claude models we use run on an Australian profile today); a global profile can go anywhere. Your stored data stays in your account. Know which profile you're on, and write it down.
What's retained. Bedrock lets you configure whether prompts and outputs are retained, and some models require retention for abuse detection. So the honest statement isn't "it keeps nothing". It's "here is the model we chose, here is its retention setting, and here is where that's configured."
3. Cloudflare at the front door
Everything staff touch comes through Cloudflare: DNS, certificates, the tunnel into your systems, bot protection and access rules. No open ports. The only way in is the one you designed.
4. One login, with roles
Staff sign in with the Google or Microsoft account they already have. Every person has a role. The AI acts as that person: it can only answer from what the asker could already open. That one rule prevents most of the horror stories, and it's what the security people recommend too. OWASP's guidance on excessive agency comes down to: limit the tools, limit the permissions, act in the user's own security context, and get separate approval for anything consequential.
The rules that make it safe
Read-only by default. The AI can look at your job system. It cannot change a booking. It can read an inbox. It cannot send. Anything that changes the world is a person's click, with the draft and its evidence in front of them.
Evidence, not a confidence score. I used to show "high, medium, low". A model's opinion of its own answer isn't worth much. What's worth something is the evidence on screen: the procedure, the job note, the email the answer came from. And a source is something to check, not a guarantee. A wrong document quoted accurately is still wrong.
Tested rules for when to stop. If the system can't find a source, or the records disagree, or the question touches pricing, safety or a legal commitment, it says so and routes to a person. Those rules are written down and tested, not left to the model's mood.
Spend caps. A limit per person, per month, enforced by the system and tested, not a policy in a handbook.
An audit trail built to be tamper-evident. Who asked what, what it read, what it answered. Append-only and exportable, so when something goes wrong you can see exactly what happened.
Customer data never leaves its lane. Shared code, templates and examples hold placeholders. A customer's names, systems and suppliers live only in that customer's own records. A scanner runs on every change and refuses to ship if real customer details land in shared code. It exists because I got that wrong once.
Tested every release. A set of real questions with known right answers and known refusals. If a change makes the system sound more sure and be less right, it doesn't go out.
The model drafts. A person carries the risk. I've watched a large corporate let staff send AI-written material to customers unchecked. The AI knew a little. The staff member knew nothing. The customer got fiction. AI can hold everything there is to know about being a sparky. It still can't run the wire.
I used to say the reviewer must be someone who could have written the answer themselves. The research changed my mind a little: a well-built assistant genuinely helps less experienced people do better work, and that's the point of it. So the rule is now this. Whoever sends it must be able to verify the claims that matter, and must know when it needs a specialist. The system's job is to make both of those easy.
What's yours and what isn't
The business owns its account, its data, its configuration, its content, and anything built specifically for it. The engine underneath, sign-in, permissions, the audit trail, the ask engine, the connectors, is a set of building blocks I've already built and maintain across everything I build. Nobody should rebuild a second sign-in per customer. The blocks are used, not copied, so a fix ships once and everyone has it.
You don't have to use mine. A business can build its own. It takes the best part of a year, and then the upkeep is theirs: keeping it current, keeping it to ISO 27001 and SOC 2 standards, and watching it. That upkeep is the real cost of AI in a business, and it never shows up in the subscription pitch.
And if you ever want to walk away, the product and the data sit on top of the engine, not inside it. Moving across is a known, bounded job, not a rebuild.
Keep a record of what you build
Capture the decisions as you go: the idea, the choice, the date, who made it. For anything that matters as a trade secret, seal a fingerprint of the record and anchor it to a public blockchain. The secret stays in your register. The anchor shows that exact record existed, unchanged, on that date. It doesn't replace a lawyer. It means nobody is reconstructing history from memory a year later when there's money on the table.
What this costs
The cloud side is pay-per-use, so it depends on the workload: how many staff, how many tasks a day, which model. For a small business it is usually a modest monthly figure with hard caps so it can't surprise you, but anyone quoting you a number without asking about your workload is guessing. Cloudflare is cheap or free at this scale.
The larger costs are the ones people leave out: the build, the testing, the monitoring, and someone looking after it. Budget for those honestly. They're the difference between a system and a demo.
The short version
- Start with the task. Make the useful action one button, not a blank chat box.
- Give it the right records, under that person's own permissions.
- Separate read, draft, approve and send. Read-only by default.
- Show the evidence. Write down and test when it must stop and ask.
- Run it in an account you own. Know which model, where it's processed and what's retained.
- Spend caps, an audit trail, customer data in its own lane.
- The model drafts. The sender can verify what matters and knows when to call a specialist.
- Measure the whole task: time, checking, corrections, cost, outcome.
Make the useful action easy, enforce the boundaries in the system, and measure whether the whole job got better. That's AI in a business. Not a chat window with a company card attached.

