Gradient Labs
Also known as: Otto, Lending Agent, Disputes Agent, Gradient Labs AI
Specialist AI agents for regulated financial services, covering frontline support, outbound and back office work across lending, disputes, KYC and KYB, that follow a bank's own written procedures end to end with financial services guardrails on every turn.
Gradient Labs builds AI native customer operations for financial services: a suite of specialist agents for lending, disputes and KYC, with a platform that runs the operations between them. Founded in London in 2023 by Dimitri Masin, Neal Lathia and Danai Antoniou, who built Monzo's AI and data teams, it has a New York office and names Wise, Monzo, Zego, Pockit, Current, Stash, Rho, Yonder, SteadyPay and Plum among its customers. It increased its Series A to $26 million in June 2026, led by Octopus Ventures and CommerzVentures with follow on backing from Redpoint Ventures and Exceptional Capital, and says its agents serve more than 32 million end users.
The agents cover three kinds of work. Frontline support handles customer conversations across phone, chat, email, SMS and social media, reasons through each case and takes actions such as freezing a lost card or tracing a missing payment; customers know this agent as Otto. The outbound agent starts conversations itself over text, email and voice for document collection, payment failures, fraud alerts and collections, and can leave voicemails and honour callbacks. Back office agents investigate flagged cases and escalate with a case file.
The Disputes Agent takes a dispute in by voice, text, email or a form, classifies it against card scheme reason codes, reviews the evidence, chases what is missing, decides the outcome, waits for a person to approve or override with a full audit trail of its decisions, then files the chargeback directly with Mastercard or Visa and tells the customer. The Lending Agent runs the borrower lifecycle from abandoned applications to collections and, by Gradient Labs' own count, handles more than 100,000 voice calls a month across lending customers, in 30 plus languages.
Each agent follows plain language procedures the operations team writes with no code, and Collaborate, launched in August 2026, keeps those procedures, tools and test cases version controlled like code: every change can be drafted, tested against scenario based test cases, previewed and approved, nothing goes live without approval, a change can be rolled back with one click, and the agent proposes fixes from what it sees in production for a person to approve.
More than 20 financial services guardrails run on every turn. Customer guardrails detect complaints, vulnerability and financial difficulty and hand the conversation to a person or to a dedicated procedure, and agent guardrails inspect each drafted reply for tipping off, advice that cannot be given and disallowed terms, and edit it to comply. Gradient Labs states full SOC 2 Type 2 certification and says it is fully GDPR compliant, with a Vanta trust centre, SSO, audit logs and role based permissions, and says it is built for US, UK and EU rules including FCA Consumer Duty, CONC, Reg E, Reg Z, PSD2 and the EU AI Act.
It runs on models from OpenAI, Anthropic and Google and switches between them when a provider has an issue. On Intercom, Zendesk or Freshdesk it sits on the existing support platform from day one with no technical integration, and it can run headless through its Conversations API inside a bank's own channels. Pricing is outcomes based with no platform fee, with no figures published and a model worked out on a demo request, and once a use case is scoped Gradient Labs guarantees the deployment, with the customer's money back if it does not deliver.
From our comparison video
Gradient Labs vs Lorikeet for Regulated Support Teams
Our 7 minute video puts Gradient Labs side by side with Lorikeet. It covers what each one does and how each is priced, then when to choose which.
Recorded in November 2026. The profile here is live and updates as research is verified.
Vendor details
Canonical URL
https://gradient-labs.ai
Category
Customer support agent
Company status
independent
Use cases & customers
Primary use cases
Target customers
In practice
A customer disputes a card charge. Otto gathers and verifies the evidence, decides the outcome, escalates to a human for sign off, and files the chargeback with the card scheme, running the dispute end to end.
A borrower falls into arrears. Otto contacts them at the time they are most likely to respond, personalizes the message from their account history, and secures a promise to pay, all without a human agent placing the call.
A fintech onboarding a business needs verification. Otto checks each document against internal policy, requests anything missing from the customer, verifies the business, and routes it to onboarding, automating a slow manual compliance step.
Sources & related URLs
Research sources
Agentic Index coverage score
10.0 / 14 capabilities · 71%
| Integrations & Tool Calling | Full |
|---|---|
|
The agents take authenticated action in outside systems, and those actions carry real consequences. The Disputes Agent plugs into the help desk, the card processor and the scheme APIs, Intercom, Zendesk, Mastercard and Visa Resolve Online among them, and files the chargeback directly with the card scheme. The Lending Agent plugs into a loan management system, the contact centre and the payment processor, verifies identity and captures promises to pay, and agents freeze cards, order replacements and trace payments in a bank's own systems. On Intercom, Zendesk or Freshdesk the platform sits on the existing support platform from day one with no technical integration, connects over API, takes a CSV upload of accounts, and lets a customer register its own tools, such as a fraud scoring service or a CRM lookup, through the Conversations API. Sourcegradient-labs.ai/product/disputes-agentread 2026-11-02 |
|
| Workflow Orchestration | Full |
|
Regulated workflows that would take a person days are carried out in multiple steps across multiple agents. The Disputes Agent takes the dispute in by voice, text, email or a form, classifies it against card scheme reason codes, reviews the evidence against scheme requirements and checks for prior claims, contacts the customer for anything missing and resumes when they reply, applies the bank's decisioning logic, waits for a person to approve or override, files the chargeback with the scheme and tells the customer the outcome. The Lending Agent runs the borrower lifecycle from abandoned applications and onboarding calls to hardship reviews and collections. Frontline, outbound and back office agents share case context, and back office work such as KYB verification is triggered by system flags. Procedures are written in plain language in a live document, with no code, and procedure versions can be gated, capped and promoted through the API. Sourcegradient-labs.ai/product/disputes-agentread 2026-11-02 |
|
| Knowledge Grounding & RAG | Partial |
|
The agent is trained on a company's own SOPs and its best human agents: it learns from real conversations, not only the help centre, and executes natural language procedures that mirror the operation. The operations team defines in a live document, in plain language, how the agent should think through a problem, answer and act, and a bank can sync its own CMS or knowledge base into the agent through the API so new articles flow in automatically. The agent spots knowledge it was missing by watching how human agents answered and proposes additions. Gradient Labs names no index, embeddings layer or retrieval architecture, and does not say whether answers cite their sources. Sourcegradient-labs.ai/productread 2026-11-02 |
|
| Human Oversight & Guardrails | Full |
|
More than 20 financial services guardrails run on every turn of every conversation, in two suites. Customer guardrails watch what the customer says for complaints, vulnerability and financial difficulty, and when one fires the agent is interrupted and the conversation is handed to a person or rerouted to a procedure for that situation. Agent guardrails inspect each drafted reply for tipping off, advice that cannot legally be given and disallowed terminology, and edit the draft to comply; a bank can layer its own guardrails on top or route drafts through them. The team sets approval requirements for sensitive decisions, a person approves or overrides a dispute outcome before the chargeback is filed, lending conversations that need a specialist are escalated, the agent can ask a human for help mid conversation, and every change to the agent waits for sign off in Collaborate. One European digital bank reports 9 million guardrail runs. Sourcegradient-labs.ai/product/voiceread 2026-11-02 |
|
| Security, Identity & Governance | Full |
|
Gradient Labs states full SOC 2 type 2 certification and says it is fully GDPR compliant, with a Vanta trust centre linked from every page that holds the documentation; the trust centre refuses automated reading, so its detail is unconfirmed here. On the access side it names SSO, comprehensive audit logs and role based permissions as its enterprise controls, and Collaborate records who made each change to an agent, who approved it, and lets a change be rolled back. More than 20 financial services guardrails run on every turn, with 9 million guardrail runs reported at one European digital bank, and the terms of service forbid storing or using customer personal data beyond performing the service or training models on it. The subprocessor list is available on request, not published, and neither SSO nor the role model is documented beyond that one line. Sourcegradient-labs.airead 2026-11-02 |
|
| Observability & Auditability | Full |
|
The agent's own decisions are on the record. On a dispute, the agent hands the team a full audit trail of its decisions before a person approves or overrides the outcome and the chargeback is filed. When a guardrail fires, the decision and the chain of thought reasoning behind it are visible to the staff who operate the agent, who can leave feedback if they disagree, and Gradient Labs says those figures are reported to banks' model risk committees. Collaborate keeps the version of the agent that was live, the change that introduced a behaviour and the approval behind it on file for a regulator, and scorers watch live traffic and raise alerts when quality dips. Real time monitoring tracks resolution rates and customer satisfaction per deployment, and a European digital bank reports a 98% QA score with guardrail run counts. No exportable audit trail or step level trace of every run is described. Sourcegradient-labs.ai/product/disputes-agentread 2026-11-02 |
|
| Memory & State Persistence | Partial |
|
The frontline agent remembers past conversations when a customer gets in touch, and on lending, conversations and commitments are tracked across the borrower's lifetime and passed as context to the next call, with voicemails, callbacks and text follow ups carrying the history forward. Gradient Labs says its specialist agents work as one system, sharing context and memory and passing work between each other. It names no memory layer that a customer can see, and does not say where memory lives, how long it lasts or how to delete one customer's memory without touching the record. Sourcegradient-labs.ai/productread 2026-11-02 |
|
| Deployment & Data Residency | Not documented |
|
No hosting region is published, and Gradient Labs names no residency commitment, VPC, single tenant or self hosted option. The Lending Agent page lists data residency only as one of the EU rules the agent is built for, the terms of service are under English law and name no region, and the privacy policy covers site visitors rather than the product. The subprocessor list is available on request, not published. The Vanta trust centre may name hosting and residency, but it refuses automated reading, so a buyer should ask for it directly. Sourcegradient-labs.ai/terms-of-service/latestread 2026-11-02 |
|
| Prebuilt Agents, Templates & Packs | Full |
|
Customers adopt a named set of prebuilt specialist agents as units, each with its own product page: the Lending Agent for the borrower lifecycle, the Disputes Agent for the chargeback lifecycle, a KYB agent for business verification, the Voice agent and the Outbound agent, over a platform that runs the operations between them, and each comes with the guardrails, compliance checks and test scenarios for its domain. A use case library lists seventeen ready use cases across retail banking, lending, insurance, remittances, investing and crypto, from freezing a lost card to first notification of loss and ISA transfers. Gradient Labs says the Lending Agent is live at SteadyPay and the Disputes Agent at Yonder, and a lending deployment can start from a CSV of accounts in a day. Sourcegradient-labs.ai/productread 2026-11-02 |
|
| Triggers & Channel Coverage | Full |
|
Customers reach the agent across phone, chat, email, SMS and social media, one agent on every channel, in multiple languages, and the agent also starts conversations itself. The outbound agent texts, emails and calls customers for document collection, payment failures, fraud alerts, collections and account activation, leaves voicemails, honours callback requests, and is started by the team, by system alerts or by team workflows. The Lending Agent places more than 100,000 outbound voice calls a month across lending customers and chooses who to contact, when and on which channel. Back office investigations such as disputes and KYB arrive as cases, not messages, and the agent can also receive a dispute from a form on the bank's site or app. Sourcegradient-labs.ai/product/outboundread 2026-11-02 |
|
| Model Flexibility & Routing | Partial |
|
The agent is not tied to one model provider. Gradient Labs runs it on models from Anthropic, Google and OpenAI, adopts new releases quickly, and switches to another provider when one has an issue, with failover across multiple cloud and model providers. The choice and the routing are Gradient Labs' own; no customer facing model selection, routing control or bring your own model option is described. Sourcegradient-labs.ai/pricingread 2026-11-02 |
|
| APIs, SDKs & MCP Extensibility | Partial |
|
The Conversations API lets a bank run the agent headless inside its own channels: the customer's backend creates a conversation when a customer opens in app chat or calls, forwards each message, and receives the agent's replies and handoff events by webhook, with the same API covering agent initiated voice calls. Through it engineers can add and update knowledge, gate and cap a new procedure version, run two variants and read the difference, and register custom tools and resources the agent can call. The pricing page offers the public API on request. Collaborate also exposes procedures, tools and test cases as files in a git repository. No API reference, SDK or developer portal is published, so the surface is asserted and described rather than documented. Sourcegradient-labs.ai/blog/how-and-why-to-use-headless-ai-agentsread 2026-11-02 |
|
| Testing, Debugging & Optimization | Full |
|
Collaborate, launched in August 2026, tests every change before it goes live. Test suites are scenario based synthesised conversations, built from real past conversations with predefined tool responses, and every test case runs against a change in one batch; a green suite is the bar for launch, runs can be triggered from the customer's own CI, and a draft can be previewed by chatting to the new version before a customer sees it. New capabilities can roll out gradually to a share of live traffic, scorers watch live traffic and fire alerts when quality dips, and the agent proposes fixes from the conversations it escalated and turns every failure into a permanent test case, with nothing going live without approval. Any AI agent such as Claude or Codex can draft a procedure change and write test cases for it on a branch. The product page also offers testing every scenario in the web app before deployment and continuous automated quality checks. Sourcegradient-labs.ai/blog/introducing-collaborate-improve-your-ai-agent-like-coderead 2026-11-02 |
|
| Browser & Computer Use | Not documented |
|
The agents act programmatically, through integrations into a customer's help desk, loan systems, payment processors and the card schemes, and through the Conversations API. Gradient Labs names no browser control, hosted session, virtual desktop or computer use. Sourcegradient-labs.ai/productread 2026-11-02 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
Gradient Labs released Collaborate, a new workspace that allows operators, engineers, and AI agents to work together as peers. The update introduces built-in version control, evaluations, and continual learning capabilities for agent management.
Bears on: Observability / auditability
View sourcePricing
Contact for pricing
What is public
Gradient Labs publishes the shape of its pricing and no figures: outcomes based, no platform fees, paid only for successful resolutions the agent delivers, quoted through sales. Its agent pages offer a free pilot, its $10K challenge page offers $10,000 if its agent loses a head to head test against another vendor, and its Series A post (June 2026) states a money back guarantee on a scoped deployment.
Sales call required
Yes, required for paid access
Free / trial
Free pilot
Missing data
No rate per resolution, minimum, plan structure, contract term or cancellation terms are published, and what counts as a successful resolution is not defined on the site.
Related vendors
- Decagon — AI agent platform purpose-built for customer support, with deep…
- 5.Y — 5
- Ada — Agentic customer experience platform whose AI agents autonomously…
- Aisera — Enterprise agentic AI platform that orchestrates specialized agents…
- ASAPP — Generative AI for enterprise contact centers that plugs into…
- Assembled — AI customer support orchestration platform that unifies autonomous…
Alternatives to Gradient Labs
The closest documented capability profiles to Gradient Labs among customer support agents tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Cresta11.0 / 14Fuller documented coverage on Knowledge Grounding & RAG and APIs, SDKs & MCP Extensibility
- Kustomer11.0 / 14Fuller documented coverage on Knowledge Grounding & RAG and APIs, SDKs & MCP Extensibility
- Sprinklr11.0 / 14Fuller documented coverage on Knowledge Grounding & RAG and APIs, SDKs & MCP Extensibility
- Five911.5 / 14Adds documented Deployment & Data Residency
- Gallabox11.5 / 14Fuller documented coverage on Knowledge Grounding & RAG and Model Flexibility & Routing
- Outverse8.5 / 14A lighter documented profile than Gradient Labs
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded