In May 2026 Ken Griffin, the CEO of Citadel, said: “For the first time, AI is real. These are extraordinarily high skilled jobs being automated by agentic AI.”[1] He wasn’t talking about call-centre scripts or data entry. He meant professional judgement, the kind of work IFAs, paraplanners and wealth managers spend careers building.
Agents are already changing how that work gets done. What a firm controls is whether it manages the change on purpose or stumbles into it.
What has changed in the past twelve months
AI tools in financial services used to be copilots: they suggested, you decided. Agents don’t wait for approval at each step. Coding, workflow and research agents now carry out a sequence of tasks on their own, report back, and in some set-ups hand parallel pieces of work to sub-agents.[2]
The infrastructure is consolidating around orchestration. Asana’s acquisition of StackAI is one sign of it.[3] The products arriving at your door assume people will be managing several agents at once, not using a single assistant.
That matters in an advice firm because the tasks agents take on (research, drafting, compliance checks, pulling client data together) sit close to regulated outputs, and most firms’ governance was written for people doing the work.
The human sets the brief and signs off the result
The simplest way to picture where people fit: the human gives the agent its context and constraints at the start, and checks and approves the output at the end. The agent does the execution in between. Some people call this the “agentic human sandwich”.
In a firm that runs agents, the two management skills that count are writing a precise brief and reviewing the output properly.
Both are harder than they sound. Writing precise agent instructions (described as “prompting less and steering more”[4]) can be learned. A review process that catches agent errors before they reach a client has to be designed.
An agent that drafts a suitability letter still needs a person to read it and take responsibility for it. An agent that screens a client document for AML flags still needs a person to make the call. In a regulated firm the review step is where your professional liability sits, so it can’t be cosmetic.
What changes about management when agents do the work
Firms are redesigning whole workflows around AI rather than adding tools to old ones. Gadoci Consulting puts 70% of enterprise AI adoption in 2026 in that category[5]. For a firm with advisers, paraplanners and admin staff, that redesign changes how you deploy people as much as software. Three things shift.
The bottleneck moves to review. In a human workflow the limit is capacity: how many people you have and how fast they work. As agents get faster, the limit becomes how quickly people can approve, steer and correct their output.[6] A paraplanner who spent six hours drafting a report might spend ninety minutes reviewing one an agent produced. That’s a real gain, but it asks the paraplanner to exercise judgement the whole way through, not only at the end.
Roles are defined by outcomes rather than tasks. If the agent executes, the person’s job is to define a good outcome, set the agent’s constraints and judge whether the output meets the standard. That’s closer to a principal’s job than a practitioner’s. Not everyone makes that move easily, and helping them is a leadership problem more than a technology one.
Oversight has to be active. Managing agents means reading logs, checking outputs against what you expected and catching drift when results start to go subtly wrong. VentureBeat reported that 54% of enterprises deploying agents have already had a security or reliability incident, and most still run their agents the same way.[7] Knowing the risk and not changing anything is a management failure, not a technical one.
Five controls to put in place now
Firms running agents in client-adjacent work need governance now, not on a five-year plan. This is a practical starting point.
1. List every agent or automation touching client data or regulated output. It doesn’t need to be a formal project, but it does need to be complete. If an agent drafts documents, runs compliance checks or touches client records, it goes on the list. You can’t govern what you don’t know is running. (Our governance pack checklist covers what else belongs alongside it.)
2. Write down the review step for each one. Who reads the output before it’s used, what are they checking, and what are they allowed to approve? “Whoever’s around” or “the person who set it up” is an assumption, not a review process.
3. Test against edge cases, not just typical inputs. Agents tend to do well on common cases and fail unexpectedly on unusual ones, and in financial services the unusual cases are often where errors matter most. Shadow testing, running the agent alongside your existing process before you rely on it, is the practical control.[8]
4. Set access limits before you scale. An agent that needs to read client files shouldn’t be able to write to your CRM or reach credentials the task doesn’t need. Give each agent exactly the permissions its task requires. This is where most firms are exposed today, and it’s cheap to fix.
5. Name who is accountable. Under SM&CR someone in your firm is responsible for the systems and controls around regulated activity, and that covers automated systems. Each significant agent deployment should have a named owner who will answer for it. Not the IT team, and not the vendor.
The shift in how senior people see their role
The harder change is how experienced people think about what they’re for. Griffin’s point was that the jobs now being automated took years to learn. Someone whose professional identity is built on building cashflow models, drafting research notes or reviewing compliance checklists will find an agent can produce a credible version in minutes.
What’s left for the person is judging whether the work is right, the relationship with the client who receives it, and the professional accountability no agent can carry. None of that is small. People still need help seeing their job that way, though, and it falls to the firm’s leaders to say so out loud and give them time to adjust. Start with the list in control 1: it’s the quickest way to see how much of your firm’s work is already being done this way.
This article is for informational purposes only and does not constitute regulated financial advice or a compliance opinion. Consult a qualified compliance professional for advice specific to your firm.
Sources
[1] AI Daily Brief, reporting Ken Griffin (CEO, Citadel), public remarks on agentic AI, May 2026. [Cited for Griffin quote on agentic AI and high-skilled job automation.]
[2] van Riel, Z., ‘AI Coding Agents Tutorial’, Zen van Riel, AI Engineer Blog, 2026. Available at: https://zenvanriel.com/ai-engineer-blog/ai-coding-agents-tutorial. [Cited for shift from copilot/autocomplete to autonomous task-based execution.]
[3] TechCrunch, ‘Asana acquires no-code agent builder StackAI’, 28 May 2026. Available at: https://techcrunch.com/2026/05/28/asana-acquires-no-code-agent-builder-stackai. [Cited for consolidation of agent infrastructure around orchestration tooling.]
[4] Everyday AI, commentary on shift from prompt-box interaction to agent consumption, June 2026. [Cited for “prompting less, steering more” framing; no primary URL available.]
[5] Gadoci Consulting, ‘What We’ve Learned About AI and the Enterprise as of Today, June 5 2026’, 2026. Available at: https://gadociconsulting.com/articles/what-we-ve-learned-about-ai-and-the-enterprise-as-of-today-june-5-2026. Vendor-sourced. [Cited for 70% of enterprise AI adoption moving toward integrated workflow redesigns.]
[6] AI Daily Brief, framework note on human review cycle as production constraint in agentic workflows, May 2026. [Cited for observation that human approval and steering becomes the rate-limiting step as agents accelerate.]
[7] VentureBeat, ‘The agent security gap: 54% of enterprises have already had an AI agent incident, and most still let agents share credentials’, 18 July 2026. Available at: https://venturebeat.com/ai/the-agent-security-gap-54-of-enterprises-have-already-had-an-ai-agent-incident-and-most-still-let-agents-share-credentials. [Cited for 54% enterprise AI agent incident statistic.]
[8] van Riel, Z., ‘Deploy Production AI in 2026, Cut Errors by 50% Fast’, Zen van Riel, AI Engineer Blog, 18 July 2026. Available at: https://zenvanriel.com/ai-engineer-blog/deploy-production-ai-2026-cut-errors-50-percent. [Cited for shadow testing as a pre-deployment validation control.]