How Gradient Labs turned banking SOPs into audited AI workflows

Trajectory accuracy is the number they actually care about

Most support-automation pitches lead with satisfaction scores. Gradient Labs leads with something narrower and harder: whether the system follows the correct procedure from start to finish across a whole conversation. They call it trajectory accuracy, and they benchmark providers on their most difficult banking procedures rather than on average cases.

On one initial eval, they report GPT-4.1 hit 97% trajectory accuracy and consistency, with the next-best provider at 88%. Co-founder Danai Antoniou frames the gap in operational terms rather than as a leaderboard result.

In financial services, that's the difference between resolving a call and creating a compliance incident.Montana Labs

The distinction matters because a card-cancellation call can be individually correct at each turn and still end in the wrong place. Measuring the full path — verify identity, freeze the card, initiate replacement, answer follow-ups — is closer to how a bank's own auditors would judge an agent than any single-response score.

A hybrid architecture built around latency and state

The eval result shaped the design. Gradient Labs routes reasoning-intensive steps to larger OpenAI models and pushes faster, deterministic tasks to smaller ones, with routing that adapts to complexity and latency budgets. A central reasoning agent orchestrates specialized skills so a complex case can move across workflows without losing context.

Voice is the constraint driving the current migration. The company says it is seeing 500-millisecond latency on GPT-5.4 mini and nano and is moving a significant portion of production traffic onto them. Antoniou describes the requirement bluntly: the model has to hold procedure state across interruptions, backchannels, and topic switches while still generating responses fast enough for a natural conversation.

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.Montana Labs

Reliability is engineered as a rollout process, not a claim

The more transferable part of this announcement is the deployment discipline. Fifteen or more guardrail systems run in parallel on every interaction, checking for financial-advice detection, vulnerability signals, complaints, and attempts to bypass verification. That is a design that assumes the model will occasionally drift and catches it, rather than trusting it not to.

Evaluation replays real customer conversations against expected procedures and generates synthetic conversations to probe rare edge cases before anything ships. Banks then choose which issue categories the AI handles, starting with lower-risk workflows, simulate conversations first, and begin with a small slice of live traffic under continuous monitoring. Coverage expands only as performance holds.

This is what earns trust in a regulated environment: not a demo, but a visible path from historical-data analysis to simulation to a metered production ramp with human review on flagged conversations.

What automating procedures — not answers — unlocks next

The reported outcomes are strong: CSAT as high as 98%, over 50% resolution rates on day one for workflows as sensitive as disputes and fraud, and more than 10x revenue growth over the year as the company expanded from inbound support into outbound and back-office work.

The specific implication is that Gradient Labs has bet on procedure-following as the automatable unit, and that bet scales with model capability in a direct way. Every improvement in a model's ability to hold state and follow instructions expands the set of SOPs that can be safely handled — which is why their roadmap targets carrying context across interactions: customer history, ongoing issues, and picking up where a previous conversation ended. For applied teams, the lesson is that the durable moat is not the model choice but the eval harness and guardrail scaffolding that let you widen automation coverage one auditable procedure at a time.