Can an AI persona actually model a general counsel?

Short answer: partly — and the part it misses is the part that matters.

A well-built synthetic persona can model how a general counsel talks about choosing a firm. It is far weaker on the unflattering reasons behind the choice, and it gets least reliable exactly where you need it most — the narrow, specific segment you actually sell to. Whether a vendor has engineered around those two failures is the whole difference between useful intelligence and confident fiction.

Here’s what’s really going on beneath “yes, we can model your clients.”

“General counsel” is not a person

The first problem is that a job title is not a client. A GC steering a mid-market energy firm through its first cross-border acquisition and a GC in a listed pharmaceutical company with a settled panel are different people making different decisions for different reasons. A tool that models “general counsel, 52, London” hands you the average of both — and the average corresponds to neither.

This isn’t a matter of opinion. In the largest study of its kind — 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) — AI agents built from demographics alone were 9 percentage points less accurate than agents grounded in a real two-hour interview with the actual person (74% against 83%), and 12 points behind agents given the interview and the person’s survey answers together (86%), running on the same models. Agents built from a persona description did worse still, at 71%. The label leaves most of the human on the table.

And it gets worse, not better, as the segment narrows. “General counsel in mid-market German technology firms” is precisely the kind of small, specific cell where synthetic tools are least reliable — and it’s also exactly the client a firm most wants to understand. The more valuable the segment is to you, the thinner the ground you’re standing on.

The fix is to stop modelling a job title. Vary what moves the decision: sector, matter type, panel or direct instruction, the state of the existing relationship, whether the firm is stable or mid-transition. Do that across hundreds of simulated clients and there’s no single “GC average” left to hand you. You read where they agree and where they split instead. Taleist builds its Client Proxies this way.

The respectable-answer problem

The second failure is subtler and more dangerous, because it produces answers that look perfect.

Ask a model why someone chose a particular option and it gives you the respectable reason. The way you catch it is with a control: keep the question, swap the subject, and watch the same answer come back. We first saw this on an early project — every simulated parent said they chose a children’s activity for “confidence building.” We didn’t believe the uniformity, so we swapped the activity for a completely different one and asked again. Same answer. The model wasn’t telling us what parents valued about that activity. It was telling us what a good parent is supposed to say.

A synthetic general counsel does exactly the same thing. It won’t readily admit it kept the incumbent firm because switching was a risk to its own standing with the board. It won’t admit it chose the recognised name over the better-fitted boutique to feel safe, or that a directory ranking reassures it more than it would ever say out loud. Instead it hands you the respectable list — “expertise, responsiveness, commercial understanding” — the same list every firm already claims and no firm can act on. The real drivers, the ones that actually decide instructions, are risk, reputation-cover and relationship inertia. They are exactly what the model is built to hide, because it wants to sound like a competent GC.

This is a property of the models themselves, not a flaw one vendor happens to have. Language models are trained to be helpful and to avoid the controversial or the unflattering, so they systematically under-report the self-protective, professionally risky reasons that drive real decisions. A provider who doesn’t know this reports the respectable answer as a finding — and you build a pitch on a decision nobody actually makes.

The way past it is a question that makes the respectable answer cost something. Tell the simulated GC its current firm has proposed a 12% fee rise and an equally credentialed rival will hold the old rate. Then ask what it does. “Expertise and responsiveness” can’t answer that. Risk, reputation-cover and inertia can, and they show up in the reasons.

What it takes to model a GC properly

None of this means a synthetic GC is worthless. It means the modelling is hard, and the hard parts are where a serious provider earns their fee.

It takes a willingness to ask the questions you could never put to a real client (“would you still instruct them if they dropped out of Band 1? why does a ranking matter after eight years of good work?”), which is the one real advantage of a synthetic client: there’s no relationship to damage. And it takes each simulated client run on its own, so a hundred of them can’t drift towards one safe consensus.

So — can it?

Here’s the honest verdict.

A properly built synthetic persona can model the attitudinal landscape a general counsel’s decision sits within — the considerations in play, how they differ across client types, which language persuades and which falls flat — at a breadth and candour you could never get from real GCs without making them feel cross-examined rather than consulted. That is genuinely valuable, and it’s the right use of the tool.

It cannot model the specific human, and it cannot predict the specific instruction. Used to understand the shape of how your market decides, it’s powerful. Sold to you as a machine that tells you what your clients will do, it’s dangerous — and the vendor selling it that way has told you something about how much they understand their own product.