Most AI initiatives in regulated finance do not fail because the technology was not ready; they fail on a handful of avoidable mistakes made long before the model is ever chosen. The pattern is consistent enough to name. Across the institutions that stall — a fund management company, an asset manager, an investment foundation — the same errors recur, and each one has a prevention move that is neither expensive nor technical. What follows is eight of them, in the rough order they tend to appear, from the first scoping conversation to the year after go-live. None is about algorithms. All of them are about the discipline of touching a controlled process, which is the part a generic AI vendor rarely respects and a regulated institution cannot afford to skip.
The first mistake is starting with the tool. An institution decides it needs "an AI solution", shortlists platforms, and books demos before anyone has written down the workflow the tool is meant to serve — so the software arrives looking for a problem, and the problem it finds is rarely the one worth solving. The prevention is to start with the process, not the product: map the workflow end to end and let it dictate the tool, not the reverse. The second mistake follows directly — automating a process that was never sound to begin with. An automation built on an undocumented workflow does not fix the confusion; it industrialises it, producing the same unreliable output faster and with the appearance of authority. Describe the process, name its control points and sign-offs, and repair what only ever held together through one experienced person's judgment, before a line of it is automated.
The third mistake is building with no rollback — designing only the path where everything works. In a FINMA-regulated process the failure design is the design: what catches a wrong figure before it reaches an investor, how you revert to the previous process within the hour, who is notified. A partner who only ever describes the happy path has not built for the day the answer is wrong, and in regulated finance that day arrives. The fourth mistake is ignoring data quality. AI inherits the data foundation it is given, and in a controlled process it inherits the audit exposure with it — the same figure carried at two values in two systems, a lineage no one can trace, a spreadsheet no one fully trusts. Audit the data before building on it: the cleanup folded into the plan is cheap, the cleanup discovered in production is not.
The fifth mistake is skipping the assessment — committing to a build before anyone has measured a baseline or scoped the real work. Without a starting number the initiative can never prove its return, detect a regression, or answer the first question an internal auditor asks; the assessment is what turns a hopeful budget into a defensible one. The sixth mistake is hiring a freelancer for a partner's job. A quick contractor can build a clever demo, but a regulated automation is a running commitment that has to be monitored, kept current, and made ready to answer a FINMA request long after go-live — and a process whose only author has moved on is a finding waiting to be written. Match the engagement to the accountability the process carries, not to the lowest quote.
The seventh mistake is picking the loudest vendor. The firm with the most polished pitch and the widest claims is rarely the one that will still be answerable two years on; the signal worth weighting is evidence of practice — a partner who writes concretely about the failure modes of a regulated automation, who names a boundary and a reference that is not their flagship — not evidence of presence. The eighth and most expensive mistake is treating AI as a one-time project. A regulated automation is not bought and finished; it drifts, the underlying systems change, and the audit requests keep arriving after the invoice is paid. A partner who hands over a "completed" project rather than a maintained process has quietly moved the ownership cost onto your desk. Ask who owns the workflow after go-live, and what keeping it audit-ready costs each year, before you sign.
Read together, the eight share a single root: the temptation to move fast by skipping the unglamorous parts — the process map, the baseline, the failure design, the question of who owns it afterwards. The institutions that avoid all eight are not the ones with the biggest budget or the newest model; they are the ones that narrowed the work rather than widening it, started with the single workflow whose baseline they could prove, measured it, and let the proven case fund the next. Boiling the ocean is not ambition; it is the mistake that contains the other seven.
That measured, one-workflow-at-a-time discipline is exactly how reporting across 39+ funds at a leading Zurich investment foundation was built — not a sweeping transformation, but a single provable process taken through this rigour before the next was attempted, each earning the one after it. If you want to avoid these mistakes before they cost anything, an AI Audit is where it starts: it maps the workflow, establishes the baseline, and scopes the real work before a single decision is committed. It is the same discipline that runs through everything we build.