TL;DR: Before AI touches a customer, four answers must be clean: it can’t make a bad moment worse, you know exactly where their data goes, there’s always a fast path to a human, and you’d be comfortable if the customer knew what was happening. Any wobble — don’t do it yet.
Before you let any AI touch a customer — a chatbot, an auto-reply, an “AI agent” handling enquiries — ask yourself four plain questions: can this make a bad situation worse, where does my customer’s data go, is there always a fast and obvious path to a human, and would I be comfortable if my customer knew exactly what was happening? If any one of those answers wobbles, the honest move is to not do it yet. Customer-facing AI is where the hype does the most damage, because the cost of getting it wrong lands on the people who trust you.
Why customer-facing is the riskiest place to start
Internal admin — the quoting, the reporting, the scheduling — is forgiving. If AI drafts a clumsy internal note, you fix it and move on. But the moment AI faces a customer, the stakes change. A bad answer to a frustrated person, a privacy slip, a bot that traps someone who needed help — those don’t stay internal. They become the experience your customer remembers. So I treat customer-facing AI as the last place to deploy it, not the first, and only after these questions are answered cleanly.
The four questions
1. Can this make a bad situation worse? The customers most likely to hit your AI are often the ones already annoyed, confused or in difficulty. Will the tool help them, or hand a distressed person a wall? If it can turn a small problem into a furious one, it’s not ready to face customers.
2. Where does my customer’s data go? Whatever a customer types into an AI tool, where does it end up? Stored where? Seen by whom? Used to train what? If you can’t answer that plainly, you can’t responsibly put it in front of the people who trusted you with their information.
3. Is there always a fast, obvious path to a human? This is the one I won’t budge on. Every customer-facing AI must have a visible, easy escape hatch to a real person — no maze, no bot gauntlet. “Press 0”, “call this number”, “message a human here”. If the design makes reaching a person hard, the design is wrong.
4. Would I be comfortable if the customer knew exactly what was happening? The honesty test. If you’d hide from your customer that they’re talking to a bot, or that their data is going somewhere, that discomfort is telling you something. Good customer AI survives the customer knowing about it.
Customer-facing AI should make help easier to reach, never harder. The day it stands between a person and a human who could help them, you’ve built the wrong thing.
The neurodivergent-led reason this matters to me
I’ll be straight about why I push this so hard. FL3P is neurodivergent-owned, and I know what it’s like to be overwhelmed and meet a chat widget that doesn’t understand you and won’t let you out. That experience is why every page we build has a fast path to a calm human. Designing for the worst moment your customer might be having — not the happy-path demo — is the whole job. Most “customer AI” is designed for the demo. That’s the difference.
What I’d actually do
Start AI away from your customers. Prove it on internal work first, where mistakes are cheap and nobody’s trust is on the line. Only bring it customer-facing once all four questions answer cleanly — and even then, lead with the human path, not the bot. The businesses that get this right use AI to get help to their customers faster. The ones that get it wrong use it to keep customers at arm’s length and wonder why loyalty drops.
If you’re weighing up putting AI in front of your customers and want an honest gut-check first — including a straight “not yet” — that’s exactly the kind of thing worth running past a real person before you commit.