Maik Taro Wehmeyer

The through-line

Wehmeyer's consistent argument is that the decisions a lender makes are its actual intellectual property, and that this IP belongs inside the risk team rather than in a legacy system or a consultant's model. That conviction came out of building machine learning models for banks and fintechs, where he saw two things: large institutions lacked the infrastructure to actually operate modern models, and whenever an outside team built a credit policy, there was "always a question around who no understands how it works and who owns it" [1]. Taktile was the answer to both: software that lets internal risk and credit teams "create more lending IP faster on their own without having any external dependencies" [1]. He returns to this repeatedly, including when describing how his own company sells: "ultimately we believe that risk policies are the core lending IP of fintech or of a bank" [1].

The framing has widened over time from credit decisioning to every mission-critical risk decision in a financial institution, and from statistical models to AI. He now describes Taktile as a vertical AI solution for banking and insurance, covering fraud checks, KYC, credit checks and AML along the whole customer journey [2], with the same idea extended to fintechs, banks and telcos [7], and with the platform positioned as decisioning infrastructure built in Europe for automated decision processes at banks, insurers and fintechs [3][8].

On unblocking the credit team

The bottleneck he attacks is the gap between when a risk expert designs something and when it runs in production. His test for a lender is simple: "how fast can you actually change your scorecards" without depending on the IT or engineering department [1]. The old three, six or twelve month delay is no longer acceptable because, as he puts it, "the world is changing so fast at the moment" and lenders want to segment risk far more precisely, so "having one scorecard for all of your customers is just not sufficient anymore" [1]. The remedy is a low-code, no-code environment where "risk experts who are not software Engineers but who have deep knowledge on the risk base" can work directly [1].

He calls the resulting practice interactive decision design, and defines it as "ensuring that the right people in the organization can collaborate to quickly build tests and then iterate on automated decision flows" [1]. It has three components in his telling. First, technical and non-technical colleagues use the same tool: at one customer in India offering a buy now pay later product for doctor's offices, engineers build the machine learning models while credit risk experts adjust the pricing logic sitting on top, both on the same platform [1]. Second, off-the-shelf data integrations remove the engineering dependency entirely. Third, sophistication can grow over time, with customers starting on simple rules and later combining "those expert rules with predictive models" [1]. The payoff he wants is operational: if a risk committee decides it has a bigger risk appetite, "can you actually start an A B test the next morning" [1].

On data as the route to more approvals

His argument for a modern decision engine rests on data access. "The amount of data that is available now has allowed lenders to really level up the accuracy of their decisions and you know most lenders don't have the infrastructure they need to harness these data sources" [1]. The commercial consequence he claims is that "we're seeing lenders increase the amount of customers they approve without taking any more risk" [1], achieved by experimenting with new sources to reach segments that were previously unscoreable. His worked example is a US B2B credit card company moving from traditional businesses, easily underwritten on bureau data, into tech startups, where open banking data reveals when a startup raised funding and how much it burns each month, giving a far better read on financial health than a bureau can produce for a young company [1].

He is emphatic that access itself drives behaviour: "pure accessibility to new data sources is something which just leads automatically to more experimentation and to much faster changing cycles of scorecards" [1]. It is not only about tuning a threshold but about testing whether payroll or accounting data would improve a default predictor at all [1]. His survey work pointed the same way, with lenders "grasping for new data sources that they have not used before"; his conclusion is that in 2023 there is enough data in the market and the problem is distribution, "you just got to get it to the risk teams in an easy manner" [1]. He has also examined how embedded lending lets businesses offer financing inside their own customer journeys, with its benefits, challenges and future opportunities [5], and how lenders can restore profitability under a tighter status quo [9].

On what AI actually changes in a regulated industry

Wehmeyer's case for AI is that models "have learned a lot of things that humans don't know about," so augmenting human intelligence with machine intelligence produces less fraud, less credit loss, higher automation rates and fewer people [2]. He is equally clear about why it is hard: "AI in many cases is just probabilistic. It's not deterministic meaning whatever input input comes out it's not 100% clear what output you're getting which the regulator doesn't necessarily like" [2]. Because banking and insurance sit close to social services in some countries, the regulator wants decisions that are fair, honest and transparent [2].

His resolution is a deliberate blend rather than an all-in bet: "mixing AI models with human intelligence and deterministic rule sets is what we do on the platform and that normally gives a lot of comfort to the regulator of like slowly introducing AI across the path until it's like being understood a bit better" [2]. For borrowers, the benefit he emphasises is speed rather than leniency. A small business applying for a loan at a large bank might wait two or three months, "by the time your business might be bankrupt or your competitors ahead," and faster acceptance is what consumers and SMBs value; fraudulent applicants, he notes, are the exception who will not like the outcome [2].

On why the money is in vertical B2B applications

He relays a Y Combinator fireside conversation in which Sam Altman argued that LLMs will take over essentially every human task but cannot build enterprise-grade applications, and that the opportunity for founders therefore lies in vertical AI solutions [2]. Wehmeyer endorses the conclusion directly: "that is actually where the big money is B2B. That's where we all should be going and that's what founders should be thinking about very heavily" [2]. His reasoning is the distance between a model and a bank: between an LLM and the legacy systems of a large institution sit legacy databases, politics and procurement, and closing that gap is what creates high barriers to entry [2]. On the model layer, he reads Anthropic as more aggressive on B2B and better at understanding the enterprise opportunity, hiring industry heavyweights and partnering with B2B AI companies, while OpenAI remains consumer-focused and is trying to hire 500 forward deployed engineers to catch up [2].

On enterprise sales into US banks

Wehmeyer moved to New York because the US is the largest market in financial services, with 8,000 banks, and selling into it requires a US sales team, which in turn requires the CEO on the ground [2]. That geographic bet also shaped how Taktile won its first American customers and prepared for the next phase of scale [6]. What he found there was a market that runs on relationships: "It's incredibly relationship driven uh and very trust driven, especially for our product, which you know, we ripping out the heart of a bank or of an insurance company" [2]. Replacing the decisioning core is a trust decision, so a small vendor has to make the buyer comfortable, and "the best proof point of you know show trust is the founder and CEO actually like being there and saying hi" [2]. His illustration is a large bank deal in Denver won after a thirty-minute meeting and a shared evening at the US Open, where the CEO told him he would win the deal "because I know you and I trust you" [2].

On the long grind to product market fit

He is blunt about the timeline: "It took us 3 years after founding the company towards signing the first customer," because "you cannot come with a halfbaked product" and there are too many bells and whistles required in this category [2][4]. He sees this as structurally at odds with investor expectations of revenue at six or twelve months, and says it is simply not possible for this type of industry and product [2]. The compensation is durability: once the product is in, "it's pretty sticky" and "we've never lost a customer" [2]. Of the first hundred pitches to banks, one said yes, which he describes as genuinely rough, since getting a no is not something most people are practised at, and concludes that "being resilient as a founder is probably the uh the most important thing" [2].

The advice he took from Y Combinator is deceptively plain: build a product your customers love, which his group partner Michael translated into spending four and a half days a week with customers rather than closing partnerships or taking investor calls, because those are "small gratifying things that by the end of the day doesn't really really help uh in order to get to product market fit" [2]. The same partner gave him his working definition: pushing a boulder up a hill means no product market fit, while chasing a boulder down the hill, where you cannot hire fast enough or serve customers or handle investor calls, is what it feels like when you have it [2]. Asked which he is doing now, he answered: "I'm chasing" [2]. He also pushes back on the idea that YC is only a first-time founder programme, pointing to continued investment in later rounds, post Series A content and small AI gatherings [2].

On the geography of building a fintech

Wehmeyer treats Berlin, London and New York as complementary rather than competing. Berlin offers some of Europe's strongest engineering talent, a growing set of licensed neobanks that make it easier to build and passport a fintech across the EU since Brexit, and a regulatory environment where building to GDPR and German information security standards means shipping to what customers in Africa, India, LatAm and the US treat as "the gold standard" [1]. London retains a uniquely concentrated financial ecosystem, with major institutions and internationally successful fintechs still headquartered there, and remains one of the most interesting fintech hubs in the world despite Brexit [1]. New York is the financial centre of the world, with a depth of lending and risk knowledge he considers "hard to copy and hard to grow into" from Berlin in a short period [1].

The distributed setup was not entirely chosen. After Y Combinator in San Francisco, Covid travel bans kept the team out of the US, so product teams were built in London, Berlin and Romania, an arrangement he now regards as a feature rather than a bug given what a dollar of engineering talent buys there [2][4]. The early cross-Atlantic reality was harsh, including going live with their first customer, one of the world's largest micro lenders, at 2am Berlin time, which he describes as thrilling but not a sustainable motion for building a company [1]. Their first fundraise landed in the first week of March 2020, when the stock market fell more than 30 percent and investors told them nobody would raise for years, a prediction that missed the coming collapse in interest rates [1].

Takeaways

• Credit policies are a lender's core IP and should be owned and iterated on in-house, not outsourced to a vendor or buried in a legacy system [1].
• Judge a lending operation by how fast a risk leader can change a scorecard without waiting on engineering; a single scorecard for all customers is no longer sufficient [1].
• Easy access to new data sources, such as open banking, payroll or accounting data, automatically produces more experimentation and lets lenders approve more customers without taking more risk [1].
• Regulators dislike probabilistic systems, so introduce AI gradually by combining models with human judgement and deterministic rule sets [2].
• The largest founder opportunity in AI is vertical, enterprise-grade B2B applications, because bridging an LLM to a bank's legacy systems, politics and procurement is the hard part [2].
• Selling mission-critical decisioning to US banks is a trust business; the founder showing up in person is the strongest proof point available to a small vendor [2].
• Deep infrastructure for regulated industries takes years, not quarters: three years to a first customer, one yes in a hundred bank pitches, and no customers lost since [2].
• Product market fit feels like chasing a boulder downhill; getting there means spending nearly all your time with customers rather than on partnerships or investors [2].