AI in Finance — Deployment Digest – Week of 2 September 2026

Twenty-two items from the monitoring run of 1 September 2026. The through-line this week is not a single event but the texture of a saturated field — six advisor-tech launches in five days — and the turn that saturation forces: when every tool has agents, the question is no longer whether a tool has them but what they may do without a person.

The price of admission

The week's most useful sentence is not in a product launch. It is a line of trade commentary the run picked up: "AI is no longer the differentiator; it is the price of admission." Read the twenty-two items against it and they arrange themselves. The contest has moved. A year ago a firm shipped "AI" and that was the news; this week six advisor-tech tools launched in five days, all of them agentic, and having agents is simply the cost of being in the room. What separates one item from the next is no longer presence but permission — how far the agent is allowed to go before a person is asked. That is the axis this digest reads. It runs cleanly this week from one end, where the permission is named and mostly held, to the other, where it is left unspoken.

The boundary named, and mostly held

Scalable Capital (accounts wired to external AI, a gate at every trade) — The Munich digital bank and broker — over a million clients, some €60bn on the platform — let customers connect their accounts to external, general-purpose AI assistants (Claude, ChatGPT, Grok) over an MCP server, to place trades, manage savings plans, run portfolio analysis and pull their own data. The press did the expected thing; a Fortune headline read "a major German bank just let Claude and ChatGPT trade for customers." The mechanism underneath the headline is narrower and more interesting: a per-action human approval gate, the customer approving each trade before it executes. This is the same downward pressure toward money-movement the log has tracked all summer, but with the opposite decision about where the gate sits — the mirror image of eToro's authorise-once grant (Week of 18 August). eToro moved the consent up front, once, at the grant of authority; Scalable keeps it at each trade. The design choice is the story.

Confirmed (launch). The per-trade approval gate is described in the bank's own materials, not independently demonstrated, and the run flags that its friction is minimal — a client can rubber-stamp an AI-proposed trade as fast as reading it. What the gate protects against is a bad trade reaching the market unseen; what it does not touch is the deciding beneath the trade.

There is a quieter thing to notice here, and it is the one worth carrying. The intelligence doing the work is not the bank's and not the client's — it is a general-purpose consumer model, reaching through the bank's platform, under the bank's gate, into the client's account. It acts on the customer's behalf and moves, the whole time, inside the institution's systems. Acting on your behalf and being on your side are not the same thing; the model wired into the trading platform is neither the bank's instrument nor yours, and the only intelligence in the chain formed from the client — with no institution behind it, reaching into no one's systems — is not the one placing the trade. We leave that where it sits for now; the positioning note is where it belongs at length.

Advisor360° ("the Autonomous Financial Platform," every action restated) — Advisor360° shipped an August release under a brand that states the maximal claim — "the Autonomous Financial Platform" — carrying a Client Intelligence agent that summarises, flags and suggests next steps; VIDA agents that answer plain-language questions and "take action on the advisor's behalf, with every action restated for the advisor to confirm"; an "agentic" Meeting Prep agent it credits with cutting prep time some 60%; native Schwab account-opening in pilot; and Fidelity straight-through trade processing that can be held for review. The brand register and the mechanism pull opposite ways in the same release — "autonomous" on the marquee, "restated … to confirm" in the body — and the detail worth keeping is the trade-review hold: it is configurable by the home office, not on by default. Governance available, not guaranteed — the same shape as nCino's Mortgage-MCP safeguard-as-a-setting (Week of 18 August).

Confirmed (release), several features pilot-stage. The −60% prep-time figure is self-reported; the span of a VIDA "action" — a data edit versus something that touches a client — is unstated; and the release does not give the default state of the Fidelity review hold if an administrator never sets it, which is precisely the state most firms will run in.

Repodo (the audit automated, the signature left human) — Repodo, founded by Lunar alumni on an €8.2m pre-seed, calls itself "an operating system for financial auditing" and "the last major professional service the modern world forgot to rebuild," running agentic AI across the mechanical span of an audit — collect, reconcile, document, analyse — with, in the same breath, "final sign-off left to human judgement." It is the audit line the log has held since the June baseline: agents may draft the whole of the work, but the opinion is signed by a named human, because that is the one point where the line legally cannot move.

Confirmed (launch / funding) as to the raise. Pre-seed; nothing is demonstrated. And "final sign-off" is doing quiet work — the interesting judgements in an audit (materiality, what to sample, how to resolve an exception) are made long before the signature, and whether those sit inside the automated span or outside it is not stated.

Three items, one shape: in each the marketed verb outruns the mechanism, and in each the mechanism turns out to be a stated human gate. That is the strong half of the week — not because the tools are timid, but because the boundary is spoken and can therefore be checked.

Against the grain: the agent that sells the brakes

Caddi (an agent that builds the agents, marketing restraint) — Caddi launched an agent that builds and governs a firm's other back-office agents: it reads the existing systems, ranks a firm's processes by frequency and cost, and turns a demonstrated task into agent rules in what it calls "Loop Studio." What sets it against the week's grain is the pitch. Where Practifi, Marloo and Growhill (below) market more autonomy and say nothing of the human, Caddi markets less — it "call[s] AI models only where judgment is required and run[s] deterministic code elsewhere," and keeps "a replayable record of each execution," explicitly to stop model error compounding (it cites 95%-per-step accuracy falling to about 75% over six steps if a model decides every step). It rhymes with last week's Playbook — both are agents that build agents, the autonomy line moving up into meta-orchestration rather than down toward the account — but Playbook sold the autonomy and Caddi sells the brakes.

Confirmed (launch). SOC 2 Type II, 100+ connectors and the named customer outcomes are company-cited, not independently verified. The edge the pitch leaves unexamined: who reviews an agent that Caddi's agent has built before it runs against live client data, and what "governs" enforces versus merely records. Deterministic-where-possible narrows the surface where a model can err; it does not answer for the steps where, by Caddi's own account, a model still decides.

The boundary left unspoken

Practifi, Marloo, Growhill (capability named, the human step not) — The other half of the six-in-five-days runs the same register with the gate left out. Practifi launched "Sentir," an "AI-native intelligent CRM" it describes as carrying sixteen named agents spanning tax, planning and client sentiment — the inventory stated, which agents act versus surface, and where an adviser confirms, not. Marloo entered the US framed — and URL-slugged — as "designed to replace traditional financial advice workflows," automating meeting prep, advice documents and compliance "rather than just recording them"; the listed functions read as preparatory and imply a review, but the marketed verb is replace, and no confirmation gate is named. Growhill Wealth launched an "agentic AI workforce" that monitors portfolios, flags issues and preps research for independent advisers across Asia — functions that are surfacing by nature, under a label ("workforce") stronger than the mechanics. The common thread is not that these tools are more autonomous than the ones above; it is that their autonomy is unspecified, and in Marloo's and Practifi's case the unspecified span reaches into compliance and tax — the domains where the review question bites hardest.

Each Confirmed (launch) as to the launch; capability asserted, nothing demonstrated. Marloo's 900-firm / 8-country reach and Practifi's agent count are self-reported. The gap in all three is not a claim of unattended execution — it is the absence of any statement either way.

Set against them, and worth naming for the contrast, are the run's clearly-assistive items, where the human is spoken plainly: DBS Treasures paired AI onboarding with more than 600 new hires; april surfaced IRS-transcript-based tax-planning insights to the adviser, consent-gated by the client's own authorisation; Decade framed its raise around "custom financial AI models with human advisors." The register these use — augment, with the person named — is the one the mid-tier launches drop.

The consumer assistant reaches regulated ground

Flanks / Perplexity (regulated holdings inside a general-purpose assistant) — Slightly to the side of the autonomy axis, and pairing almost too neatly with Scalable, Flanks agreed to pipe regulated portfolio data from 700+ institutions across 33 countries into Perplexity's connector ecosystem, putting that data inside natural-language advisor workflows. Where Scalable lets an external consumer model reach into a regulated platform, this moves regulated holdings out into a consumer-grade assistant. The worry it raises is not autonomy but governance: regulated portfolio data sitting inside a general AI assistant, with permissioning, and whether the surfaced output is advice or information, unaddressed. It is the same underlying question from the other direction — what happens when the intelligence touching a client's regulated financial life is a general-purpose model that belongs to neither the client nor their institution.

Confirmed (partnership); a data-connector layer, no execution claim. Whether "regulated data" carries its permissioning into the assistant, and who is accountable if the assistant's output is read as advice, are not stated.

Capital in the plumbing

Underneath the launches, an unusually concentrated week of capital and consolidation at the layer the agents will eventually route through — and, tellingly, none of it carrying a specific agentic claim.

Vanguard / Altruist — Vanguard agreed to acquire Altruist, the "AI-forward" self-clearing custody platform serving 6,000+ independent advisers, in a deal reported around $4bn (the WSJ's ~$4bn and the aggregator's ~$4.6bn differ). Custody and clearing — the rails, not an agent. Confirmed (announced) / Reported (price). RQD* Clearing took a $74m growth investment led by Bain Capital — clearing capitalised, again the substrate. Confirmed (funding). Deutsche Bank selected Thought Machine's Vault Core on a ten-year term to re-platform its Private Bank, around €600m and carrying no AI claim at all — which is exactly why it belongs here: it is the clean-core-data precondition the log keeps naming, the condition an agent needs before it can be trusted with anything. Confirmed (deal), no AI dimension. Alongside it, FE fundinfo launched a "Product Mastering Core," a productised, "governed" fund-data layer it pitches as "the on-ramp for everything else in the stack" — the same precondition sold as a product. Confirmed (launch); the 64% efficiency figure is cited by the interested party, and whether "governed" is enforced or merely descriptive is unaddressed.

The register across this row is "AI-forward," "technology platform," "governed"; the deals themselves demonstrate no agentic capability. What they demonstrate is where the money is actually going — into the custody, clearing and clean-data substrate beneath the tooling, on the wager that whoever owns the rails and the structured data owns the ground the agents will stand on.

Intent still far ahead of value — now in the CFO office

The standing gap between intent and realisation surfaced again this run, this time in the CFO's own consultancy figures. Gartner projects roughly ten margin points by 2029 from AI while noting "moderate current value and early adoption" — 84% of finance functions implementing or planning AI, only 7% reporting high or very-high impact. Deloitte puts 63% "fully deployed" against 21% seeing tangible value and 14% with agents fully integrated. Near-universal intent, thin realised integration — the KPMG 99%-plan / 11%-deployed shape (Weeks of 18 and 25 August), now one storey up in the office of the CFO.

Reported (survey) / Estimate (projections). The consultancies sell AI-transformation and governance advisory; the 2029 margin number is a projection, and "autonomous finance" is left undefined per respondent — assistive or executing, counted the same.

What the run did not carry

The silences set the week's shape as much as the launches. No downward move of the execution line at the money-movement end of the eToro authorise-once kind: this run's money-movement item (Scalable) kept the gate at per-trade confirmation — the opposite choice — and Advisor360° restated each action; no new one-time-authority grant appeared. No net-new family-office-specific agentic launch of the FamilyOfficer.com kind: the family-office action sat, again, one layer down, in a funded platform play (Nexedge's raise to run "a family's whole balance sheet") and a data layer (FE fundinfo), not in named agents. No binding AI regulatory rule: the only rulemaking this run was an SEC proposal on crypto-asset exemptions carrying no AI-specific provision, so the governance edge stayed on paper — CFO survey gaps and, elsewhere in the run, a forthcoming external "autonomy" yardstick, described but unpublished. The personnel register — "AI Teammate," "digital employees" (Citi Sky, Wells Fargo, BNY) — stayed comparatively quiet; the closest tell was Growhill's "AI workforce," while the dominant vendor register shifted to "autonomous platform / autonomous investing" and, counter-current, Caddi's restraint. And no macro or financial-stability AI-bubble item of the BIS / FSB / IMF kind; the nearest thing to a systemic number was the CFO margin projection, a consultancy estimate, not a warning.

Regrouped in, not absent: the regulatory edge did not vanish so much as thin to survey-and-directory form — from prior runs' SEC "AI-washing" test and the EU AI Act's Article 50 disclosure edge to this run's CFO intent-versus-value figures. The data-substrate precondition recurred with new instruments — from Astraeus's ontology to this run's governed fund-data layer (FE fundinfo), a ten-year core re-platform (Deutsche Bank) and a consumer-assistant data connector (Flanks / Perplexity).

Assembled from a structured weekly monitoring run of 1 September 2026, with the sourcing composition noted at the top. The run's data connectors were unauthorised and unused this week, so every thread rests on web search and primary and trade sources; every capability here is asserted by an interested party or self-reported, and every safeguard — Scalable's approval gate, Advisor360°'s confirm-and-hold, Caddi's replayable record, Repodo's human sign-off — is described rather than independently demonstrated in this run's sources. Items reflect what surfaced in monitoring during the week; where an item describes an earlier event, its date is given inline.

SMA Crown Confidential

Digital Confidantes: Bespoke AI Intelligence for Private Decision-Makers