AI's five weeks come with an unquantified bill

AssetMark's latest advisor survey puts AI adoption at 85% and the recovered time at four hours a week, or five weeks a year. The industry got the number it wanted, but it's also the number that hides the variable deciding which practices keep those hours: the survey leaves the error rate unquantified. The compliance hurdle advisors keep naming as their reason for not going further is the shape of the risk the new capital is already pricing, and Pave's $15 million round treats that gap as a product rather than a footnote.

Five weeks a year is real capacity. For a solo advisor running a lean book, it's the difference between quarterly client calls and semi-annual ones; for a team, it's the slack that lets a senior planner sit in on a decumulation meeting without surrendering the email queue. But capacity is only worth the trust behind the output, and a survey that measures hours saved does not attempt to measure the rate at which those outputs are wrong. That omission is the denominator of the entire practice economics, the number every compliance officer will eventually have to put in front of an examiner.

Advisors who have gone further with AI keep naming compliance as the wall, usually some version of an inability to document the decision trail or attest to the output. That is an accurate description of how an RIA's errors are priced: a recommendation that cannot be reconstructed under exam is a finding waiting for its date. A firm that cannot state its error rate cannot underwrite its own AI exposure, and one that can't underwrite its own exposure is already paying a premium it hasn't measured.

Compliance friction shows up as slower adoption for larger books, because an error on a larger book is more expensive than an error on a smaller one, and the survey's 85% likely includes many firms that use AI for drafting but stop short of client-facing output. The compliance wall sits at the final sign-off, long after the first prompt. The firms with the most to gain from AI hours are the ones least able to use them without a validation layer.

Pave's $15 million round lands at exactly this seam: the advisory firms that wrote checks were not renting another seat license but buying a cheaper portfolio-construction layer and a position in the channel that sells it. The custodian layer is where an error rate becomes either an owned risk or a rented one. A firm that owns infrastructure sitting between the portfolio and the custodian can set the validation rules, capture the exception log, and prove to a regulator which outputs were reviewed. A firm that rents a seat inherits whatever error rate the vendor chooses to publish — and the AssetMark survey, for all its adoption numbers, does not publish one.

Schwab's $240-per-seat Claude integration shows the rental side of that equation. The custodian has turned AI into a line item, and the process work around it will cost more than the license, because the tool still requires a human review layer. The $240 is the cover charge; the bill for validation, exception handling, and the occasional bad output lands on the firm's P&L whether or not the vendor quantifies it. The difference between $240 a seat and $15 million of equity is the difference between renting a room in someone else's compliance framework and owning the building where the rules get written.

The presence of advisory firms on a vendor's cap table changes the governance dynamic. A seat license gives a firm no say in the product roadmap; an equity check gives it a reason to demand the error-rate data the industry has not yet seen. The advisors who wrote Pave's checks bought a cheaper portfolio-construction stack and the right to ask the question the AssetMark survey could not answer.

The unquantified error rate is now the unmodeled risk in the advisory P&L. If an advisor recovers five weeks but spends one of them correcting AI outputs the client never sees, the net is four weeks and a hidden liability. If the error rate compounds across 85% adoption, the industry's collective hours saved may be smaller than the hours spent defending outputs in a compliance review. The survey that publishes the error rate will do more for practice management than any adoption number, because it will finally let firms price the hours they are being told they have saved.

For the advisors who did not write a check in the Pave round, the implication is simpler. The hours saved are already in the P&L, but the error rate is not, and every quarter that gap persists is a quarter in which the practice is carrying an unhedged operational short. The firms that move first to define their own error rate will have a disclosure advantage in due diligence that the rest cannot match.

None of this is an argument against AI. The five weeks are real, and the firms using it are not wrong to bank them. The next split in the industry will run between firms that can assure an AI output and firms that cannot. The firms that own or control the validation layer will be able to tell a client, a regulator, and a buyer exactly what their error rate is. The firms that rent seats will be able to tell them what their vendor says, and the difference will show up in valuations, exams, and recruiting. A buyer of an RIA will ask for the error rate on AI-assisted recommendations the same way it asks for revenue concentration, and the seller that has no answer will take the discount.

Pave's $15 million is a small number for the wealth industry, but it is a specific bet that the custodian layer is where AI risk gets priced. The advisory firms that wrote checks now have both a cheaper portfolio-construction stack and a reason to insist on published error rates. The five weeks were the easy part; the next survey that publishes an error rate will set the price of AI hours for every firm that has already banked five weeks.