Smells like AI: Why customer trust breaks when automation shows up in the wrong moment

This post is sponsored by Maker Lab.

There’s a specific moment that happens when a customer realises they’re talking to a machine. The tone shifts, the response cadence is wrong, the answer doesn’t quite track what they actually asked. From that point, every future interaction gets filtered through that lens. Fair or not.

The number one thing said in AI chat bots on a bank or airline website is “agent” (this is probably true). Chatbot agents have been around for about 10 years at this point and people still hate them as much as they did when they came out (even though they probably work).

Trust breaks asymmetrically. Years of good human service can compress into “this company just uses AI now” after one poor interaction. The recovery from that label is slow, and the reputational cost sticks around long after the technology is pulled back.

Trust disappears when AI shows up in moments where customers expected judgement, rather than simply because brands use AI. That’s an important distinction.

Why the recent chatbot roll-outs went the way they did

Several fast-food chains, including McDonald's, piloted AI voice ordering in 2024.

The demos were genuinely impressive. The real-world roll-outs, however, became something else entirely. From misheard orders and repeated questions to viral videos showing systems adding ketchup and butter to ice-cream, the technology produced high-profile errors while live.

Then the AI-enabled systems were pulled back, typically disappearing without a public announcement. Just quietly vanishing like their absence wouldn’t be noticed. The complaints and videos, though? They kept circulating.

The customer experience has a much lower tolerance for failure than internal productivity does, which is the main lesson from these voice-ordering roll-outs.

The tool cut internal costs, but it also created external friction in a setting where friction is catastrophic. A customer who has to repeat themselves three times just to get a meal doesn’t make a rational calculation about your “cost-per-transaction”. They’re counting the time they’ve wasted.

Then they leave. And they tell people.

A few seconds saved per transaction means very little if every failed interaction chips away at trust.

We watched those roll-outs from the outside and thought about how we’d have framed the brief differently.

Most organisations still ask, “Can AI handle this interaction?”, and the answer is often “yes”. The better question is: “Can this interaction survive an AI misjudgement, failure or mistake?” Every customer touchpoint carries a different trust threshold. Some are purely transactional, others depend almost entirely on human judgement. For high-frequency, friction-sensitive touchpoints, the answer is almost always no.

What actually smells like AI: The filters

Not every interaction carries the same risk if a customer clocks it as an LLM. What matters is the primary objective of the interaction, rather than model capabilities alone.

Before delegating an external touchpoint to automation, there are four questions worth asking:

1. What are the emotional stakes? A parking reminder can be automated invisibly. A complaint, a cancellation, a bereavement notification cannot.
2. How personal is the interaction meant to feel? Anything marketed as “we know you” (loyalty outreach, a birthday message, a “we noticed you might like...”) breaks harder when it’s from a bot that got it quite wrong. The premise was individual attention, and that the brand cared, but automation turns it transactional and hollow.
3. What does failure cost the customer, not just the company? A wrong answer on store hours is a minor inconvenience. A wrong answer on a medical, financial, or safety question is not (such as Air Canada's chatbot misinformation case), becoming an immediate legal and reputational crisis.
4. Is the interaction high-frequency or a one-off? High-frequency, low-stakes interactions (order status, simple FAQs) build tolerance for AI over time. One-off, high-stakes ones (a complaint, a first sales conversation) don’t get a second chance to build that trust.

The harder problem: Who owns the outcome

With that said, even a brand that nails every filter above can still fail at this, because each of them is a per-interaction test, whereas trust is an experience across the entire CX journey.

Marketing owns the content. Technology owns the platform. Operations owns the efficiency. Customer service owns the complaints. And each of them can make a perfectly sensible call to automate their own piece. Marketing puts the birthday message on autopilot. Operations automates delivery updates. Customer service lets a bot triage first-line complaints and escalates the rest.

All sensible, low-risk calls on their own, and nobody sat down and said “let’s make this brand feel like a robot”.

But also, no one is looking at the customer’s week, only at their own piece of it: One customer can get an automated birthday message, an automated delivery update, and an automated complaint response in the same few days – three defensible decisions made by three different teams, yet ultimately adding up to “every time I deal with this brand, I get a bot”.

Beyond just being a technology gap, it’s also an accountability gap. Closing it takes someone positioned to see the whole journey, not just their piece of it. Otherwise, the AI deployment becomes a collection of disconnected optimisation projects instead of one coherent customer journey.

Being intentional with your AI footprint

For some interactions, discovery costs nothing. Chatbots handling straightforward queries are widely accepted. For others, it’s damaging: anything meant to feel personal, responsive or emotionally attuned carries a much higher bar.

Brands that get this right are deeply intentional about where AI belongs, treating automation as a clear fork in the road: either they invest until the AI is good enough, that is, accurate, and genuinely personal, to earn a place in a sensitive interaction. Or they deliberately hold it back because the cost of a misfire is more than the business or the customer can absorb.

What trips brands up is skipping that fork entirely: treating “we technically can automate this” as the same question as “should we, here, right now”. Those aren’t the same call.

The future will belong to the companies using AI with the most intention, rather than those using the most of it. Customers rarely judge AI on benchmark scores or model names; they judge the experience.

And when something smells like AI in the wrong moment, trust disappears much faster than the technology improves.

The author is James Tugman, AI strategy and implementation manager at Maker Lab.

Maker Lab is a human-led, AI-powered in-house agency. In-housing is fast becoming the new model for marketing. Maker Lab specialises in making it work: building the teams, systems, capabilities and AI behind in-house marketing across APAC.

Want to learn more, get in touch: john@wearemakerlab.com