Picture the scene: a team needs to know whether a new product idea will land. Instead of booking interviews, they ask an AI to simulate the customer conversation. The AI says yes. Nobody in the room checks whether a real customer would agree, because the answer already sounds confident enough to move on.

That's synthetic research: AI systems that simulate customer responses instead of collecting them from real people. It's fast, and it's exactly why AI governance in Australia needs an answer to a harder question: when does an AI-generated result count as evidence, and when does it still need a real person to check it?

Instead of recruiting participants and running interviews over several weeks, teams ask AI to roleplay a customer segment and generate the responses that segment might give. Qualtrics reports that 71% of market researchers believe the majority of market research will be synthetic within three years.

Where it actually helps

Exploratory work is where synthetic research earns its keep: stress testing a proposition, building an early persona, getting a team aligned on who the customer is before a full study starts. What used to be a month-long process, getting baseline qualitative and quantitative data in front of a customer journey team, can now happen in a day.

That's not a shortcut around research. It's a faster way into it. A team that walks into a full study already aligned on the proposition and the persona spends less time relitigating assumptions and more time testing the ones that matter. The research budget goes further because it's spent on the questions that survived the first pass, not on rediscovering what a synthetic run could have flagged in an afternoon.

Where it goes wrong

The problem starts the moment a synthetic answer gets treated as evidence. A simulated customer will validate a proposition a real one wouldn't, because it has none of the friction a real person brings. It doesn't get confused. It doesn't change its mind halfway through a question. It never reveals the gap between what someone says they'd do and what they actually do once the product is in front of them.

Companies can get lazy here. It's easy to open a simulator, ask whether a product is valid, and treat the answer as settled. Real customers stutter. They reconsider. They say one thing and do another. That gap between stated and actual behaviour is often the most valuable data in the room, and synthetic research will never surface it. A simulator has no incentive to tell a team its idea is wrong. A real customer, mid-interview, often does exactly that, without being asked to.

Models like Anthropic's Claude are genuinely good at generating a plausible, coherent answer. Plausible isn't the same as true, and a team that can't tell the difference will act on the wrong signal with more confidence than the evidence earns.

What this means for customer experience teams

The decision that matters is what's actually being decided, and that call usually sits with whoever owns the customer relationship, not whoever owns the model. Exploratory work and early pressure testing are a good fit for synthetic research. A launch decision, a pricing decision, or a build decision needs a real person's answer first, not a simulated one.

The speed synthetic research offers is real. So is the risk of mistaking a confident-sounding answer for something a team has actually earned the right to act on.

Not sure whether your last research sprint told you the truth, or just agreed with you?

The Omnia Collective's customer experience team pairs synthetic research with structured customer validation, so you know before you build whether you're testing real behaviour or a plausible-sounding answer. It takes a couple of hours to map where your current process sits on that spectrum. See our customer experience services or get in touch.

This post is adapted from Episode 5 of our Nap Stack podcast on AI, data, and building a business. Listen on Spotify or Apple Podcasts.