Karsten Wenzlaff, Advisor
August 26th, 2025
Karsten Wenzlaff is a Crowdfunding and Fintech Expert from Germany. He is Secretary-General of Digit[...]
Do Autonomous Financial Agents Need A Different AI Architecture?
September 25, 2026 | NCFA Fintech Intelligence Question | Artificial Intelligence And Data, Risk Compliance And Regtech, Payments Infrastructure And Money Movement
When An LLM Understands The Request, Should It Also Make The Decision?
Last Updated: September 25, 2026
Status: Emerging
Organizations: TypeSafe AI, International Monetary Fund, Monetary Authority of Singapore, Form3, GoCardless, OneVest, Questrade, Circle, Mastercard, Flybits, Rogers Bank
On September 15, 2026, TypeSafe AI introduced Jev, an AI model built to make structured decisions inside software rather than write free form answers for people. Instead of generating prose, it returns predefined choices with probabilities and confidence scores.
TypeSafe calls Jev a System One Model and argues that this approach is better suited to autonomous software. That's the company's thesis, not an established industry conclusion. TypeSafe produced the benchmarks but acknowledges limitations in its testing and says Jev can still make wrong decisions.
AI agents can already read brokerage accounts, prepare trades, open wealth accounts, hold digital money and make payments. Giving an agent access is one thing. Letting it decide what to do with someone's money carries a different level of responsibility.
Do autonomous financial agents need a different AI architecture?
As agents progress from answering questions to spending, trading and making decisions, an error can become an action before a person sees it. If one AI interprets the request, makes the judgment and initiates the transaction, several distinct financial controls can end up depending on the same model.
Some firms are already separating what an agent understands from what it is allowed to do and what ultimately gets executed. The unanswered part is whether financial judgment should also be separated from the general purpose language model.
A specialized decision model is one option. An LLM with constrained outputs, external permissions and conventional software making the final call is another. There isn't enough independent testing in regulated financial workflows to know which approach performs better.
NCFA has already tracked AI's growing compliance burden, agents entering governed financial workflows and consent and liability when AI makes payments. Those controls become more consequential once an agent can act on a financial decision rather than simply recommend one.
Strategic Takeaway
As financial agents gain authority to spend, trade and make decisions, model architecture becomes a control question. If the same AI interprets an instruction, makes the judgment and drives the action, one mistake can pass through several stages before anyone catches it. Separating those jobs may reduce that exposure. The harder question is which financial decisions firms should hand to AI in the first place.
As financial agents gain authority to spend, trade and make decisions, model architecture becomes a control question. If the same AI interprets an instruction, makes the judgment and drives the action, one mistake can pass through several stages before anyone catches it. Separating those jobs may reduce that exposure. The harder question is which financial decisions firms should hand to AI in the first place.
Market Evidence
Click each item to expand
1. Form3 Lets The Agent Look, But Not Pay September 23, 2026, United Kingdom ⓘ
Form3's Model Context Protocol adaptor gives authorized AI agents access to payment information through a controlled interface. Initial access is read only, and the agent stays outside the payment flow.
An agent can investigate a payment or help someone understand what happened without gaining the authority to send money. Form3 has separated access from execution at the infrastructure level.
2. OneVest Automates The Work Without Taking Over The Whole Decision September 23, 2026, Canada ⓘ
OneVest launched agentic account opening for Canadian wealth firms on September 23. Its agents coordinate onboarding across custodians, track progress and handle administrative work that normally moves between people and systems.
Advisors and operations teams remain in control, and the system preserves an audit trail. The launch doesn't indicate that the agent independently approves a client or makes the final suitability decision.
3. GoCardless Lets An Agent Complete The Payment September 22, 2026, United Kingdom ⓘ
GoCardless processed an AI driven account to account payment on September 22 through work developed in the FCA AI Live Testing programme.
The agent completed the transaction inside an AI conversation, but its authority came from an existing Direct Debit mandate. It could handle more of the transaction without receiving an open ended right to move money.
4. A Canadian Agent Has Already Shopped And Paid September 21, 2026, Canada ⓘ
Mastercard, Flybits and Rogers Bank announced a completed Canadian agentic commerce transaction on September 21.
A Rogers Red Mastercard holder gave the Flybits powered assistant instructions and a spending limit. The agent selected an eligible product and completed the purchase using Mastercard Agent Pay.
The AI chose and paid, but only inside authority the customer had already defined.
5. Jev Is Built To Make A Choice, Not Write An Answer September 15, 2026, United States ⓘ
TypeSafe describes Jev as its first System One Model. A developer defines the possible output types in advance. Jev returns a decision with probabilities and a measure of confidence rather than generating an unrestricted text response.
- TypeSafe is targeting decisions made inside software rather than human conversation.
- Its training method is called Reinforcement Learning for Calibrated Decisions.
- Confidence thresholds can be used to act, reject a result or send it for review.
- Predefined outputs constrain what the model can return. They don't stop it from making the wrong decision.
A response can therefore be perfectly formatted and still be wrong, an important distinction when the output can trigger a financial action.
6. Questrade Lets The Agent Prepare A Trade July 23, 2026, Canada ⓘ
Questrade's MCP connection lets approved AI agents work with brokerage account data and prepare orders when trading permission is enabled.
The customer still reviews the security, direction, order type, quantity, price, estimated cost and commission before submitting the trade. The agent can do much of the work without getting the final say.
7. Circle Gives Agents Money With Rules Attached May 11, 2026, Global ⓘ
Circle Agent Stack gives AI agents wallets that can hold and move USDC without asking a person to approve every transaction.
Developers can limit spending, restrict services, use allowlists and put time limits around an agent's authority. The agent may decide that it wants to pay, while the wallet determines whether that payment is permitted.
Policy Evidence
Click each item to expand
8. MAS Wants Controls Running While The Agent Acts July 3, 2026, Singapore ⓘ
The Monetary Authority of Singapore and industry participants published Safeguards for Agentic Finance at Runtime, or SAFR, on July 3.
SAFR examines controls that operate while an agent is active, including whether its actions comply with policy, whether a transaction can be stopped, what records are kept and whether controls work across different systems.
Those safeguards operate outside the model rather than depending on the agent to police itself.
9. Financial AI Is Becoming An Accountability Problem Too May 6, 2026, Global ⓘ
As financial agents gain access to accounts and transactions, governance extends beyond whether the model gives a good answer. Firms also need approved data, permissioned tools, review points, records, escalation and someone accountable when things go wrong.
Once an agent can act, the choice of which AI makes a financial decision becomes part of that control structure.
10. IMF Keeps The AI Away From Final Settlement April 22, 2026, Global ⓘ
An April IMF Note divides agentic payments into three parts: intent, authorization and settlement.
AI can interpret what someone wants and turn that request into instructions. A separate authorization layer checks mandates, permissions and regulatory requirements before the approved transaction reaches payment infrastructure for settlement.
The IMF doesn't argue that LLMs should be replaced. Its model separates probabilistic AI from some of the controls governing the transaction.
How Much Should One Model Be Trusted To Do?
There isn't one financial agent architecture yet. Form3 stops at information. Questrade lets an agent prepare a trade. OneVest automates much of an account opening. Circle lets agents spend within wallet rules. GoCardless and Mastercard take agents into live payments, but inside authority established before the transaction.
None of those examples establishes a need for specialized decision models. An LLM with strong permissions and conventional software controls may be enough for many jobs.
A fraud decision can be uncertain even when the permitted actions are tightly controlled. The available actions may be tightly defined even when the right answer isn't obvious: approve it, reject it, ask for more information or send it to a person. A model designed to choose among those outcomes and report its confidence could have a role between language interpretation and a hard software rule.
A model can perform well across thousands of decisions and still get one customer's transaction badly wrong. In credit, fraud, sanctions, trading or client money, the consequences of that one decision can matter more than the average benchmark.
Financial firms already separate customer instructions, risk decisions, permissions, execution and records. As AI enters governed financial workflows, firms have to decide which of those jobs AI should perform and which should remain outside its discretion.
What Would Make Specialized Models More Convincing?
Adoption by regulated firms would test TypeSafe's benchmarks against real credit, fraud, trading and payment decisions. Independent testing could then compare specialized models with LLM based agents and conventional systems, especially where a wrong decision can cost someone money.
Regulators may influence the architecture without choosing the model. Requirements that separate interpretation from authorization, limit an agent's authority or send uncertain decisions to a person could favour some designs over others. Those boundaries are already part of the debate around AI payment consent and liability.
LLMs may keep improving while structured outputs, wallet controls and policy engines handle the hard boundaries. Specialized models could prove valuable for particular financial decisions without becoming standard infrastructure. Understanding a customer's instruction is one job. Deciding whether their money should move can be another.
Do you agree the evidence supports this answer?
Click Agree or Disagree. Your vote is recorded anonymously and aggregate results are tracked.
Explore Trending Questions
The National Crowdfunding & Fintech Association (NCFA Canada) is a financial innovation ecosystem that provides education, market intelligence, industry stewardship, networking and funding opportunities and services to thousands of community members and works closely with industry, government, partners and affiliates to create a vibrant and innovative fintech and funding industry in Canada. Decentralized and distributed, NCFA is engaged with global stakeholders and helps incubate projects and investment in fintech, alternative finance, crowdfunding, peer-to-peer finance, payments, digital assets and tokens, artificial intelligence, blockchain, cryptocurrency, regtech, and insurtech sectors. Join Canada's Fintech & Funding Community today FREE! Or become a contributing member and get perks. For more information, please visit: www.ncfacanada.org
 |  |  |
|---|---|---|
 |  |  |
Leave a Reply