Choosing an AI consultant is one of the few procurement decisions a regulated institution makes where the wrong answer is not a wasted budget line but a governance liability — because whoever touches a regulated process inherits its requirements. The deciding signal is almost never in the pitch deck; it is in the answers to a handful of specific questions. Twelve of them, grouped into four areas — technical depth, implementation process, track record, and commercial structure — will tell you more in one meeting than a month of proposals. Each has an answer that should end the conversation.
Start with technical depth, and make the first question about data. Ask where your data goes and whether it ever leaves your environment or Switzerland; a vague answer, or a reassurance without a diagram, is the red flag. Ask what catches a wrong answer before it reaches an investor and how you roll it back — a consultant who only ever describes the happy path has no failure design, and in a FINMA-regulated process the failure design is the design. Then ask whether the calculation stays deterministic and auditable, with the model sitting around it rather than producing the numbers itself; if the model is the calculation, the process cannot be reconstructed after the fact, and reconstruction is exactly what an audit demands.
The second area is how they actually work. Ask what baseline they would measure on day one and how they would show you the improvement six months later; if the answer is a feeling rather than a metric — cycle time, exception rate, the number of manual touches between the data and the investor — walk. Ask how the new workflow runs in parallel with the existing one before anything goes live; a consultant who proposes a single cutover switch has never had to reconcile an automated output against a manual one under audit. And ask for the three success criteria and the kill criteria of the proof of concept, agreed before the work starts — a proof of concept with no pre-defined way to fail is not a test, it is a commitment in disguise.
Third, press on track record, because this is where confidence and competence diverge. Ask them to walk you through one engagement end to end — the process before, exactly what changed, and the single number that moved; a real practitioner answers in specifics, a bluffer retreats into adjectives. Ask which regulated processes they have actually delivered and whether you can speak to a reference who is not their flagship client; expertise in every industry is expertise in none, and the constraints of regulated finance — data residency, auditability, sign-off — are not something a generalist discovers gracefully at your expense. Ask who specifically builds the work and where they sit while your data is in play; in regulated finance that is an outsourcing and data-residency question, not an administrative detail, and you are entitled to a straight answer before a single file changes hands.
The fourth area is the one most institutions underweight: the commercial structure. Ask what they deliberately do not take on — a serious partner has a boundary and will name it, and a consultant who answers every scoping question with "yes, we can do that too" has shown you the most important red flag of all. Ask whether you own the workflow, the code, and the documentation at the end; a process you cannot run or audit without the vendor in the room is a dependency, not an asset. And ask who owns the workflow after go-live and what ongoing accountability looks like, because a partner who treats a regulated automation as a one-time project has not planned for the monitoring, the drift, and the audit requests that arrive long after the invoice is paid.
Behind the twelve sits the one question we wish more institutions asked: what workflow are you not recommending we automate, and why? The best partner narrows the scope rather than widening it — starts with the single workflow whose baseline is provable, measures it, and lets the proven case fund the next one. When reporting across 39+ funds at a leading Zurich investment foundation was automated, that was the whole method: not a sweeping transformation, but one narrow, measurable workflow taken through exactly this discipline before anything else was attempted. A consultant eager to automate everything at once is not ambitious; they are unpracticed in what regulated delivery actually costs.
None of these questions is technical, and that is the point. Choosing an AI consultant in regulated finance is a test of judgment, not of jargon — the partner worth hiring is the one who measures before they build, designs for the answer being wrong, and hands you an audited process you own outright. Ask the twelve, listen for the red flags, and let the answers do the vetting. That is the same discipline that runs through everything we build.