Same question, two answers
Two wealth firms of similar size run the same AI model. Both adopted it this year, through channels that did not exist eighteen months ago, and both spent months connecting it to the systems where their clients live. On a Tuesday morning the head of advisory at each firm types the same request: which clients could meet a $2 million capital call next month without selling anything at a loss?
At the first firm the answer arrives in seconds. It is a clean table of fourteen names, each with an amount of available cash, a liquid reserve and a short note on suitability, and it reads like the work of a careful analyst. At the second firm the answer takes a little longer and lists nine names. Beside every figure sits the date it was measured and the engine that calculated it, and two families are absent because the person asking has no permission to see one of their structures.
On a screen the two answers look alike, and only one of them would survive contact with an auditor, a lender or the client. The first table counts custodial cash already promised to other capital calls, treats a gated evergreen fund as money on hand, values a building at an appraisal from last spring, leaves out a loan recorded in another system and proposes action to a client whose risk profile expired in July. None of these errors shows in the answer. Each was made by a model doing exactly what it was asked, with exactly the information it was given.
The firms are hypothetical. The moment is real. Over the past year, AI providers and the platforms advisors already use have begun packaging AI for wealth firms, and adoption has followed quickly. A study of 68 RIA firms by Cerulli Associates and Vista Equity Partners, published in September, found that 64% already see AI reducing manual and administrative work and that AI-specific budgets are set to nearly double by the end of 2026. Deloitte, drawing on an industry survey, puts the share of advisory firms using AI in some form at 73%, against 6% using agentic tools and 5% that have integrated AI across their systems (Deloitte Insights).
Those figures describe an industry that has decided to adopt AI and is still deciding what to trust it with. The distance between the two firms in the scene lives inside that hesitation, and the model has almost nothing to do with it, since the model is the same. What separates them is two things that only work together. The first is a core that holds the whole client, front to back: positions and transactions from every custodian, cash, mandates and rules, suitability profiles, plans, fees and reports, and also much of what most platforms leave out, to name only a few: the ownership structures through which a family actually holds its wealth, the assets that sit with no custodian at all, commitments and liabilities, the date behind every valuation, the tax residency of each member, the heirs, the consents and the permissions. The second is knowing how to read that core and what to take from it for each question, a capacity the model acquires only through teaching aimed entirely at that core and at the way that particular firm works, with its own policies, products and clients, and its own definition of what counts as free cash or a suitable proposal. A complete core read by an untaught model is read badly, and a well-taught model working on an incomplete core explains, very fluently, a picture with pieces missing. Both take far more work than the announcements suggest, and both depend on the permanent attention of the people who built the core, wired the AI to it and wrote the teaching.
What travels through a connector
Connectors have made the plumbing easier, though by less than the announcements suggest. Where a custodian or software vendor publishes one, a model can sign in, see what is there and read it, and common standards such as the Model Context Protocol spare each firm from writing that integration itself. What remains is still serious work. Each connector has to be authorised firm by firm and user by user. It exposes what its vendor chose to expose, often for reading only. A firm with three custodians receives three versions of the same position and still has to reconcile them, and nothing goes live before security and compliance have agreed on which data may leave which system. Deloitte's figure from the opening, 5% of advisory firms with AI integrated across their systems, measures how far most of the industry still is from the end of that road. And at the end of it, what travels down the pipe is a set of fields. A position has a quantity, a price and a date. A client has a name, a risk score and an advisor. A cash balance has a currency and an amount.
Everything a decision depends on lives around those fields, in rules the source system applies and seldom exposes. The price on a property company may be a market close or an appraisal formed nine months earlier. A cash balance may be free or already earmarked for a commitment that only the fund administrator has recorded, the gap at the centre of The Asset Outside the Mandate. "Quarterly liquidity" on an evergreen fund means little in a year when the fund is paying out a fraction of every request. The person on the account may still sign while authority has passed to an attorney-in-fact and the heir lives abroad (The Heirs Are Not on the Platform). A risk profile may have been contradicted by the client's own transactions since the day it was signed. Net worth on the screen may leave out the loan secured on the family's buildings (The Other Side of the Balance Sheet).
A model reading those fields has to infer what they mean from names, formats and context, and inference is what it does best, which is the source of the trouble. A field called market_value will be read as a market value, and a balance called available will be read as available. The answer that comes back is consistent with itself, properly formatted and written in the assured voice of a senior analyst, while the error sits where no reader can see it, in the space between what the field says and what the source system knows about it.
Those errors compound. A simple question touches one domain, such as a client's equity allocation. The questions firms most want an AI to answer cross many at once: positions and transactions across custodians, cash and liquidity, commitments and liabilities, valuations and the dates behind them, mandates and rules, suitability, tax, fees, ownership structures, authority and permissions, among others. If every domain an answer passes through carries even a modest chance of being misread, the chance that every part of the answer is right falls quickly as the domains add up. The chart uses assumed error rates, chosen only to make the shape visible, and the shape holds for any rates one picks as long as one layer misreads more often than the other.
The capital-call question from the opening scene crosses seven domains. A small chance of misreading each field, repeated across every domain an answer touches, leaves a generic layer right on only part of its answers, while a complete core read by a layer taught to understand it keeps the accumulated error low. The distance between the two lines grows with every domain added, so it is widest on exactly the questions a firm most wants answered, and on those questions the wrong answers look the most finished.
Who knows what the field means
Connectors have a real place: common protocols lower the cost of giving a model access to a system that once needed a bespoke integration, and for a firm running many separate systems, an assistant that reaches across them is a reasonable first step. A first step across a fragmented stack leaves the fragmentation where it was, though. Every seam between two systems is a place where meaning is lost, and an assistant laid across those seams inherits every one of them, now answering in fluent prose.
What it leaves open is meaning. A protocol settles how a model calls a tool. It is silent on what the tool should return, which rules it must respect and what the model should be kept from computing on its own. Somebody has to decide that net worth comes from the consolidation engine and never from the model's arithmetic, that a private valuation always travels with its date and basis, that an evergreen holding is reported with its redemption terms and the proration actually being paid, that a permission is enforced before data reaches the model and well before an answer is drawn, and that a proposed trade is run against the mandate rules before anyone is asked to approve it. Every one of those decisions depends on knowing how the core reconciles, what it treats as authoritative when two sources disagree, where its edge cases live and what changed in the last release.
The team that designed the data model already knows all of this, because it wrote the rules. It can give the AI tools that call the platform's own calculation and rule engines, so that every figure in an answer is one the firm could reproduce line by line in an audit. It can apply to the AI the same permission model that already governs every screen and every report. It can test the AI against the scenarios the core was built to handle, and run those tests each time the core changes, inside the same release cycle. A third-party connector sees the core from the outside. It is usually accurate on the day it is written, and it begins to drift the first time a field is added, a calculation refined or a rule tightened, with nobody on either side whose job it is to notice.
None of this happens by itself, and it is worth being precise about what the model needs. A frontier model does not have to be retrained to work in wealth management, and an offering packaged for advisors already arrives with connectors and workflow skills that teach it the general shape of the profession. What no general package can carry is the meaning of one particular core: which of its numbers are final and which are estimates, which engine owns each of them and which rules bind them, down to the firm's own policies on which products it will not offer to which profiles. That knowledge has to be written into the layer, in the instructions the model follows, in the plain-language descriptions attached to each tool, which are what the model reads to decide what to call and with which parameters, and in the tests that check whether it chose well. Much of the craft of building an AI layer goes into this teaching, and it is done well only by people who know the core from the inside.
Teaching has a limit, and the limit is the core itself. A model taught perfectly still answers from what the platform holds, so a core that misses the assets held outside any custodian, cannot say which of three custodian feeds is right or keeps each family member's tax residency in a spreadsheet will produce well-taught answers about an incomplete picture. The two conditions travel together: a core complete and stable enough to describe clients as they actually are, and a layer taught by the people who built it. A firm with only the second has an articulate guide to data it does not have. A firm with only the first has accurate books that the model will misread.
Retrieval over the firm's documents narrows the gap less than it seems to. It works well on text: an investment policy, a meeting note, a prospectus, a side letter. A reconciled position, a time-weighted return or a fee accrual exists in no document until an engine calculates it, so a model asked for one without a tool that computes it will assemble a plausible figure from whatever fragments it can find. The result is close often enough to earn trust and wrong often enough to matter, which is the most expensive combination there is.
The worry about lock-in deserves a direct answer too. A firm protects its independence by owning its data in a form it can take elsewhere and by keeping the freedom to choose which model sits on top. Both are achievable on a single core. What gains a firm nothing is having the meaning of its business rebuilt, slightly differently and slightly wrongly, inside every integration it runs.
Errors that persuade
Before AI, a fragmented stack announced itself through friction. An analyst reconciled two reports by hand, an advisor phoned operations to ask which balance was right, a number failed to tie and was therefore checked. That friction was where a great many errors were caught. A generic model laid across the same systems hides it. The inconsistency is still there, beneath an answer that arrives looking reconciled, and it never reaches the one person who might have spotted it.
Scale turns this from an inconvenience into a property of the firm. A blind spot in a platform used to cost money one client at a time, whenever someone happened to act on it. Wired to an AI that every advisor consults all day, the same blind spot is repeated in thousands of answers. A firm that never mapped which entity owns which account will be told, each time it asks, that a family is well diversified when most of its wealth sits in a single holding company. A firm that carries appraisals as prices will see a calm property allocation in every year except the one that matters.
The industry is also adopting AI for a particular reason, and that reason sharpens the problem. McKinsey estimates that at current productivity the US industry could be short of 90,000 to 110,000 advisors by 2034, and that some 110,000 of today's advisors, 38% of the total and responsible for 42% of the industry's assets, will retire within a decade. Firms are counting on AI to fill that gap. Much of what leaves with those advisors was never written into any platform: which sibling really makes the decisions, why a position was never sold, what was promised about the family company. A model can only reason over what the platform holds, so the firms leaning hardest on AI to replace departing capacity are often relying on it for the very knowledge it will never be able to see.
The standard safeguard is human review, and the people who study AI governance in this industry are clear about what it needs. Deloitte's 2026 investment management outlook lists clear data lineage, defined thresholds for human-in-the-loop review and audit trails among the artifacts firms should have in place, and its May 2026 analysis of agentic AI in wealth management observes that efforts to scale tend to stall on the basics: scattered content, access rights nobody has defined and thin testing. Review protects a firm only to the extent that the person reviewing can check what is in front of them. A reviewer shown a fluent table has no means of knowing which cash was committed, which value was stale or which structure was quietly filtered out, unless the answer carries its own lineage: the source, the date, the basis and the engine behind every figure. That lineage can only be supplied by a layer that knows where each number was born.
Taken together, these demands describe an AI layer built as part of the core. It works from a semantic model of the business, in which a valuation, a commitment, a liability, a role and a permission are objects in their own right, each with dates and rules attached. It hands every calculation to the engines the firm would rely on without AI, so the model explains numbers and has no occasion to invent them. It applies permissions inside the data, before anything reaches the model. It can show where each figure in any answer came from. It is tested against the firm's real scenarios every time the core changes. And every action it proposes, a trade, a rebalance, a client email, waits for a person to confirm it, with the rules checked before the confirmation is even requested.
None of this is finished on the day it goes live. Model providers release new versions regularly, and the same question can come back with a different answer after an update. The core moves as well, with new fields, new products, new rules and new custodian feeds, and the client base moves with it, bringing cases the layer was never shown. A layer that was right in March can be quietly wrong by September without anyone having touched it. Keeping it right is permanent work: watching which tools the model actually calls and what it answers, re-running the firm's scenarios after every change on either side, rewriting the teaching whenever the core changes, and catching drift before an advisor acts on it. That work belongs with the people who built the core, made the connections and wrote the teaching, because they are the only ones who can tell whether an answer that changed has become better or has become wrong.
How Pivolt approaches it
Pivolt was built on one data model, front to back. Clients, portfolios, plans, trades, rules, fees, reports and custodian data live in the same model, alongside the family and holding structures through which wealth is actually owned and the private assets whose valuations keep their dates and their basis. Pivolt's AI Advisor works across that model directly, without a layer of connectors in between. Because the same team built the core and the AI layer, the layer knows what each field means and which engine owns each number, and it changes in step with the core every time the core changes. The team that wired the AI to the core and wrote its teaching is also the team that watches how it behaves, release after release, so the firm never has to assemble that responsibility from several vendors who each see only their own piece.
Three guarantees hold in every answer. Every number the platform calculates comes from Pivolt's deterministic engines, never from the language model. Every action waits for a person's confirmation, with the rule engine checking mandate limits before and after. Every answer stays within the permissions and approvals in force for the person asking. The ambition is deliberately modest: an AI whose answers a firm can reproduce, explain and defend, because they were produced by the same machinery the firm already trusts for its reports and its invoices.
On a platform built this way, the head of advisory in the opening scene asks the same question and gets nine names instead of fourteen. The other five drop out for plain reasons: three only looked able to meet the $2 million capital call until every figure was checked against the rest of their position, and two belong to structures the person asking is not allowed to see. Each of the nine comes with the numbers behind it and the source of each one, so the advisor can confirm the answer before calling the client.
What separated the two firms in the opening scene sat underneath the model: a core that holds the whole client, teaching written for that core and for that firm, and people who keep checking the answers long after the launch. That work is slow, it does not show in a demonstration, and it is the part of an AI project that no competitor can buy ready-made.
Sources
- Cerulli Associates, Advisor Headcount Set to Grow as AI Expands Capacity, 9 September 2026
- Cerulli Associates and Vista Equity Partners, State of Wealth Management AI Adoption, September 2026
- Deloitte Insights, Jeffrey A. Levi and Samia Hazuria, The agentic AI productivity wave is heading for wealth management, May 2026 (adoption figures cited from Orion's 2026 Advisor Wealthtech Survey)
- Deloitte Insights, 2026 investment management outlook
- McKinsey & Company, The looming advisor shortage in US wealth management, February 2025
- The two firms, the capital-call question and the error rates in the chart are hypothetical and were constructed for this article.
