Fini

Fini layers onto an existing helpdesk through one-click OAuth, with marketplace-verified listings on Zendesk, Intercom, HubSpot, Front and Gorgias plus Salesforce, Freshdesk, LiveChat, Slack and Discord, and connects backend billing, CRM and internal systems so the agent takes real actions, with a worked example calling Stripe customer search, charge listing and refund creation before closing the ticket.

Sourceusefini.com/product/integrationsread 2026-09-05

A named orchestration layer with a documented decision and execution path. The LLM Supervisor is published as layer two of the architecture and is described as choosing the optimal path between retrieve, reason, reply and act, using conversation context, custom rules and tone instructions, and access to enterprise systems including CRM and order management.

Beneath it, work routes to specialized skill modules: Search and Retrieve, Write Response, Follow Process running internal flows such as refunds, cancellations and verifications with automated rule enforcement, and Take Action triggering API calls into business systems with no human handoff required. Fini's own worked examples show the chain end to end, checking a refund policy, verifying a delivery timeline by API and initiating the refund in one pass.

Sourceusefini.com/product/platformread 2026-09-05

Knowledge Atlas is a named, vendor-maintained knowledge structure: it ingests the customer's knowledge base and historical tickets into a graph, scans resolved tickets to detect gaps against that graph, generates the missing article and commits the update, so the retrieval structure persists and repairs itself rather than being assembled per conversation. Answers are grounded in it, with the reasoning traced back to approved internal knowledge.

Sourceusefini.com/knowledge-atlasread 2026-09-05

Guardrails are layer one of a published five-layer architecture, sitting ahead of the action rather than reviewing after it. A real-time safety check blocks harmful, off-topic and policy-violating content, masks personally identifiable information such as phone numbers and emails, and enforces company-specific rules including refund policy, with only safe valid queries moving forward.

Inside the response module every generated reply is evaluated for accuracy, tone and safety before being sent, which Fini states as every answer being LLM-reviewed rather than only LLM-generated, and the Follow Process module enforces policy automatically on refunds, cancellations and verifications.

A published case study names four implemented checks: Polite Language, Competitor, Topic escalation for complex cases such as refunds, and Truth Check verifying accuracy before responses reach customers, with output rails using LLM judges to flag, correct or escalate to a human.

Sourceusefini.com/product/platformread 2026-09-05

A live SafeBase trust portal publishes GDPR, ISO/IEC 27001 and SOC 2 with the SOC 2 report, ISO certificate, penetration test report, DPA and cyber insurance obtainable on request, alongside named product controls covering audit logging, access monitoring, data erasure, credential management and password security, and a subprocessor list. PCI DSS Level 1 and HIPAA with BAA appear as homepage badges but are absent from the trust portal's certification list, and ISO 42001 is not shown on any Fini page.

Sourcesecurity.usefini.comread 2026-09-05

Voice, chat and email run on one reasoning layer sharing the same policies and a single audit trail, and every answer is classified, confidence-scored and policy-checked with the result written to an audit log, so which policy governed a given reply is recoverable; the trust portal carries audit logging and access monitoring as named product controls. Audit log export, and any export format, is not documented.

Sourceusefini.comread 2026-09-05

The platform is published as five named layers: Guardrails, LLM Supervisor, Knowledge and Action Modules, Live Feedback and Evaluation, and a Traceability Layer. Not one is a memory layer, and the only state named is conversation context feeding the LLM Supervisor within a request, held for the session rather than across sessions.

Past tickets are ingested as a knowledge source, and the adaptive learning loop trains on real-world interactions, which is learning absorbed into a model rather than state an agent retains. A Fini Labs blog post mentions policy-aware memory, but the architecture page lists the components and omits it, and the documentation at docs.usefini.com is gated.

Sourceusefini.com/product/platformread 2026-09-05

There is no documented choice of where Fini runs. The trust portal names Google Cloud Platform as the infrastructure and documents backups and disaster recovery but no region selection, EU-only processing, VPC, single-tenant or on-premises option, and the product site and pricing page carry no residency commitment either.

Sourcesecurity.usefini.comread 2026-09-05

One prebuilt agent, adopted as a unit, with no catalog behind it.

Sophie ships as a single named agent covering high-volume support journeys out of the box, with published worked examples spanning refunds, duplicate charges, subscription cancellation and product questions, and the platform is marketed as production-ready from day one with self-learning ingestion of docs, tickets and knowledge base requiring no manual tagging, so time to value does not depend on the customer authoring flows.

No template library, agent catalog or marketplace is documented, and the skill modules the agent routes to are internal architecture rather than assets a customer browses and selects.

Sourceusefini.com/product/platformread 2026-09-05

One agent runs across voice, chat and email on the same reasoning and policies, with real-time voice handling bookings end to end and intake arriving through the connected helpdesk, Slack or Discord. The site states 130 plus supported languages.

Sourceusefini.comread 2026-09-05

One disclosed model provider, with no customer choice. The trust portal's subprocessor list names OpenAI as the only model provider, with Supabase, AWS, Microsoft Azure and Google Cloud appearing as infrastructure rather than models. No model selection, routing or bring-your-own-model option is documented.

Sourcesecurity.usefini.comread 2026-09-05

A documented API for Fini's own platform exists, behind a customer gate. Fini publicly links Docs and an API Library from its site footer, pointing at docs.usefini.com and an authentication page under api-reference. The documentation site redirects to an access-code gate, so the surface is open only to customers and its scope is not public, and no SDK or MCP server is advertised.

Sourcedocs.usefini.com/api-reference/authenticationread 2026-09-05

A live optimization loop is documented and a pre-release gate is not. Live Feedback and Evaluation ships as layer four of the architecture, auto-validating outputs, detecting errors and contradictions, and learning from corrections over time, with Automated QA described as built-in feedback and validation that improves accuracy without human review, and real-time conflict resolution spotting and fixing inconsistencies in the knowledge base.

No customer-facing test environment, simulation harness or scored evaluation of a proposed change before it reaches customers is documented, and no readable comparable verdict is surfaced to the buyer; the published accuracy and resolution figures are vendor-reported. Every response is also LLM-reviewed for accuracy, tone and safety before sending.

Sourceusefini.com/product/platformread 2026-09-05

No browser control, hosted session or computer use appears anywhere on the site. Instead, the Take Action module triggers API calls into business systems, with Stripe and Zendesk named, and the Follow Process module runs internal flows through automated rule enforcement rather than by operating an interface.

Sourceusefini.com/product/platformread 2026-09-05