Gradient Labs

What it is

Gradient Labs builds AI agents that run regulated customer operations at banks, lenders, insurers and payments companies. The scope is narrower than a horizontal support agent and deeper: alongside frontline chat, email, SMS and voice, it runs the case work behind the conversation — chargeback disputes, KYC and KYB verification, collections outreach, lending follow-up, complaint handling. The vendor’s own framing is “we handle what others hand off,” and that is the actual product claim: not deflecting the first message, but closing the file.

The founders came out of Monzo. CEO Dimitri Masin was the bank’s VP of Data Science, Financial Crime and Fraud, and worked at Google before that; he co-founded the company in 2023 with Neal Lathia and Danai Antoniou, also early Monzo employees. That history explains the product shape. The people who built fraud and financial-crime operations inside a UK bank built a vendor for the queue they used to staff, which is why disputes and vulnerable-customer handling are first-class agents here rather than roadmap items.

Funding tracks the same arc. A $13M Series A led by Redpoint Ventures landed in July 2025. On 1 June 2026 the round doubled to $26M, led by Octopus Ventures and CommerzVentures with Redpoint and Exceptional Capital following, taking total funding to roughly $42.6M. The company entered the US market on 27 October 2025 and now names Current, Stash and Rho alongside its European base of Wise, Monzo, Zego, Plum, Pockit, Nala, Yonder, SteadyPay, Penfold and nsave.

Why it shows up in regulated CX stacks

• It clears the compliance review that stalls horizontal agents. The platform runs more than 20 financial-services guardrails on every conversation turn — detecting vulnerable customers, blocking unauthorised financial advice, preventing promises the firm cannot keep, and avoiding disclosures that would compromise a live fraud investigation. That is the specific artifact a second-line compliance function asks for, and it is the reason a Gradient Labs pilot reaches production in shops where a general-purpose agent gets parked pending a control assessment.
• The specialist agents are in production, not in beta. The Lending Agent is anchored at SteadyPay, the Disputes Agent runs the full chargeback lifecycle at Yonder, and there are KYB and Outbound agents for verification and for agent-initiated work like collections, document gathering and fraud alerts. Vendor-reported outcomes: 150% faster dispute cycles at Yonder, a 60% collections success rate with a 20% lift in re-engagement.
• Voice shipped on 1 December 2025 and carries the same guardrails. The agent asks clarifying questions, holds regulated call sequences in order and completes back-end tasks mid-call rather than routing out. The company reports hundreds of thousands of voice interactions a month in live environments. Voice is where regulated automation usually breaks, and it is the part of the stack most vendors are still demoing.
• It sits on top of the helpdesk you already run. Deployment is an overlay rather than a replatform, with a web app for testing before anything touches production traffic, and integration by API or CSV. One customer, Plum, is cited as live in 30 minutes with no engineering involvement — the low end of a range that is measured in weeks for a bank with real integrations.
• The model choice is documented. Anthropic’s published case study describes Gradient Labs running Claude for intent classification against its knowledge graph, standard-operating-procedure execution across multi-step processes, and response generation. Masin’s stated reason was that responses read as more natural than the alternatives tested — a rare instance of a vendor showing the layer underneath instead of calling it proprietary.
• It supports 32 million end users across its deployments, with vendor-reported revenue growth of 900% over the year to June 2026.

Pricing reality

Nothing is published. The pricing page states an outcomes-based model with no platform fees — you pay only for query resolutions the agent actually delivered — and routes to a demo request.

The unit economics are unusually legible for a contact-sales vendor, because Masin has described the model publicly. Pricing is tiered by the resolution percentage achieved, and the target is to charge around 30% of what the same volume costs in human support, leaving roughly 70% savings before you account for ticket-mix differences. He contrasts that explicitly with per-conversation billing — Salesforce’s roughly $2 per AI conversation regardless of outcome — and the company backs the model with a money-back commitment if it misses what it committed to.

Run the arithmetic against your own fully-loaded cost per contact, not against a software licence. Intercom Fin bills $0.99 per resolution and Zendesk $1.50 per automated resolution, and both numbers buy software your team configures and tunes. Gradient Labs prices against the labour line instead, which makes it look expensive next to Fin’s per-unit rate and cheap next to a regulated operations team in London or New York. The comparison that decides the deal is Gradient Labs’ effective rate against your blended agent cost per contact for the regulated queues specifically — disputes and KYC cases, where handle times run long and the people are expensive — rather than against your password-reset volume, where it will lose.

The one number to pin down in writing is the definition of a resolution, because the tier you land in and the invoice both key off it.

Best for

Support, operations and CX leaders at regulated fintechs, digital banks, lenders, insurers and payments companies — roughly Series B and up, with hundreds of thousands to tens of millions of end users — where a meaningful share of the queue is regulated case work rather than FAQ deflection. It is the strongest pick when compliance has already blocked or slowed a Decagon or Sierra evaluation, when disputes and KYC backlogs are the constraint rather than first-response time, and when you want a vendor delivery team to own the migration instead of hiring agent engineers.

Skip it if your support surface is not regulated — the guardrails you are paying for are the whole differentiator, and outside financial services they are overhead on a narrower feature set. Skip it if you intend to own and tune the agent in-house with your own engineers, if procurement cannot approve spend without a published rate card, or if your regulated case volume is low enough that the incumbent helpdesk’s agent plus a small ops team already clears it.

Versus the alternatives

Intercom Fin and Zendesk are the volume incumbents, and for most teams the cheapest first test remains the agent inside the helpdesk you already pay for — take that path when the queue is conversation-shaped, the knowledge base is current, and nothing in the flow needs a regulated procedure executed in a fixed order. Sierra is the fastest-growing entrant in the segment and the better pick for consumer brands at enterprise volume that want to own agent design themselves; Decagon is the pick when you have engineering capacity to invest and the work resolves inside single conversations. Lorikeet is the closest genuine substitute — also built for complex, high-stakes support rather than deflection — and it is the head-to-head to run if you want a second quote on the same problem. Parloa wins when the mandate is voice-first contact-center transformation across languages rather than case resolution. Crescendo is the answer when you need humans contracted underneath the agent rather than guardrails on top of it. The honest baseline is your current outsourcer plus the helpdesk agent you already own, and Gradient Labs’ argument against it is specific: it is the option that automates the investigation, not just the reply.

Watch-outs

• Every headline performance number is vendor-published, and each one defines its own metric. 80-90% resolution, 98% QA pass rate, 98% CSAT, 150% faster disputes, 900% revenue growth — all company material, none independently audited. The skeptical counterweight is worth holding next to it: Salesforce’s own researchers measured LLM agents succeeding on 58% of single-turn tasks and 35% of multi-turn requests against marketing claims of 83%. Guard: write your own definition of resolution into the contract, measure it on your own traffic across a two-week shadow period before go-live, and make the acceptance test that number rather than the reference deck.
• The day-one rate and the mature rate are far apart, and the gap is work you staff. The company states 40-60% resolution without customisation, reaching 80-90% after three to five months of data integration. That interval is integration engineering, knowledge curation and QA review on your side, running while you pay for the resolutions the agent does land. Guard: build the ramp into the commercial terms — a milestone schedule tied to the resolution curve, with a defined remedy if the 80% band is not reached by month six — and ask for reference customers at your contact volume who are actually holding the mature rate.
• Outcomes-based pricing aligns the incentive but the vendor defines and meters the outcome. Gradient Labs decides what counts as a resolution and its agents generate the volume being billed. Guard: define a resolution in the contract — no reopen within seven days, no re-contact on the same issue, no credit for a transfer to a human — cap billable resolutions per contact, and require a monthly line-item report you can reconcile against your own helpdesk data.
• Handing regulated customer operations to a three-year-old vendor is an operational-resilience decision, not a software purchase. EU financial entities fall under DORA’s ICT third-party risk regime, and UK firms under the FCA’s operational resilience rules; an AI agent running disputes and KYC is an important business service by any reasonable mapping. Guard: run the assessment before the pilot, not after — documented exit plan and substitutability analysis, sub-processor list and data locations in writing, contractual audit and supervisory access rights, and a tested manual fallback for the queues the agent will own.
• The vertical focus that gets it through compliance also caps where it can go. The guardrails, the agents and the delivery team are all built for financial services. If your business expands into non-regulated support surfaces, or you acquire a product line outside finance, you are running a second agent vendor rather than extending this one. Guard: scope the contract to the regulated queues explicitly, keep the helpdesk-native agent live on general support instead of decommissioning it, and re-test the split at renewal rather than consolidating on either side by default.