Eighteen months ago, agentic AI in wealth management was almost entirely a slide-deck conversation. Today, in conversations with twenty-eight firms across Europe, North America, and APAC, the picture has shifted. The deployments are real, the volume is small but growing, and, perhaps most importantly, the operating model around them has started to settle.
We define "agentic" the way the firms we spoke to do: workflows that include retrieval, reasoning, and the autonomy to take a small number of bounded actions on behalf of an advisor or a back-office team member. That definition is narrower than the marketing claims but wider than a chatbot. It is the layer where most of the productivity wins are now happening.
Three patterns kept showing up across the sample. They are not surprising in isolation; they are surprising in how consistently they recur. Firms that ignored any one of them ended up rolling back deployments.
1 / From single-purpose tools to multi-step workflows
The first generation of AI deployments in advisory firms looked a lot like the early generation of dashboards: each one solved a specific, scoped problem, meeting summarisation, KYC document extraction, a knowledge-base assistant for advisors. Each one delivered a measurable win. None of them changed how the work was done.
The firms making real progress in the past two quarters have rebuilt around workflows, not tools. The output of one model becomes the input of another. The output of one human becomes the input of the next agent. A typical example, in a Switzerland-based private bank we spoke to: the post-meeting summarisation agent now generates a follow-up draft email, queues a calendar invite, attaches the relevant published research, and routes any compliance-sensitive lines for review. Eight months ago, those were four different tools.
"We stopped buying point solutions. We started designing workflows that happened to use AI in three or four steps."
The implication is unglamorous but important. The firms winning here are spending more time on workflow design than on model selection. Several have stood up a small "advisor workflow" team, usually two product people, one engineering, one compliance, whose only job is to map and reshape the end-to-end work, then decide where the agentic layer plugs in.
2 / What firms are putting into production
Across our sample, the deployments that are genuinely live (not in pilot) cluster into five categories. Pilot-only deployments cluster differently, we'll come back to that.
- Meeting capture & follow-up. Live for 22 of 28 firms. Mature, predictable, regulated where required, and now consistently saving an advisor 90–120 minutes a day.
- Document extraction (KYC, onboarding, account opening). Live for 19 of 28. The economics are clear; the variance is mostly around regional document quality.
- Internal knowledge retrieval. Live for 17 of 28. Less ROI-quantifiable, but high adoption from advisors when configured well, and high abandonment when not.
- Compliance pre-review of communications. Live for 11 of 28. The most contested category. Three firms told us this was the single largest productivity unlock; two told us it had eroded trust with compliance.
- Portfolio commentary drafting. Live for 9 of 28. Notable because it touches the client-facing surface, and the firms doing this well have invested heavily in tone-of-voice tuning.
Pilot-only categories were dominated by direct client-facing agents. Of the 28 firms, only 3 had any agentic surface that a client could actually interact with in production. The risk appetite for client-facing autonomy is, predictably, well below the risk appetite for internal-facing autonomy.
3 / The compliance layer is the operating model
Every firm in our sample that had successfully moved beyond two production workflows had done something specific: they had built, or seriously rethought, an automated review and archival layer. Not as an afterthought. Not as a vendor checkbox. As the spine of the deployment.
Two patterns dominated. The first: a per-output classification step that grades every agentic action against the firm's policy library (with the policy library itself maintained outside the model). The second: a real-time archival pipeline that captures the full chain, prompt, retrieved context, model output, action taken, into the firm's existing supervision and records infrastructure. The firms running both consistently described their compliance officers as "supportive" or "co-owners". The firms running neither described compliance as "the blocker".
The pattern that did not work, in every case we found: trusting the model vendor's safety claims and hoping for the best. Three firms had attempted this. All three had rolled back deployments.
The new role: the agent steward
Eleven of the twenty-eight firms had hired or named someone to a specific role that did not exist a year ago. Titles varied, "AI deployment lead", "agentic operations manager", "advisor-AI partnership lead", but the job was the same. Maintain the deployed agents. Catch drift. Tune retrieval. Manage the policy library. Liaise with compliance and the affected advisor pods.
The firms that had hired this role were also the firms with the highest number of production workflows. The correlation is not subtle. The capability does not maintain itself.
Three operational shifts to watch
Beyond the deployment patterns, three operating-model shifts came up repeatedly in our conversations. Each one is in progress; none is fully resolved.
- Procurement is becoming product management. The traditional procurement function is increasingly unable to evaluate models the way it can evaluate static SaaS. Firms that have figured this out have moved evaluation into a small joint product/risk function.
- The training budget is moving. Onboarding curricula are being rewritten to assume agentic tooling. New-hire ramp on agent-assisted firms is now 30–40% shorter, according to two firms that measured it carefully, but it requires a very different curriculum.
- The vendor relationship is changing shape. Firms are pushing back on closed-ended contracts that don't allow them to bring in updated models. The renegotiation cycle is shorter; the integration footprint is more conservative.
What we're watching next
The next twelve months are likely to bring two adjacent shifts. First, the client-facing surface. Three of our twenty-eight firms have a roadmap to put a properly scoped agentic capability in front of high-net-worth clients within the year. None has decided how to surface it. The design and disclosure work here is more interesting than the technical work.
Second, the consolidation of the toolkit. Several firms have realised they have four or five overlapping agent platforms in production, usually a vendor product, an internal build, a model-provider-supplied SDK, and a piece of the CRM. The simplification work is starting. The firms doing it well are reducing surface area before adding more agents, not in parallel.
If there is a single line to take from these conversations, it is this: the firms moving fastest are the ones who treated agentic AI as a new operating model rather than a new product category. The toolset is interchangeable. The discipline is not.
