slack · #cx-questions
"Just refund it, she's a repeat customer and the box was crushed."
For Shopify support teams · concierge pilot
Let AI draft refund decisions without guessing your policy.
A two-week concierge pilot for Shopify teams with messy refund and return tickets. We draft replies and Shopify actions in Slack, your team approves everything, and every edit becomes a policy rule.
Send 10 weird tickets
plus your current refund policy and macros
We map the judgment
what to draft, escalate, refund, reship, or reject
You approve every reply
no customer-facing automation until you trust the file
The failure mode
Refund rules live everywhere except where agents can use them.
The real policy is scattered across old tickets, Slack threads, manager memory, helpdesk macros, and one-off exceptions. Generic AI support tools can draft a reply — but they still have to guess how this merchant actually decides.
gorgias macro
if (tag:vip) { full_refund; skip_RMA; } else { offer_15pct; }
manager memory
"We don't do returns on clearance — unless they bought during the warehouse sale."
ticket #48219
"Approved $90 refund, no return. Lost in transit. Mark as one-off."
post-it on monitor
Damaged in transit → refund + reship. Don't make them ship it back.
What's in the file
A policy file is just three things, kept honest by humans.
Not a black-box prompt. A reviewable document an operator can read, edit, and defend. Every line has provenance — who wrote it, when, and from which ticket.
Plain-English rules with thresholds.
Refund windows, exclusions, escalation paths, approval limits, fraud signals.
Past tickets become decision examples.
So the agent models the judgment behind the reply — not just the words.
Every human correction becomes a proposal.
Append-only. Repeated edits surface as plain-English rule changes you can accept.
The approval loop
Every draft goes through a human first.
Policyfile reads the ticket, drafts the reply and the operational action against the current policy file, and posts both in Slack. Your team approves, edits, or rejects. Nothing reaches the customer until you say so.
- Drafts the customer reply and the proposed Shopify action
- Cites which rule and example each decision came from
- Flags edge cases the file can't yet handle — for a human, not a guess
- Turns repeated corrections into rule proposals you can accept in one click
No auto-send
No helpdesk migration
Human approval by default
Two weeks · concierge
We prove the workflow manually before asking you to trust automation.
Send the messy inputs.
You share your refund policy, macros, and 10 recent edge-case tickets. We define what should be drafted, escalated, refunded, reshipped, or rejected.
Approve every draft in Slack.
We propose the customer reply and Shopify action for each refund or return ticket. Your team approves, edits, or rejects before anything reaches the customer.
Decide from the numbers.
You get the policy file, the edits we learned from, and a short report on approval rate, handling time, and inconsistent refund decisions caught.
What we measure
By week two, you should know if this saves real support time.
We report against three practical numbers from day one. If the workflow is not getting faster or safer, there is no reason to keep going.
01
Approval without edit
The share of drafts your team approves untouched. The line should move up from week one to week two.
02
Minutes per refund ticket
The time from ticket review to approved customer reply and Shopify action.
03
Bad decisions prevented
Over-refunds, inconsistent exceptions, and escalation misses caught before they reach the customer.
FAQ
Built for trust before autonomy.
Does this send replies to customers automatically?
No — not during the pilot, and not by default after. Policyfile drafts the reply and the proposed action; your team approves or edits before anything reaches the customer. Autonomy is a setting you turn on per rule, when you're ready.
Which helpdesks and stacks do you support first?
The first wedge is Shopify brands using Gorgias or Zendesk, with Slack as the approval surface. If you're on something else, talk to us — but expect the first integration to be hands-on.
Is this only for refunds?
Refunds and returns are the first workflow because the feedback loop is fast, the rules are messy, and the impact is measurable. The underlying primitive — a reviewable policy file with rules, examples, and edits — extends to any repeated judgment-heavy decision.
What do you actually need from us to start?
A short workflow review, a link to your public refund policy, a handful of recent edge-case tickets, and one person who can approve proposed decisions in Slack during the pilot. That's it.
What does the file look like once it's "done"?
It never is — and that's the point. The file keeps capturing edits as your business changes (new SKUs, seasonal policies, fraud patterns). What changes is that the agent stops needing humans for the routine 70% and only escalates the ambiguous tail.
Now onboarding design partners
Have refund decisions that still depend on judgment?
20-minute workflow review · no obligation · we will tell you if it is not a fit