Brainfish targets the part of a support queue that ordinary AI agents hand back to humans. Its own framing is that roughly 20–30% of tickets can be answered from a help article, and any AI agent can do that — the other 70–80% require someone to check a system and take an action. Brainfish pairs an AI layer that understands the customer and writes the reply with a rules layer that checks your systems, applies your policy, and executes the outcome.
The distinguishing design choice is the separation of conversation from decision. You describe a process in plain English and Brainfish compiles it into a workflow that runs the same way every time, so refund caps and approval thresholds live in the workflow rather than in a prompt the model can improvise around. Typical automated flows include verifying a customer, checking a refund policy window, issuing a Stripe refund, cancelling a plan in Recurly or Salesforce, resetting identity access through Okta or Duo, and replying before closing the ticket.
Brainfish connects to Zendesk, Salesforce, Intercom, Stripe, Duo, Splunk "and anything else with an API." Support operations teams own the workflows end to end and can see every run, with each step, system, and source logged for audit. Security documentation covers SOC 2 Type II, ISO 27001, and GDPR, with US, EU, or AU data residency, SSO/SAML, and a published commitment that customer data is never used to train models. Pricing is not published; Brainfish routes prospects to a demo where a workflow is built live against their own systems.
