The pitfalls of running your business on ChatGPT and Claude.

Half the businesses we talk to in 2026 have already wired ChatGPT, Claude, or Gemini into their daily work. Not as a product — as a habit. The marketing manager pastes a customer email into Claude. The accountant uploads a P&L and asks for "the punchlines." The realtor writes listing descriptions in ChatGPT. The lawyer drafts NDAs in a private Claude project.

It works. It's faster than not using it. And it is, in many specific cases, the single biggest productivity unlock anyone in that role has had in a decade.

It is also, when used as the backbone of a business workflow, a slow-motion mistake. Here's where the costs hide.

1. The "free" trap — $20 a month looks cheaper than it is

The seductive thing about a $20/month ChatGPT seat is that it looks like the entire AI bill. It isn't. The real costs are paid in time, quality, and risk.

Time, because every time the assistant doesn't remember last week, someone has to re-explain the business — the products, the tone, the customer list — at the start of the new conversation. Multiply that by every person on your team, every day.

Quality, because two of your people will end up asking the same question and getting two subtly different answers — and you'll only catch the divergence when a customer points it out.

Risk, because every paste is a tiny act of trust: you are sending a snippet of your business into someone else's data centre. Sometimes a contract clause. Sometimes a customer's email. Sometimes (and this is the one that hurts) the contents of the spreadsheet your CFO sent you "just for you, please."

2. Where your data actually ends up

Most people read "we don't train on your conversations" and stop there. That's a real commitment, and the major vendors do honour it on paid tiers. But it is not the whole story.

Your conversation is still retained — typically for 30 days, longer if a content-policy flag is triggered. It is still readable by the vendor's incident-response team under specific conditions. It still leaves your country. And it is still subject to the law of their jurisdiction, not yours.

If you are a Canadian SMB regulated under PIPEDA — or worse, under PHIPA, FINTRAC, or law-society rules — that distinction matters. "We don't train on it" is not the same as "your client's name never crossed the border."

3. The workflow problem — chats are not pipelines

The single most expensive mistake we see is treating a chat as a workflow. They aren't the same thing.

A chat is brilliant for exploration. You don't know what you're going to ask next. You're talking. You're iterating. You're the human-in-the-loop, in the most literal sense.

A workflow is the opposite. It runs every Tuesday at 8am. It does the same thing every time. It has inputs from your CRM, outputs to your accounting tool, an escalation path when it gets stuck, and a place to look when something goes wrong.

If you've ever caught yourself writing the same prompt three weeks in a row, you've already discovered a workflow. The chat window is just the wrong place to run it.

4. Memory is not a feature, it's an architecture

Both ChatGPT and Claude now have "memory" features. They are useful, and they are not enough.

The memory in a consumer AI app is, by design, scoped to you. It remembers that you have a dog named Bagel. It does not, and should not, remember that your firm has an open file for client #14823 with a billing rate of $475/hr and an outstanding trust balance of $12,500.

For a business, "memory" means: this AI knows our clients, our products, our pricing, our policies, our precedents, and our brand voice — and it knows all of it the same way, every time, regardless of who's asking. That is not a feature. That is a data architecture. It lives in a vector index, a structured database, a retrieval layer, and a permission system. It is the part the consumer app cannot build for you.

5. Hallucinations get worse in a workflow, not better

In a chat, you can spot a hallucination because you're reading the answer. In a workflow, the answer becomes the input to the next step. The hallucination compounds.

Picture three plausible failure modes: the realtor whose listing description "remembers" a fireplace the property doesn't have. The lawyer whose AI summary of a discovery doc invents a witness. The accountant whose AI categorisation moves legitimate expenses into a personal-use bucket. None of these failure modes happen because the model is bad — they happen because there's no checking layer between the model and the action.

6. No audit trail, no defence

Here's the one almost nobody asks until it's too late: when something goes wrong, can you reconstruct who asked what, when, with what context, and what the AI said?

In a native chat app, the answer is usually: kind of, if the person remembers to look, and if the conversation hasn't been deleted, and if you have admin access to that user's seat. That's not a defence. That's an excuse.

A business workflow needs the same audit trail your CRM has. Who, when, why, what changed. The native AI apps weren't built for this — they were built for a person talking to a model. Asking them to be the system of record is asking the wrong tool to do the wrong job.

So what's the alternative?

This is where we'd usually be expected to pitch the product. We're not going to — at least not directly. But we will tell you what the shape of the alternative looks like, because it's worth understanding even if you build it yourself.

1. Put the AI inside the workflow, not next to it. The model should be triggered by something in your CRM, accounting tool, or matter system — not by a person opening a tab.
2. Give it your data, not their data. The model should query a retrieval layer that contains your clients, your precedents, your brand voice. Not the open internet.
3. Log everything. Every call, every input, every output, every action taken downstream. This is the audit trail you'll wish you had when something goes sideways.
4. Build a check layer. The output should not be the final action. It should be a draft a human signs off on — or a small, well-scoped action a programmatic guardrail validated.
5. Be honest about where the model lives. If you're in regulated work, your model needs to run in a way that doesn't move your data across the border in ways that get you in trouble. There are now ways to do this without a $2M enterprise contract.

A short checklist before your next Monday morning

• Audit your team's AI tabs. Make a list of every recurring task someone is doing inside ChatGPT or Claude this week.
• Find the three with the longest prompts. Those are the ones most worth moving into something real — they're the workflows hiding in plain sight.
• Ask where the data goes. For each of those three, ask: what is the most sensitive thing being pasted into the chat? Is that OK with your customers, your regulators, and your future self?
• Decide what "good" looks like for each. If you could wave a wand, what should the workflow do automatically, every time, with audit, with escalation, with the same answer every time?
• Then build it — or pick a tool that already has. This is the part where, full disclosure, we'd love to help. But the first four steps are worth doing regardless.

The native AI apps did everyone a favour by making AI usable overnight. They were the on-ramp. They were not the destination. Don't park your business on the on-ramp.