Brainfish

Across many classes of tools, the agent acts rather than only retrieving: Brainfish states plainly that it reads from and writes to the tools a team already runs, and the named set spans helpdesk (Zendesk, Intercom, Help Scout), CRM (Salesforce, HubSpot), messaging (Slack, Microsoft Teams), documentation (Notion, Confluence, Mintlify, ReadMe), storage (Google Drive), mail (Gmail), issue tracking (Linear, Jira), revenue intelligence (Gong) and video (Loom), plus OpenAPI specs and any HTTP source.

Sourcebrainfishai.comread 2026-09-05

Beyond answering, an AI Actions add-on automates workflows and handles advanced support tasks through custom API integrations, tool calls, MCP and callbacks, configured with no code or low code and purchased in blocks of five actions, and a Workflows surface is marked as a new capability, so execution goes further than retrieve, answer and escalate. No multi agent execution, branching runtime or agent to agent handoff is documented, and what the Workflows surface does is not described.

Sourcebrainfishai.com/pricingread 2026-09-05

A maintained knowledge layer rather than context assembled per question: Brainfish ingests from Notion, Confluence, Google Drive, Loom, Zendesk, Salesforce Knowledge, Help Scout, Helpjuice, Mintlify, ReadMe, OpenAPI specs, websites and any HTTP source into a single store, keeps it current by flagging stale articles and drafting new ones from calls and Slack threads, and routes the slice each agent needs at the moment of a question. Every answer is drawn only from approved sources and cites the exact documents behind it with a confidence score.

Sourcebrainfishai.com/pricingread 2026-09-05

Brainfish offers two mechanisms, neither of them an approval surface of its own. Grounding is a constraint rather than a gate: the agent answers only from approved sources and cites them, which bounds what it can say but reviews nothing.

Autonomy is set by a trust threshold, with the agent sending directly for cases a team trusts and escalating the rest to a person with full conversation context under configurable rules, which routes rather than holds for sign-off.

Where a human does approve, it is on a draft reply inside Zendesk or Intercom, and that gate belongs to the helpdesk rather than to Brainfish.

Sourcebrainfishai.comread 2026-09-05

A Trust and Safety center at trust.brainfi.sh, linked from the site header and footer, sets out both a compliance posture and access controls. On compliance it lists SOC 2 Type II, audited annually by an independent third party, ISO/IEC 27001:2022 certification, GDPR with right to delete built in, and a stated commitment that models are not trained on customer data.

Access controls include single sign-on with Okta and Azure AD named as supported identity providers, private tenants, and customer-managed encryption keys for enterprise, and sensitive fields are obfuscated before any snapshot, with snapshots processed then discarded by default. The route to obtaining the SOC 2 report is not published.

Sourcetrust.brainfi.shread 2026-09-05

Brainfish reports resolution rate, deflection and customer satisfaction, surfaces the questions its knowledge layer is failing to answer, and cites the exact source documents behind every answer with a confidence score, which is genuine answer provenance: a reader can see which document produced a given reply. That is provenance for the output, not reconstruction of the run; no step-level decision trace, tool-call log or configuration audit trail is documented.

Sourcebrainfishai.comread 2026-09-05

No memory layer with a stated scope or lifetime is documented. Answers are personalized with live account, product and user context supplied at the moment of the question, which is per-request assembly, and the pricing FAQ states that sensitive fields are obfuscated before any snapshot is taken and that snapshots are processed and discarded by default, which describes state deliberately not retained. The persistent store the product does keep is the knowledge layer, which holds the team's content rather than agent memory.

Sourcebrainfishai.com/pricingread 2026-09-05

Hosting runs in the United States, European Union and Australia, with the region chosen by the customer. Brainfish states that data stays in-region, that EU hosting carries no quota, SLA or support difference from the US region, and that it does not egress to other regions for inference or storage. Private tenants and customer-managed encryption keys are offered on top. No self hosted or on premises build is offered, and EU hosting is gated to Platform and above.

Sourcebrainfishai.com/pricingread 2026-09-05

No prebuilt agent, template gallery, cloneable starter or marketplace of packaged assets is documented. The product ships as a single configurable support agent pointed at the customer's own knowledge, and the onboarding pitch is a working agent built against your own content rather than one chosen from a set.

Sourcebrainfishai.comread 2026-09-05

Inbound conversation queues bring work to the agent as well as channels: Intercom live conversations and Zendesk tickets flow in from the helpdesk the team already runs, alongside asking in Slack, answers in Microsoft Teams, the embedded chat widget, the AI help center, an in-product agent and email, with proactive prompts surfaced before a user asks. No scheduled or cron style trigger is documented.

Sourcebrainfishai.comread 2026-09-05

No model selection, routing or bring-your-own-model control appears anywhere in the customer-facing configuration surface, and no provider is named; the vendor states only that models are not trained on customer data.

Sourcebrainfishai.comread 2026-09-05

Brainfish ships a native Model Context Protocol server so engineering teams can plug it into their own AI surfaces, with Claude named as a supported client, and the AI Actions add-on is sold on APIs, tool calls, MCP and callbacks.

The APIs in that description are the customer's own systems being called by the agent, and no public REST API reference or SDK for Brainfish's own platform is published; the developer resources published are a changelog and a status page.

ReadMe, Mintlify and OpenAPI specs appear on the site as knowledge sources Brainfish ingests from, not as its own developer documentation.

Sourcebrainfishai.com/pricingread 2026-09-05

Live performance feeds back into content: the product surfaces the questions its knowledge layer is failing to answer and recommends improvements to specific agent responses, an optimization loop after deployment measured against resolution and deflection. The artifact being improved is the knowledge, not the agent's configuration, and no test suite, conversation simulation, scenario set or scored evaluation of the agent itself is documented.

Sourcebrainfishai.comread 2026-09-05

Its route out to other systems is programmatic: Brainfish states that it reads from and writes to the tools a team already runs, naming those tools and their integrations. No browser control, hosted session, virtual desktop or computer use is documented.

Sourcebrainfishai.comread 2026-09-05