Last week the founder story was $9B flowing into forward-deployed engineers — the labs conceding that integration, not intelligence, is the bottleneck. This week's story is the same lesson from a different angle: a Berlin startup raised $110 million not to make a smarter agent, but to let agents make decisions banks can be audited on.

Taktile's Series C — led by Goldman Sachs Alternatives, with Tiger Global, Index Ventures, Balderton, Y Combinator and Dig Ventures in the round — brings the company to $184M raised since 2020. Read the funding for what it tells you about the shape of the market, not the size of the check. The valuable agent just moved from the chat box to the loan desk.

The 30-second version#

1. The surface moved, and the surface is the whole story#

For three years "AI agent" mostly meant a thing that produces text a human then checks: a draft, a summary, a suggested reply. The human was the safety net, and the agent's mistakes were cheap because nothing happened until a person clicked.

A credit approval is not that. When an agent clears an AML alert or denies a loan, the action is the output. There's no draft stage where a human quietly fixes it — the decision moves money, rejects a customer, or files a report to a regulator. The cost of a wrong answer stops being "redo it" and becomes "bad loan on the book," "fraud that got through," or "an examiner asking why."

That's why the money is interesting. Investors didn't bet $110M on a better model — GLM-5.2 and DeepSeek V4 are a sixth of the cost and downloadable. They bet on the machinery around the model that makes its decisions safe to act on. On a decision surface, that machinery is the company.

What to do: before you build the agent, name the surface. Is your agent producing something a human reviews, or taking the action itself? The second case is a different product — price it, staff it, and architect it as one.

2. On a decision surface, "sounds right" is disqualifying#

A chatbot can be plausibly, fluently wrong and the worst case is an annoyed user. A decision agent that's plausibly, fluently wrong approves a fraudulent account with a confident rationale attached. Fluency is not a feature here; it's a hazard, because it makes bad decisions look well-reasoned.

So the credible platforms don't ship a bare model against these queues. Taktile's own framing is a hybrid: hard-coded rules for the parts of policy that must be deterministic ("never auto-approve above $X," "always escalate a sanctions hit"), agents for the judgment in between, and human oversight at the boundary. The rules aren't a lack of ambition — they're what lets you put a model anywhere near a regulated decision at all.

A chatbot that's confidently wrong wastes a minute. A decision agent that's confidently wrong books the loss and writes a tidy paragraph explaining why it was right. On this surface, "sounds right" is the failure mode, not the goal.

What to do: split your agent's logic in two. Anything that is policy — legal limits, mandatory escalations, hard caps — belongs in deterministic rules you can read and test, not in a prompt. Reserve the model for genuine judgment, and log which side made each call.

3. The audit trail is the product, not the compliance tax#

Here's the part founders under-build. On a decision surface, the requirement that feels like overhead — be able to explain, months later, exactly why the system decided what it decided — is the same capability that makes the product good.

A complete decision record (what data came in, which rules fired, what the model reasoned, whether a human overrode it, and the final action) is simultaneously your regulatory defense, your debugging tool when a decision goes wrong, your evaluation dataset for improving the agent, and your trust story to the customer whose money is on the line. Taktile can claim 95% underwriting automation because the other 5% and every automated call are inspectable — the trail is what makes the automation sellable.

This is the how-to that pairs with this week's what-happened: give the agent a replayable decision audit trail before you give it the authority to act. Most teams bolt logging on after an incident. On a decision surface, the log is load-bearing from commit one.

What to do: build the decision record before the decision logic. If you can't reconstruct why an agent did something from your own data, you don't have a product a bank — or an examiner, or your future self at 2am — can trust.

The takeaway for founders#

The dollar figure is not the news. The news is where the frontier of "useful agent" moved: from generating content a human reviews to making decisions a human, and an auditor, must be able to check. That surface exists far beyond banking — hiring, healthcare intake, insurance claims, trust-and-safety, infrastructure changes — anywhere a wrong action is expensive.

If you're building there, internalize the inversion the fintech players already have: the model is the cheap, commodity input; the durable product is the governance around it — the split between rules and judgment, the human at the boundary, and above all the replayable trail that lets you say this is exactly why. Chatbots had to sound right. Decision agents have to prove it. Build the proof first.