Ask an established advisor what changed their practice most, and the answer is rarely a piece of software. It is usually a person: the first associate they hired.
A good associate does not replace the advisor's judgment. They make room for it. They take the work that has to be done carefully but does not require the advisor's licence, and they do it well enough that the advisor can spend their time on the conversation, the recommendation and the relationship.
That role is worth understanding in detail, because it is the most useful template we have for how AI should work inside financial planning software.
What the Associate Actually Does
The job varies by firm, but the shape is consistent. A planning associate does four kinds of work.
They gather. Before a plan can be built, someone has to collect the T4s, the notices of assessment, the CPP Statement of Contributions or QPP Statement of Participation, the pension statements, the insurance policies and, for incorporated clients, the T2 and the financial statements of each company. Then they have to read them and enter the right figure in the right place. It is unglamorous work, and errors here flow through every number downstream.
They build and check. With the data in, the associate assembles the plan in the firm's planning software and then looks for what is wrong or missing. The spouse whose RRIF has no successor annuitant. The holding company that appears in the balance sheet but not in the estate projection. The client in Ontario whose estate administration tax was never estimated.
They prepare for the meeting. They anticipate the questions the client is likely to ask and have the scenarios ready: CPP at 60 versus 65 versus 70, retiring at 62 instead of 65, selling the cottage now or leaving it in the estate.
They draft. After the meeting, they turn the decisions into a report the client can read, and they update the file so the next person who opens it knows what happened.
What the Associate Does Not Do
Just as important is what stays with the advisor.
The associate does not make the recommendation. They do not present a number to a client that the advisor has not seen. They do not change the plan without the advisor knowing. And they do not invent figures: every projection they show comes from the planning software, not from an estimate in their head.
These boundaries are not a sign of distrust. They are what makes the arrangement work. The advisor can delegate a great deal precisely because the boundaries are clear.
Why This Is the Right Model for AI
Most AI in planning software is marketed on autonomy: how much of the plan the system can produce on its own. That framing misses what makes delegation safe in the first place.
An advisor does not trust their associate because the associate is autonomous. They trust them because the work comes back for review, the numbers come from a system of record, and the file shows who did what. Remove any of those three and the advisor is no longer delegating; they are hoping.
The same three conditions are what make AI useful rather than risky.
Review before it reaches the plan. An AI associate can read a stack of statements and propose a mapping of every figure. The advisor confirms the mapping before it becomes part of the client file, the same way they would check a new associate's data entry.
Numbers from the engine, not the model. Language models are good at reading and explaining. They are not built to guarantee that the same inputs produce the same outputs. The projection itself should come from a deterministic planning engine, and the AI should cite that engine's figures rather than compute its own. We go further into why in where the numbers should come from.
A record of the work. One reviewable thread per client, showing what the AI was asked, what it proposed and what the advisor approved, does for AI what a well-kept file does for a human associate.
The Associate Across the Whole File
A useful associate is not confined to one corner of the plan. The work spans retirement, tax, estate, insurance and, for business owners, the corporation. An AI associate should be judged the same way.
Consider a typical Canadian file: a couple in their late fifties in Quebec, one spouse with a defined benefit pension, the other with an operating company and a holding company. The associate's work on that file touches nearly everything.
They need QPP, not CPP, and they need to know that QPP now allows a start date as late as 72. They need the holdco's passive income, because once it passes $50,000 a year it starts to erode the operating company's small business deduction. They need the RRIF conversion at 71 and the minimum withdrawal of 5.28% that follows. They need the OAS recovery tax, which takes back 15 cents of OAS for every dollar of net income above the threshold, because pension income, RRIF minimums and holdco dividends can stack in exactly the years the couple turns 65. And because the clients are in Quebec, the report almost certainly needs to be in French.
An AI that reviews only the retirement projection is helpful. An AI that works across the whole file, under the advisor's supervision, is closer to an associate.
What Changes for the Advisor
When the associate model works, the advisor's day shifts. Less of it goes to data entry and reconciliation. More of it goes to the questions only the advisor can answer: whether the client should defer QPP, whether the holdco should be wound down or kept for the next generation, how much risk the couple can live with.
That shift does not come from removing the advisor from the process. It comes from removing the advisor from the parts of the process that do not need them, while keeping them in control of every part that does.
It also helps the rest of the firm. An associate's thread is readable by the advisor's assistant, by a colleague covering a vacation, and by compliance. Delegation that leaves a record is delegation a team can share.
Questions to Ask of Any AI Associate
If you are evaluating AI in a planning tool, the associate model gives you a practical checklist.
Does it work for me or for the client? An associate works for the advisor. A tool that talks to clients directly is a different product with different risks.
Can it gather from real documents? Statements, tax slips, government benefit records and corporate returns are where the work starts.
Do I approve every change? Ask what happens if you ignore a suggestion. The plan should not move.
Where do its numbers come from? It should cite the engine, never estimate.
Can my team and compliance read what it did? If the answer is no, the work is not really delegated.
Where is client data stored, and is it used to train models? For a Canadian practice, this is a baseline question.
The Bottom Line
The most valuable person in many practices is the one who makes sure the advisor walks into every meeting prepared and walks out with a clean file. That is a specific, well-understood role, with clear boundaries.
AI in financial planning is most useful when it fills that role and respects those boundaries. Finn, PlanBase's financial planning associate, is built on that model: it gathers, builds, checks and drafts across the whole file, every figure comes from the deterministic engine, and the advisor approves every change before it reaches the plan.
Finn works for the advisor, never directly for the client. Client data is stored in Canada and never used to train AI models.