Latest / MBA Training AI / Fintech AI timelines: the evidence clock beats the build clock

Transcript

Fintech AI timelines: the evidence clock beats the build clock · MBA Training AI

• 0:05
  Host: MBA Training, AI edition. We get into Schedule fintech AI by the evidence clock.
• 0:10
  Host: A bakery can buy an oven in an afternoon, but nobody calls a sourdough starter ready until it has risen enough times to trust it. Fintech AI vendors are selling ovens right now and quoting delivery in days.
• 0:22
  Expert: And the buyers keep confusing the oven with the bread. Yesterday, October 7, Nous Research confirmed a $90 million Series B at a $1.5 billion valuation and launched Hermes for Businesses. That's their open-source agent, packaged with private deployment, single sign-on, meaning one company login for everything, plus auditing and service-level agreements, which are contractual uptime and support promises. It's exactly the checklist a bank's security chief asks for first.
• 0:53
  Host: So the build problem is solved. Why isn't everybody live next week?
• 0:56
  Expert: Building the agent was never what made these projects late. What does it is what I call the evidence clock: how long a deployed system needs to produce enough real outcomes, on the metric your control people care about, before it can act without a human checking every case.
• 1:12
  Host: Give me Nous's numbers before you lecture me.
• 1:14
  Expert: The Wall Street Journal reported roughly $36 million in annualized revenue by mid-September 2026, with an expectation to pass $100 million before year-end. There's also a claim that Hermes drives about 2.5% of global AI token usage, tokens being the chunks of text models process. That figure comes from Nous itself and hasn't been independently verified, so treat it as marketing until someone checks it.
• 1:43
  Host: Noted. And Dealroom's take?
• 1:45
  Expert: When the round was first reported, Dealroom pointed out that companies raising at those valuations have to prove measurable value fast, and autonomous agents are still unproven across many real workflows. That pressure lands on a fintech as a sales pitch: the agent's ready, so why isn't it live?
• 2:03
  Host: Fair question from a board, honestly.
• 2:06
  Expert: It's fair, and the examiner's question is fairer: what evidence justified each autonomous decision? In the US, there's no longer a template to answer it with. On April 17, 2026, the OCC, the Fed and the FDIC, which are the three main federal bank regulators, rescinded the 2011 model risk framework. The replacement, SR 26-2, explicitly excludes generative and agentic AI. If you run an LLM agent, meaning a large language model that takes actions, at a bank or a bank partner, you write your own evidence standard and then prove you met it.
• 2:44
  Host: Europe's softer?
• 2:46
  Expert: Europe moved the date. It didn't cancel anything. The Digital Omnibus, the EU's package that amends the AI Act, pushed the high-risk obligations covering credit scoring from August 2026 to December 2, 2027. Article 50, the transparency duty to tell people they're dealing with AI, has applied since August 2 this year. High-risk breaches can cost up to €15 million or 3% of global turnover.
• 3:15
  Host: Meanwhile the startups are eating the incumbents' lunch.
• 3:17
  Expert: Partly. The Cambridge Centre for Alternative Finance's 2026 global survey has fintechs ahead on advanced AI adoption, 47% to 30%, and 56% reporting higher profitability versus 34% of traditional institutions. But only 14% see AI as transformational to their strategy. Plenty of shipping, not much conviction.
• 3:43
  Host: Field notes, then. What's going wrong out there?
• 3:45
  Expert: Teams track the value half of the metric and forget the guardrail half. Value is handle time, alerts cleared per analyst, cost per resolved dispute. The guardrail is the harm a regulator cares about: missed suspicious activity, wrong adverse action reasons, which are the reasons you give a declined borrower, plus upheld complaints and customers who couldn't reach a human.
• 4:08
  Host: Name a casualty.
• 4:10
  Expert: Klarna. Its assistant launched in February 2024, handled 2.3 million conversations, two-thirds of its customer service chats, and Klarna said that equaled 700 full-time agents. Fourteen months later, in May 2025, the CEO, Sebastian Siemiatkowski, told Bloomberg that cost had been "a too predominant evaluation factor" and the result was "lower quality."
• 4:36
  Host: Did it flop?
• 4:37
  Expert: It didn't. On the Q3 2025 earnings call they reported it doing the work of 853 agents with $60 million in savings. The value metric showed up within a month. The quality evidence took more than a year to arrive. That gap is the evidence clock in a single company.
• 4:55
  Host: Who should own the guardrail?
• 4:57
  Expert: Someone outside the project team: second-line compliance, the independent risk function, or the MLRO, the money laundering reporting officer, or model risk. If the guardrail owner reports to the delivery lead, the guardrail loses to the deadline every time.
• 5:13
  Host: Now the arithmetic. Make it something I can do on a napkin.
• 5:17
  Expert: Use the rule of three. If an agent makes zero errors across n independent cases, the 95% upper bound on its true error rate is about three divided by n. To claim it misses fewer than 1% of cases that should be escalated, you need around 300 escalation-worthy cases with zero misses. You don't need 300 cases in total. You need 300 of the rare kind.
• 5:41
  Host: Walk me through a real one.
• 5:42
  Expert: This one's hypothetical, purely illustrative. A payments EMI, an electronic money institution, sends 4,000 transaction-monitoring alerts a month to an agent in shadow mode, where the agent recommends close or escalate but humans still decide. Historically 2% get escalated, so you see 80 escalation-worthy alerts a month. Three hundred divided by 80 is 3.75, so about four months of shadow running. If the sign-off committee meets quarterly, add up to three months. The build takes a few weeks. You're looking at six to seven months, and the build is the smallest part.
• 6:20
  Host: And if the board gets nervous and wants tighter?
• 6:23
  Expert: A 0.5% bound doubles the sample to 600 cases, which means about seven and a half months of shadow running. Tightening the guardrail moves the date further than any improvement to the agent can.
• 6:35
  Host: Even regulators can't be that slow.
• 6:37
  Expert: They are, and they're the ones who want adoption. The FCA, the UK's Financial Conduct Authority, has a second AI Live Testing cohort with Barclays, Experian, Lloyds and UBS covering agentic payments, anti-money laundering and KYC, know-your-customer checks. Applications opened in January 2026, testing started in April, and the evaluation report lands in Q1 2027. That's about fifteen months. FIS's Financial Crimes AI Agent, built with Anthropic, has BMO and Amalgamated Bank in development first, with general availability set for the second half of 2026.
• 7:19
  Host: So when a vendor says days, they're lying?
• 7:22
  Expert: They're describing a different clock. Gradient Labs, which sells agents to banks, says it goes live in days and offers a money-back guarantee on scoped use cases. Assistents.ai gives a range from under four weeks to six to twelve months, depending on how much custom training and integration you need. Both are vendors selling their own product, so read those as commercial claims. More to the point, both are quoting the build clock. Neither one controls your alert volumes or your base rates.
• 7:52
  Host: When is a short timeline actually honest?
• 7:54
  Expert: When a human reviews every output, the action can be reversed, and no regulated decision hangs on it. Analyst-facing KYC summaries and drafts a person edits both qualify. The bad outcome there, a rejected draft, is common, so the evidence piles up in days. Once the agent closes something, like declining credit, closing an alert, blocking a payment or resolving a complaint, the bad outcome is rare by design, and the clock is long by design.
• 8:25
  Host: Where do smart teams cheat?
• 8:26
  Expert: They swap guardrails. They measure "analyst edited the output," which happens constantly, instead of "SAR-worthy activity missed," a SAR being a suspicious activity report. Blocking that substitution is the guardrail owner's whole job. The other traps are permanent pilots with no agreed threshold, and forgetting that fraud rings cluster, which breaks the independence assumption behind the math. Treat 300 as a floor. Also budget for shadow mode honestly, because you're paying for the agent and the analysts at the same time, or your business case will look broken by month three.
• 9:03
  Host: One thing for Monday morning.
• 9:04
  Expert: Pick one use case and pull the monthly count of its worst outcome from the last twelve months of production data. Divide three by your target error rate, then divide that by the monthly count. If any go-live date on your roadmap is shorter than that number, cross it out before someone signs it.
• 9:22
  Host: Sources: Startup Fortune, Tokenpost and Github. Every link is in the show notes.
• 9:27
  Host: That's it from us. The reading continues at mba-training.com, new AI analysis every day.