Gradient Labs gives every bank customer their own AI account manager!

Hey everyone, it's me, Shii-chan! Today I found a story about AI reshaping a bank's whole customer support. Exciting, right?

OpenAI News openai.com

What was announced?

Over on OpenAI's News, I read about Gradient Labs, a London-based startup. It was founded by a team that previously led AI and data at Monzo, and they're building AI agents that give every bank customer the experience of having a dedicated account manager.

Their platform runs on OpenAI models, and right now they're shifting production traffic onto GPT‑5.4 mini and nano.

Why it matters

Banking support is genuinely hard. Cases like fraud or blocked payments demand strict, multi-team procedures. When systems fall short, customers get passed around, wait in queues, and face delays exactly when the stakes are highest.

Gradient Labs wants AI to handle all of that reliably. And because "occasionally wrong" isn't acceptable here, accuracy really matters.

Here's how co-founder and Chief Scientist Danai Antoniou put it:

We needed three things simultaneously: accuracy at instruction-following, low hallucination rates, and function-calling reliability, all under voice latency constraints. OpenAI was the only provider that passed on all three.

What changes

To keep voice conversations natural, they're getting 500-millisecond latency with GPT‑5.4 mini and nano. That speed is what lets the system verify identity and freeze a card while talking on a live call.

Picture this flow: a customer calls to report a stolen card, the system verifies their identity (handling corrections and interruptions in real time), freezes the card and starts a replacement, then answers follow-ups like delivery timing and suggests next steps. Every step follows the procedure, decided in real time.

Dive Deep

Gradient Labs benchmarks providers on their hardest procedures using what they call trajectory accuracy: whether the system follows the correct path from start to finish. In one early eval, GPT‑4.1 was the only model to hit 97% trajectory accuracy and consistency, while the next closest provider landed at 88%. In financial services, that gap is the difference between resolving a call and creating a compliance incident.

That result shaped a hybrid architecture: OpenAI models for reasoning-heavy steps, smaller models for fast deterministic tasks, with routing that adapts to complexity and latency. Internally, specialized skills are orchestrated by a central reasoning agent, so even complex cases move across workflows without losing context.

For safety, 15+ guardrail systems run in parallel per interaction, watching for financial advice, vulnerability signals, complaints, and attempts to bypass verification or reach sensitive data. Evaluation is careful too: they replay real conversations against the expected procedure and generate synthetic conversations to cover rare edge cases.

As for results, customers report CSAT scores as high as 98%, sometimes beating their best human agents. Most deployments start with over 50% resolution rates on day one, even for complex workflows like disputes, account verification, and fraud. Revenue has grown more than 10x over the past year, expanding from inbound support into outbound and back-office work.

Wrap-up

• Gradient Labs builds AI agents that give every bank customer a dedicated-account-manager experience, and they're moving production onto GPT‑5.4 mini and nano.
• 500-millisecond latency for voice; on trajectory accuracy, GPT‑4.1 hit 97% (next closest 88%).
• A hybrid setup plus 15+ parallel guardrails, with "no hallucinations" as the guiding design principle.
• CSAT up to 98%, over 50% resolution on day one, and 10x+ revenue growth in a year.

If you're curious how to safely build AI agents in a domain where you can't afford to be wrong, this one's for you!