FruxonGive the queue to an agent. Keep the judgment calls.
Support, success and onboarding — the work where somebody is waiting and a wrong answer costs you the relationship. The agent works your helpdesk and the systems behind it, and asks the person on your team whose call it is before it invents a policy.
Deflection is the easy half. The other half is the half that costs you.
Every vendor in this category will quote you a resolution rate. None of them will tell you what happens on the conversation the agent shouldn't have answered — the policy exception nobody wrote down, the account on a bespoke contract, the credit that needed a person. Alone, an agent has exactly one move there: answer confidently and hope. No model knows what your ops lead decided on Tuesday, and buying a better one doesn't close that gap.
What it does when it reaches the end of what it knows.
A gift bought during a promotion, returned on day 44. The written policy says 30 days and says nothing about promotions. This is the shape of every conversation that decides whether an agent is allowed to keep working.
Customer
It was a gift, bought in the November sale. Your site says 30 days — is that really it?
The agent
Knowledge
Pulls the order, the customer's history and the return policy. No passage covers a gift bought inside a promotional window. Confidence below threshold — it doesn't answer.
The agent
Directory
Returns and exceptions are Maya's lane, so the question goes to Maya on Slack — not to whoever is on the ticket queue. Dana is next on the ladder if nothing comes back in fifteen minutes.
Maya · Ops
Slack
Promo-window gifts get 60 days, not 30. Approve it, but refund to store credit.
The agent
Stripe · Email
Issues the store credit, writes the note to the ticket, and replies on the thread the customer started. Elapsed: a slightly longer wait.
You, later
Inbox
One logged gap: the return policy says nothing about promotions. Write the paragraph and it never asks again.
Nothing above is a feature you turn on. It is what the agent does by default when it doesn't know — which is the only behaviour that makes the rest of the queue safe to hand over.
Three places to point it first.
Customer Support
Works the queue in your helpdesk, acts in the systems behind it, and asks before it invents a policy.
Customer Success & Renewals
Watches every account on the same schedule and hands the at-risk ones to the AE with the reasoning attached.
Onboarding & Implementation
Chases what the customer still owes you, provisions what it's allowed to, and escalates the scoping calls.
One platform, one directory, one set of connected systems. An agent doing renewals escalates to a different person than one doing support — that's configuration, not a different product.
It acts, it stops, it asks. In that order.
It picks the work up where it already lives
A ticket in Zendesk, a chat in your product, a message on Slack, Teams or WhatsApp, a renewal date in Salesforce. The customer's history and the passages of your policy that bear on this case are retrieved at the step, not pasted into a prompt and carried on every reply.
It acts in the systems behind the conversation
Issues the refund in Stripe, opens the replacement order in Shopify, writes the note to the CRM, queries your Postgres for the shipment. Not “sends a notification for someone else to action.” Read-only until you widen the lane.
It stops when it isn't sure
A policy exception, an account flagged VIP, an amount over the ceiling you set. It neither guesses nor dumps the thread back on you, and the customer sees a slightly longer wait rather than a handover queue.
It asks the person who'd actually know
Routed by expertise rather than by queue, on the channel that person already works in. One line back and the reply goes out — with the whole exchange, question and answer and who approved it, on the record.
When it isn't sure
The escalation is the feature, not the failure.
You introduce the agent to the team once — who's on it, what each of them owns, where to reach them. From then on a question it can't answer goes to the person whose call it is, not into a queue nobody watches. A new hire wouldn't know your return policy either; on day one they'd turn around and ask, and nobody would think less of them for it.
Routed by expertise, not by queue — set once on the person, not re-tagged on every agent that might need them
Nobody answers? It falls through to the next person on the ladder, then to a safe default. Nothing hangs on someone's lunch break.
Out of its depth entirely? A person takes the conversation mid-sentence with the full history and hands it back when it's resolved
Every question it had to escalate is a gap you can see — close it and it stops asking
What you'd otherwise do about this queue.
Three of these are reasonable choices and two of them are probably already running somewhere in your stack. The column that matters is what each one leaves on the table.
What that leaves
What this does instead
A deflection bot
Tuned for the contacts it can close alone. On the ones it can't, the good outcome is a handover queue and the bad one is a confident wrong answer nobody catches until the review.
It escalates to a named person by expertise and then finishes the reply itself, so the customer waits minutes rather than starting again with someone new.
Your helpdesk's own AI add-on
It knows your tickets and your macros. It doesn't know your Stripe, your Postgres, or who is allowed to approve a credit over five hundred.
It runs inside the same helpdesk and reaches the systems behind it, with a per-action ceiling on what it can do alone.
An outsourced team
Someone always answers, and none of them can change anything in your systems without raising a ticket back to you. Cost per contact goes down and stays down at exactly one level.
The routine contacts cost what a run costs, and the judgment calls still go to someone on your payroll who already knows the answer.
Building it in-house
The agent is a fortnight. The directory, the escalation ladder, the approval gates, the traces, the cost ceiling and the eval harness are the next two quarters — and they're the part that decides whether it ships.
All of that is the platform. Your team writes the agent and the prompts; nobody writes a retry queue.
Building it yourself is the one worth pricing properly — there's a page for that at /compare/diy.
It works inside your helpdesk, not instead of it.
Connect a system once for the whole organization and every agent can use it from then on — with credentials held at the org and referenced by name, never pasted into a prompt.
Slack
GitHub
Jira
Google Drive
Salesforce
MongoDB
Notion
Linear
Grafana
Discord
Google Chat
Confluence
PostgreSQL
HubSpot
Airtable
Shopify
Stripe
Datadog
Sentry
GCP
BigQuery
Slack
GitHub
Jira
Google Drive
Salesforce
MongoDB
Notion
Linear
Grafana
Discord
Google Chat
Confluence
PostgreSQL
HubSpot
Airtable
Shopify
Stripe
Datadog
Sentry
GCP
BigQuery
Google Calendar
Gmail
Mixpanel
Monday
MySQL
SAP
Zendesk
Zoho CRM
Google Maps
Google Ads
Coralogix
Telegram
Apollo
Mailchimp
Calendly
Redis
Supabase
GCP Logging
gVisor
Loops
Typeform
Google Calendar
Gmail
Mixpanel
Monday
MySQL
SAP
Zendesk
Zoho CRM
Google Maps
Google Ads
Coralogix
Telegram
Apollo
Mailchimp
Calendly
Redis
Supabase
GCP Logging
gVisor
Loops
Typeform
Don't see yours? Import tools from any MCP server.
Start it narrow. Widen it when it earns it.
You are never choosing between fully autonomous and useless on day one. The lane is a setting, and the escalation log tells you when it's safe to move it.
Week one
Read-only, on one channel
Sandbox Mode: it reads production and its writes are captured instead of sent. You watch a week of what it would have replied, and no customer sees any of it.
Week two
One job, and permission to ask about everything
Turn it on for a single queue with a low ceiling and a liberal escalation policy. It will ask more than you'd like. That's the point — every question is a gap in what you wrote down.
Weeks three and four
Close the gaps, widen the lane
Add the passages it kept asking about, approve the actions you're comfortable with, raise the ceiling. Escalation rate falls because the knowledge improved, not because the agent got braver.
The questions support leaders actually ask
The ones that come up in the first call, answered before it.
How is this different from a deflection bot?
A deflection bot is measured on the conversations it closes without a human. That metric quietly rewards answering things it shouldn't. Fruxon agents are built to notice the boundary and ask a named person on your team before replying — so the number that goes down over time is escalations, and the number that goes up is what it can safely handle.
Do we have to replace our helpdesk?
No. The agent works inside the helpdesk you already run — it picks tickets up there, sets status there, and leaves the record there. It also reaches the systems behind the conversation, which is the part a helpdesk-native assistant generally can't do.
What happens at 2am when nobody answers the escalation?
It falls through: next person on the ladder, then a safe default you set per action — usually hold the reply and tell the customer a person is looking at it. Nothing hangs indefinitely on one person being awake, and nothing gets answered by guessing because the ladder ran out.
Won't my team be buried in questions from an agent?
In the first fortnight it asks a lot, and every question is a gap in your documented policy. Close the gap and that question never comes back. There are urgency caps, timeouts and a relevance gate so it doesn't interrupt for things it could look up, and the escalation log shows you exactly what it's asking about.
How long before it's actually handling tickets?
Connecting a helpdesk and a payment system is an afternoon. The honest answer for a first queue in production is a couple of weeks, most of which is you watching Sandbox Mode output and deciding where the ceiling sits.
In the queue
Invoices, internal tickets, order exceptions, vendor files — the work in the queue.
Finance Operations
Matches the PO, checks the terms against the contract, clears what reconciles and flags what doesn't.
IT & Employee Support
Access requests, provisioning, and the same question asked forty ways — with a gate on anything that grants permission.