Published on
November 18, 2026
Before You Add an AI Agent, Decide Who Will Manage It

Family offices and RIAs need capacity—not another system to supervise.
A family office buys technology to give its team more time.
Then someone has to implement it, fix the data, configure the workflows and check the outputs.
Now add an AI agent.
It promises to coordinate systems and take repetitive work off people's desks.
But who defines its role? Who reviews the exceptions? Who notices when a change upstream makes yesterday's reliable workflow unreliable?
The uncomfortable possibility is that a tool purchased to relieve an overstretched team becomes another responsibility for that same team.
The problem is not always missing technology
In "The Bandwidth Trap," Jessica Parker describes how lean family offices can struggle to implement the technology they buy because the people responsible are already occupied keeping the office running.
Partially adopted platforms can leave teams maintaining parallel processes and returning to familiar spreadsheets.
AI does not automatically solve that problem.
It may reduce the effort required for an individual task while introducing new work around permissions, testing, supervision and maintenance.
The question is not simply whether the technology is capable.
It is whether the office has a realistic operating model for using it.
Treat it like a hire—but do not mistake it for one
At RIA Edge Austin, a panel moderated by James Cantwell recommended introducing agents through narrowly defined assignments, probation periods and regular reviews. Trevor Hicks compared early access to the limited authority a firm would give an intern.
That is sensible advice.
But I would be careful with the phrase "agentic staff."
An agent cannot accept professional responsibility or become answerable to the client. The hiring analogy is useful for defining scope and supervision—not transferring accountability.
Give an agent one assignment, one accountable human owner and clear limits. Expand its authority only when the evidence supports doing so.
"Help with operations" is not an assignment.
"Identify missing onboarding documents and prepare a draft follow-up list—without contacting clients or changing records" is.
Measure the supervision—not just the speed
A demonstration shows how quickly an agent completes a task.
A useful pilot shows what it takes to rely on it.
Before deployment, ask:
- Who reviews the output, and how long does that take?
- Can the reviewer inspect the supporting records?
- What happens when information is missing or contradictory?
- Who handles failures when the owner is unavailable?
- Who can stop the workflow?
The relevant measure is net capacity: the work removed, less the work introduced through review, maintenance and correction.
An agent that improves coverage without saving time may still be valuable. But that is a different benefit from reducing workload.
"Human in the loop" is not a complete control if the human has neither the time nor the evidence to review properly.
Imperfect data is not permission to guess
The RIA Edge discussion also challenged the idea that firms must achieve perfect data before starting. The emphasis was on bounded use cases and knowing what to trust, verify and escalate.
Do you agree?
A firm does not need to resolve every data problem before testing one useful workflow.
It does need to establish which sources are approved, what information is missing and when uncertainty requires the agent to stop.
A draft internal summary and a payment instruction should not have the same tolerance for ambiguity.
Start with a task whose limits you understand—not one whose risks you hope the model will work out.
Not every task needs an agent
A recurring calculation may be better handled by tested code.
A routine transfer may need a conventional integration.
A stalled process may need a clear owner rather than a conversational interface.
Use agentic behaviour where interpreting variable information or selecting among permitted steps adds value.
The objective should be the simplest reliable solution—not the most autonomous one.
For solo advisors and small RIAs, that distinction matters especially. They cannot afford to become the permanent support desk for their own technology.
Decide what will stop
There is one question I would add to every pilot:
If this succeeds, what existing work will we retire?
Parallel operation can be sensible during validation. Permanent duplication is different.
If the team keeps the spreadsheet, maintains the platform and checks the agent's version, it may have created a third process rather than improved the first two.
Agree the acceptance criteria for retiring the old process before the pilot begins.
Then start small enough to finish:
One workflow. One accountable owner. Defined permissions. A measurable review burden. A clear stop condition.
The goal is not to accumulate digital colleagues.
It is to give the people already responsible for the office more capacity to do their jobs well.
Sources
- Jessica Parker, Mr Family Office — "The Bandwidth Trap." Commentary on implementation capacity and technology adoption in lean family offices.
- Alex Ortolani, WealthManagement.com — "RIAs Urged to Treat AI Agents Like New Hires." Reporting from the RIA Edge Austin panel on agent scope, supervision and data readiness.