Short answer: no — not reliably, and the honest providers won’t claim otherwise.
Synthetic research reads how your clients think well. It predicts what they’ll do far less well. And it’s weakest of all on what they’ll pay. That distinction isn’t pedantic — it’s the line between using the tool correctly and betting a real decision on fiction.
Attitude is not action
The single most useful fact about synthetic research is also the one vendors are quietest about: it is much better at attitudes than at behaviour.
The evidence is unusually clean. In Park et al., LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals (first published as Generative Agent Simulations of 1,000 People; Stanford and Google DeepMind; arXiv, version 3, June 2026) — the largest study of its kind — AI agents built from two-hour interviews with real people matched those people’s attitudes on the General Social Survey 83% as accurately as the people matched themselves two weeks later. Strong. But on five behavioural economic games — real decisions with money and stakes — the same agents fell to a normalised correlation of 0.66: about two-thirds. The gap between those two results is the gap between knowing how someone thinks and knowing what they’ll do.
How someone thinks, against what they do
For a professional services firm that gap is the whole ballgame. “How does a general counsel weigh reputation against relationship?” is an attitude question — synthetic research is good at it. “Will this particular GC move their litigation work to us next quarter?” is a behaviour question — and there the tool is far shakier.
Why the gap doesn’t close with better data
Here’s the part that should settle the matter. In the Stanford study, giving the agents much richer input — a full two-hour interview instead of a demographic sketch — lifted attitude accuracy sharply, from 74% to 83%. It did nothing measurable for behaviour. On the economic games, no agent type — interview-grounded included — did significantly better than agents built from a thin demographic description.
That tells you the behavioural gap isn’t a data problem you can spend your way out of. It’s structural. What a person actually does depends on the situation they’re in at the moment of deciding — how busy they are that week, whether a partner rang them last month, whether procurement is pushing them to consolidate the panel, whether their CEO happened to praise a rival at a board dinner, whether they’ve just had a bad experience with the alternative. A synthetic persona has the attitude layer clean. It cannot have the timing, friction and accident layer, because that layer is lived, not stored. No amount of better modelling reaches it.
So a tool that tells you why reputation matters to a GC is on solid ground. A tool that tells you whether a rankings drop will make GC X switch is guessing, and dressing the guess as data.
Price is the hardest case of all
If behaviour is shaky, price is shakier still. Synthetic methods are least reliable precisely on price sensitivity and willingness to pay — the questions where a wrong answer costs the most. Even the platform vendors concede this: the responsible ones explicitly warn against using synthetic data for major pricing commitments and other high-stakes final decisions.
Which means the one thing you must never do is set or defend your fees on the strength of what a synthetic panel “would pay”. If a vendor offers you a willingness-to-pay number, that’s not a capability — it’s a red flag.
What synthetic research can reliably tell you
None of this makes the tool weak. It makes it specific.
Synthetic research predicts considerations, not outcomes. It can tell you, at real depth and across hundreds of client types, what makes a switch feel safe or risky, which reassurances actually matter, what a “deep bench” claim means to a GC (usually nothing they can take to their board), and which of your messages land and which bounce. That is exactly the input a firm needs for positioning, pitch strategy and messaging.
What it can’t do is tell you who will win the deal, whether this client will churn, or what price clears the market. For those, the number is closer to a coin toss than a forecast — and the honest use of synthetic research is to shape the decisions those questions sit inside, not to answer them directly.
The tell in a vendor pitch
This gives you a fast, brutal test for any provider. If they hand you a behavioural or pricing figure — a percentage who’ll switch, a price your clients will bear — with a straight face, they’ve crossed the exact line the research draws, and they’ve told you they either don’t understand their own tool or are willing to oversell it. The provider worth your money volunteers the limit before you raise it: “we’ll tell you how they think about switching, not whether they’ll switch.”
That’s the first of the questions worth putting to any vendor. For the full set, see the companion guide, how professional services firms should evaluate synthetic research.
Taleist builds decision intelligence for professional services firms — research designed around how your clients actually choose, not a forecast of what they’ll do that the evidence can’t support.