Taktile

Bank statement transaction classification just became a classification problem again. When a small business applies for a loan, an underwriter sorts every transaction on its bank statements into revenue, payroll, loan repayments and so on. Since 2023 the AI answer has been to throw a frontier LLM at it. That works (Claude Opus 5.5 got 95.5% in our test), but you pay frontier prices for every single transaction. Decision models which arrived with the release of TypeSafe AI's Jev in September, change that. They pick from a fixed list of categories and return a probability for each one. Because they're built on a large pretrained model, they keep its general world knowledge and can classify new categories from definition alone without retraining. We tested 23 methods on 72 bank statements and developed a new recipe: start with just category definitions and a frontier LLM, save every underwriter correction as a label, and fine-tune a decision model once you have enough labelled data. Turns out "enough" is not that much. With about 64 three month statements, a fine-tuned 4B decision model hit 97.3% and beat every off-the-shelf model at about 1/80th of the cost of Opus. A traditional bag of words classifier needed about 4x as much labelled data to get close. Perhaps the most important finding is around confidence scoring. The fine-tuned decision model knows when it's unsure and sends only those transactions for secondary review (by another model or human reviewer). Opus is not as proficient at self scoring and sends more than half to a person because it underrates its own confidence. The takeaway: once you have labelled data, you no longer have to choose between accuracy and cost because a fine-tuned decision model gives both. Full benchmark and methodology on the Taktile Labs website. Link in comments.

Hoje lançamos o Taktile Insights. Quero contar por que isso importa para quem é dono de um KPI de crédito, sinistro ou onboarding no Brasil. Nos últimos meses sentei com dezenas de executivos de bancos e seguradoras aqui. A cena se repete. O diretor de sinistros vê o custo médio subindo. O gerente de crédito vê a aprovação caindo. Eles são os donos do número. Mas a resposta para o “por quê” mora em outro lugar: na fila do time de BI, num dashboard que ainda vai ser construído, numa extração que chega em duas semanas. Quando a resposta chega, a pergunta já mudou. A parte de agir, a gente já tinha resolvido. O Igor Viana, analista de crédito da MagaluPay, mudou uma política de crédito em 40 minutos, incluindo validação e ver ela rodando em produção. No motor anterior, isso levava de três a quatro semanas. O que ainda demorava era entender. É isso que o Insights resolve. Você pergunta em linguagem natural para os dados das suas decisões e recebe a resposta e o gráfico ali, no mesmo lugar onde as decisões rodam. Sem ticket, sem fila, sem depender de alguém que não é dono do resultado. No Brasil, velocidade de decisão deixou de ser diferencial faz tempo. O diferencial agora é quem aprende mais rápido com cada decisão que toma. Obrigado a cada cliente que usou a versão inicial e falou o que achou. Blog nos comentários.

Today we launched Taktile Insights. The idea came from a pattern we kept hearing from customers: teams that own KPIs like credit returns or claims leakage were too far from the data to act on it quickly. So we built Insights inside the decision platform. Teams can query decision data in plain language, get an answer, and implement changes - all in the place where the decisions run. No more waiting for BI teams to run one-off analysis to uncover the cause for an approval rate drop or a decrease in straight through processing rate. Not just making decisions faster, but learning and iterating faster so the next decision is better. Thanks to all the customers who worked with us to build the solution you really need. And congrats to the team for the launch 👏 Shriram Pawar Paula Guagliardi Tom Larkworthy Marten Kremser Egemen Toykan Gareth A. Davies More info via the link in the comments.

So much fear around AI in banks and insurers is that it takes control away from people. Our newest solution does the opposite. Super excited to introduce Taktile Insights - Insights empowers the people who own a decision to ask their data a question, get the answer in seconds, and act on it immediately. In one system, they can access any view of the data and shift any KPI with their own hands, from onboarding to underwriting and claims. One of my favorite stories in beta, from Igor Viana at MagaluPay: “A policy change that used to take us 3 to 4 weeks in our previous credit engine, we did in Taktile in 40 minutes — including validation and seeing it live in production.” That’s huge!! From weeks to literally minutes. Imagine that across every critical decision a bank or insurer makes - the improvement you would see in credit returns, product adoption, claims leakage etc. We’re talking about a massive shift in how the teams that drive the P&L in banking and insurance work, and how quickly institutions can adapt. When every bank and insurer can use AI to approve a loan or identify a fraudulent claim in seconds, decision speed will not be the advantage. The real frontier is using AI to enable teams to not just decide fast, but learn and improve constantly. Huge congrats to Shriram Pawar, Gareth A. Davies and Thomas Piani on an incredible build and launch. More in the blog link in comments!

Manuelle Arbeit raubt Zeit für die Entscheidungen, bei denen es wirklich auf menschliche Expertise ankommt.

Jede Woche ohne Go-live für eine neue Kreditrichtlinie heißt: eine Woche länger Entscheidungen nach veralteter Logik.