The First 90 Days: Deploying AI Inside a Wealth Firm

Most AI deployments in wealth firms do not fail. They stall — somewhere between the demo everyone liked and the security review nobody scheduled.

This is what the first ninety days look like when it goes well, including the weeks that feel like nothing is happening.

Before day one: three decisions

Settle these first, because each one changes who can bid and what the timeline looks like.

Where it runs. Inside your environment, or a vendor's. Many private banks have no latitude here, and discovering it in week six restarts the procurement.

Which questions matter. Five is enough. Written down, agreed, specific — "what is our exposure to this issuer across all mandates", not "better insight".

Who owns it. One named person with the authority to decide what a good answer looks like. Deployments without this drift until someone senior loses patience.

Weeks 1–2: connect one source

Not all of them. One — usually the portfolio system, because it answers the most questions.

This fortnight is unglamorous and tells you most of what you need to know. Whether the credentials exist, whether the data means what the documentation claims, whether anyone can get a test environment. Firms where this takes four weeks instead of two have learned something important about the rest of the programme.

A connected source and no interface at all is a better position than a polished interface over a demo dataset.

Weeks 3–4: the first real answers

Your five questions, against your actual data, in front of the people who asked them.

Expect the first answers to be wrong in instructive ways — a field that means something different in practice, a convention nobody documented, a fund that is named three ways. This is the work. Every correction here is a correction that does not reach a client later.

The thing to check is not fluency, it is traceability: can each figure be followed back to where it came from. An answer you cannot verify is not usable in a regulated firm, which is why grounding and citation are architectural rather than cosmetic.

Weeks 5–8: the security review

Start it in week one, not week five. It runs in parallel and it is the single most common reason a deployment slips a quarter.

What gets asked is predictable: where client data goes, who can see what, what is logged and for how long, who is accountable for the output, and what happens when it is wrong. Our guide to the questions compliance actually asks is the full list.

Two things shorten this more than anything else. A data-flow diagram, which answers half the questions in one picture. And a demonstration of the system refusing to answer something it cannot ground — nothing else builds confidence as quickly, because it shows the controls are real rather than described.

Weeks 6–10: one desk, real work

Overlapping the review, put it in front of a small group doing their actual job. Six people beats sixty: you get specific feedback instead of a survey, and the ones who like it become the reason everyone else tries it.

Measure something countable. Minutes to assemble a client review. Questions answered without emailing operations. Breaks cleared before lunch. "People seem to like it" will not fund phase two.

Expect usage to dip in week two of this period. The novelty passes and the habit has not formed. Firms that panic here and add features usually make it worse; the fix is almost always a missing data source, not a missing button.

Weeks 10–13: decide, then widen

By now you should be able to answer three things with evidence: did it save measurable time, did compliance sign off, and did anyone keep using it without being asked.

Three yeses means connect the second and third systems and widen to the next desk. A no on the third is the one to take seriously — it usually means the five questions were the wrong five, and the remedy is to re-pick them with the people who did not adopt it.

What actually causes the stall

Starting the security review late. It is parallel work treated as sequential, and it costs a quarter every time.

Connecting everything at once. Eight integrations before a single answer means eight things that can be late, and no evidence to show for the wait.

Rolling out firm-wide on day one. Without a desk that already vouches for it, a wide launch is a wide shrug.

No named owner. Everything above is a decision, and decisions without an owner become meetings.

Common questions

Is ninety days realistic? For one desk, a defined set of questions and two or three connected systems, yes. For a firm-wide rollout across every source, no — and any vendor promising that is describing a demo.

What if our data is messy? It is, everywhere. That is why phase one connects one source: it surfaces the mess at a scale you can fix, rather than across eight systems at once.

Do we need a data project first? Usually not a separate one. Connecting the first source is the data project, and it is better funded by a deployment with a visible outcome than by a tidy-up nobody can price.

What should we not attempt in ninety days? Anything where a wrong answer reaches a client unreviewed. Keep a person between the output and the client for the whole of the first quarter — see what to automate first and what to leave alone.

Start with the five questions

Write them down before you talk to anyone, vendors included. They determine the integrations, the cost and whether week four produces anything worth showing. Book a demo and bring them.