Automate support engineering work across your internal systems.
Any API, database or admin tool. Written as real code, so every run follows the same steps, within your limits.
Go past the help center, where other AI agents stop.
of your queue can be answered from a help article. Any AI agent can handle these.
needs someone to check a system and take action. Other AI agents pass these to your team. Brainfish resolves them.
Browse the workflows“Refund three of my four subscriptions.”
“Cancel our plan and downgrade the team.”
“What is this charge on my card?”
“Locked out after changing phones.”
“Delete our account and all our data.”
“Every export has failed since 9am.”
Run every workflow on your stack, the same way every time.
Your own APIs, databases over VPN, MCP, legacy admin tools. If it has an endpoint, a workflow can use it.
Real code, not a prompt. Same input, same steps, same result, every run.
Refund caps and approval thresholds live in the code. Outside policy, it stops for a person.
Simulate against real tickets, version every change, roll back any time. Every run is logged.
Start from a workflow that already works.
Choose from 27 templates, then run yours on your own systems.
Refund request → verify identity, evaluate every subscription against policy, refund the eligible ones and explain the rest in the same reply.
Billing ticket → classify against the refund-type decision tree, look up subscription state, execute or route to the right tier.
Duplicate subscriptions → disambiguate with a dropdown built live from the API response, confirm intent, merge, reconcile billing.
Unrecognised charge → retrieve the transaction, resolve the plan and period, check proration and trial conversion, answer in plain language.
billing paymentsDeletion request → verify identity, run pre-deletion checks for active subscriptions, balances and retention obligations, execute or return the blocker.
Ticket created → read the content rather than the submitted category, classify true intent, answer or reclassify and route.
Case created → score genuine vs junk, bucket into general, reseller or technical, apply bucket-specific handling.
Classified case → match to product pillar, check entitlement and availability, route, and attach the research the agent would have done.
Ticket arrives → read customer data, diagnose against known failure patterns, send the first reply, score the output, tune on supervisor feedback.
Technical case → identify the environment, pull the logs, run first-pass analysis against known signatures, attach the diagnosis before a human opens it.
Question in the widget → fetch the transaction and decline reason across systems, analyse against rules and history, answer live in the same session.
widget rules engine ledgerKnown error detected → surface the workaround in-product before contact, auto-create the engineering ticket, pre-investigate in monitoring.
Widget opens on a page → the workflow fires automatically and answers the question the customer was about to ask.
widget SDK any backendEntitlement request → check the requester against the authorized-contact list on the account record, confirm tier, allow self-serve or route.
Can't authenticate → verify by text, confirm against the identity record, issue the enrolment reset link or release the risk lock — on chat and phone alike.
Locked-out customer → collect evidence, hand off to the identity-verification provider by API, act on the verdict, log the decision.
“Why can’t I access my account?” → look up suspension status and the actual reason, return the specific unblock path rather than a generic article.
operational CRMAny ticket → attempt end-to-end resolution; where it can't complete, draft the agent note and customer reply in-ticket with every lookup already done and cited.
Agent opens a case → search every knowledge location at once, autocomplete the reply, surface the internal doc they'd otherwise hunt for.
Workflow can't or shouldn't finish → choose the handover — human, ticket, embedded page, email, phone — and carry the full context and everything already checked.
Daily schedule → pull tickets from ITSM and the engineering tracker plus the related docs, analyse, prioritise, publish a bottleneck report to Slack or email.
Batch of new reviews → attach store context, triage by type, draft a regionally native reply, route for approval, escalate incidents without replying.
review APIs store directoryNew customer signs → send structured pre-work, read the replies, extract the data, map it to named CRM fields via a defined API call, human confirms before write.
Document arrives → extract, apply business rules, weight priority by value and due date, assign to queue, release the worker on approval waits.
Power Automate RPA fleet on-premBatch of call recordings → transcribe, find where documented knowledge diverges from what agents actually said, propose the articles that close the gap.
recordings knowledge baseRelease ships → cross-reference the corpus, identify every affected article, draft the updates, route through review, publish to the CMS.
release notesEvery deployment → build the source-of-truth list, drop no-index and low-traffic pages, push to the catalog by API, verify the sync landed, alert on drift.
Catalog APIBuild a workflow in plain English.
Managed through chat, readable as code, priced per run.
Go live in hours, not sprints.
Questions, answered.
What is a Brainfish workflow?
A workflow resolves a ticket by doing something in your systems rather than answering from an article. It might issue a refund in Stripe, release a risk lock in your identity provider or pull logs from Splunk. Each one runs as code against your real APIs, stops for a person when it hits one of your limits, and logs every step.
How is this different from knowledge deflection?
Teams tell us articles answer 20 to 30% of their tickets, which is worth having. The rest need someone to check a system: why was I charged twice, cancel my second subscription, I can't log in. Workflows handle those.
Can one workflow span several systems?
Yes, most do. A refund workflow might read the ticket in Zendesk, find the customer in your CRM, check their subscription in your billing platform and issue the refund in Stripe, all in one run. The logic is written as code your engineers can read.
How long does a workflow take to build?
About 20 minutes from a plain-English description to a first running draft. A complex workflow that's ready for production usually takes around two hours.
How are workflows priced?
A flat monthly fee for each workflow in production, about $60 on average. Every run is priced too, so you can compare your cost per resolved ticket with what you pay per resolution today.
What stops a workflow doing the wrong thing?
Your rules. You write them in plain English, for example: refund the most recent payment if it's under $50. Anything over the limit or outside policy stops and waits for a person. Each workflow can only make the calls it was built for, you can test it against real past tickets before it goes live, and every change is versioned so you can roll back.
Meet your security team's requirements from day one.
Regional data residency, SSO, a record of every run and a source on every answer. Your data is never used to train models.
SOC 2 Type II
Audited every year by an independent third party.
ISO 27001
Certified information security management.
GDPR
EU data residency, with the right to delete built in.
Data residency
US, EU or AU hosting. Your data stays in region.
Every run on the record
Each step, system and source is logged for review and audit.
Limits AI can't override
Refund caps and approval thresholds live in the workflow, not the prompt.
SSO and SAML
Your team signs in through your identity provider.
No training on your data
Your data is never used to train models.
Bring us your trickiest support ticket.
We'll build the workflow on your systems before you spend anything.
Book a demoNo engineering sprint required.