Run · Support deflection & triage
The same six questions, answering themselves.
An agent that reads the ticket, decides which of your six it is, pulls the order, the carrier scan and the subscription record, and resolves it inside your helpdesk — then hands your team everything on a boundary list you wrote.
-
70–80%
of DTC tickets are the same six question types
-
60–80%
deflection a well-scoped build reaches, on vendor-reported numbers
Category ranges. Every deflection figure published in this category is vendor-reported — the table below shows the spread.
The problem
It is 8:40 on a Tuesday and there are forty-one tickets in the inbox. Nine are some version of “where is my order”, four of those frozen boxes that left the 3PL on Friday and have not scanned since Saturday. Two people want to skip December. One wants to swap the vanilla for the chocolate. And one — exactly one — is a customer reporting a reaction to a reformulated batch who needs a person today. That ticket gets answered forty minutes after the easy nine, because a queue is chronological and knows no difference.
That is the whole problem in one screenshot. A small number of ticket types dominate the queue: order status, returns, sizing, product questions, shipping and discount codes account for something like 70–80% of DTC support volume. For fresh and frozen brands the order-status share is brutal — a perishable box with a delivery window generates an anxiety no shelf-stable brand ever sees.
Deflection is not about replacing the team. It is about the team never seeing the sixth “where is my order” of the morning — so the ticket that needed a person gets answered like it mattered.
Order status is a data problem the inbox is being asked to solve
Answering a WISMO ticket does not require judgement. It requires four things: the order in Shopify, the tracking state from the carrier, the dispatch record from the 3PL, and enough context to know what a scan that stalled on Saturday means for a box due Tuesday. Most brands hold all four in systems that do not talk to each other — so a person opens three tabs to compose a sentence they have written four hundred times. That is what the model is actually for here. Not writing prose: deciding which of six intents a sentence is, then reading four systems and answering from what they say.
A macro library is a filing system, not automation
Nearly every brand already has a macro library in Gorgias and treats it as the solved version of this. But a macro still needs a human to read the ticket, choose the right one, paste in the tracking number and press send. The saving is seconds per ticket; the cost is tickets. A bot bolted on top of the same macros has the same problem one layer up: it can quote your shipping policy, and it cannot look at this order.
Subscription questions arrive through the wrong door
“Skip my next box.” “Push the delivery to the 14th.” “Can I swap the flavour?” None of those are support questions. They are subscription changes typed into a support widget, and a generic bot answers them by linking to an account portal the customer will not open on a phone. An agent with authority in Recharge or Skio can simply do it — and a skip that takes one message is very often a cancellation that never happened.
And nobody has drawn the escalation boundary
The reason most deflection projects stall is not the model. It is that nobody decided, in writing, what the agent is not allowed to touch: adverse reactions, anything a clinician should be answering, health claims on a supplement, refunds above a threshold, any conversation where the customer has now said “cancel” twice. Without that list the agent either answers what it should not, or gets quietly throttled back until it is a search box with a personality.
What we build
Seven pieces, built inside the helpdesk you already run. The model does two jobs here — classify the ticket, and read the systems that answer it — and the boundary list decides where it stops. Everything lives in your accounts, and that list is a document you own.
- Ticket taxonomy from your own queue Ninety days of tickets pulled out of Gorgias, Zendesk or Intercom and classified by intent by a model rather than by whatever tag somebody applied at the time — then read back and corrected by a person, because every boundary and every routing rule downstream is built on this list. You see your own six before we automate any of them.
- WISMO agent, wired to the carrier Order lookup against Shopify, tracking state from the carrier, dispatch record from the 3PL, and cold-chain-aware language for a frozen box whose scan has stalled — the answer assembled out of live systems at the moment the question is asked, rather than typed by a person with three tabs open.
- Subscription-aware answers Read and write against Recharge, Skio or Smartrr, so “skip my next box”, “push it a week” and “swap the vanilla for the chocolate” get done inside the conversation instead of linked to a portal the customer will not open. The write access is scoped to those actions and nothing else, through tools we wrote — not a shared admin login.
- Escalation boundaries, in writing The intents the agent must hand to a human, named and enforced in configuration rather than requested in a prompt and hoped for: adverse reactions, health and medical questions, refunds above your threshold, and any conversation where the customer has used the word “cancel” twice. The list is a document you own and can change without us.
- Knowledge base the agent can actually answer from Shipping and delivery windows, storage and thaw instructions, ingredient and sourcing answers, and the returns policy as it actually operates rather than as it is written on the page nobody has edited since 2023. The agent answers out of that content, so a question about your formula is answered from your formula rather than from what a general-purpose model believes about supplements.
- Triage and routing for what is left Priority, tagging and assignment on the tickets that reach a person, driven by the same intent classification, so the one that needed a human today surfaces first instead of queueing behind nine order-status questions.
- Deflection reporting you can audit Resolution rate by intent, handoff rate, CSAT on deflected conversations, and live-agent tickets per 1,000 sessions — the one figure that does not improve when somebody starts closing conversations faster.
How the build runs
Four weeks from access to a partial launch, on a fixed scope and a fixed price. Nothing answers a customer before it has answered a few hundred historic tickets in front of your team.
-
01
Ticket audit
Ninety days of tickets classified by intent, with volume, first-response time and handling time against each one. You get your own six — and their share of the queue — before anything is built.
Week 1
-
02
Boundary & content design
What the agent answers, what it never answers, and what it has to check before it hands over. Written down, reviewed by you, and signed off before a line of configuration exists. Knowledge gaps get found here, not in production.
Week 1–2
-
03
Build & integration
The agent configured inside Gorgias, Zendesk or Intercom and wired into Shopify, your subscription platform and your 3PL, so it answers from live order state rather than from a paraphrase of your shipping page.
Week 2–4
-
04
Shadow mode & partial launch
The agent drafts, your team sends. Every reply reviewed against real tickets before one customer sees an automated answer, then a live launch on the two safest intents only.
Week 4
-
05
Widen & tune
Intents added one at a time, each with its own deflection rate, handoff rate and CSAT read against the week-one baseline. An intent that does not hold its numbers goes back to a human.
Ongoing
The numbers, and who published them
Two things are true about deflection benchmarks. The marketed rate and the published case-study floor are more than thirty points apart, and none of it has been measured by anyone without something to sell. We are publishing both anyway, because the gap is the useful part.
| Claim | Rate | Kind of evidence |
|---|---|---|
| Gorgias AI, as marketed | Up to 60% | Vendor marketing claim |
| Gorgias published case studies, as a band | 26–56% | Vendor-reported |
| Psycho Bunny (Gorgias) | 26% | Vendor-reported case study |
| Shinesty (Gorgias) | 54% | Vendor-reported case study |
| The Ridge (Gorgias) | 60% | Vendor-reported case study |
| Casper (agentic AI) | 74% | Vendor-reported case study |
| Wilson (agentic AI) | 77% | Vendor-reported case study |
| Edel Optics (agentic AI) | 79% | Vendor-reported case study |
| Independent third-party measurement | — | None published — metric to confirm |
Every figure above is published by a vendor, or by a vendor’s customer inside a case study the vendor hosts. There is no independent third-party measurement of ecommerce support deflection that we are willing to cite, which is why the last row is an em dash rather than a number. Note also that the Gorgias case-study band is published as 26–56%, while the named cases we can point at span 26–60% — so we list them individually rather than smoothing them into a range. Read the whole table as the spread of what has been reported, not as a forecast for your queue.
One Gorgias case study reports live-agent tickets falling from 2.72 to 0.48 per 1,000 sessions — an 82% reduction (vendor-reported). We prefer that unit to a deflection percentage, because a deflection rate improves the moment someone starts closing conversations the agent handled badly, and tickets per 1,000 sessions does not.
Where this sits in the market
The honest version, from our own category review. The helpdesk vendors are absorbing this capability, and a service page that pretended otherwise would not survive your first hour of research.
- Maturity
- Rapidly maturing
- Saturation
- Medium, rising
- Defensibility
- Medium
What stays defensible is not the model. It is the boundary list, the subscription integration and the reporting — the parts a platform cannot ship generically because they are specific to how your business operates.
What it costs
Published ranges. Where you land inside the build range depends on how many intents you want live, how many systems the agent has to read from, and whether your subscription platform exposes what we need through its API.
Revenue Recovery Audit — ticket mix and deflection review
$1,500–$3,000
Support automation build
$4,000–$12,000
Ongoing optimisation retainer
By arrangement
Helpdesk platform and AI resolutions, paid to your vendor
$750–$3,000+/mo
That last line is not ours. Gorgias currently bills roughly $1 per AI resolution on its published pricing; across the platforms the working range is $0.40–$1.20 per resolution, which for most brands in this category lands between $750 and $3,000+ a month, paid to your vendor. Model it before you commit — below a certain ticket volume the arithmetic does not work, and we would rather say so in the audit than after you have signed a build. The audit is credited in full against any build you go ahead with.
Where to go next
Questions
Isn’t Gorgias building this in already?
Yes, and so are Zendesk and Intercom. This capability is being absorbed into the helpdesks — we read the category as rapidly maturing, saturation medium and rising, defensibility medium, and we would rather say that on our own service page than have you find it out later. What the built-in tooling does not do is the three things that decide whether deflection holds: the escalation boundaries, which nobody writes for you; the subscription-aware answers, because the helpdesk has no authority to change a Recharge subscription; and the integration work that lets the agent answer “skip my next box” by actually skipping the box. Out of the box you get a good FAQ bot with your order-status data in it. The gap between that and 70% deflection is engineering.
Is this going to replace our support team?
No, and a build sold on that premise usually ends up throttled back within a quarter. Deflection is about your team never seeing the sixth “where is my order” of the morning, so the customer reporting a reaction to a reformulated batch gets answered like it mattered rather than forty minutes behind nine order-status questions. The volume goes down; the difficulty of what is left goes up. Most brands keep the same headcount and move it onto retention work, escalations and outbound.
What deflection rate will we actually get?
We will not put a number on it before the ticket audit, and you should treat anyone who does with suspicion. Published figures in this category run from 26% to 79%, and every one of them is vendor-reported — there is no independent third-party measurement of deflection rates in ecommerce support that we are willing to cite. What moves your number is the shape of your queue: a brand whose tickets are 55% WISMO has far more headroom than one whose tickets are mostly ingredient and formulation questions. The audit tells you which of those you are.
Which helpdesk do you build in?
Gorgias, Zendesk or Intercom — whichever you already run. We do not migrate you as a precondition. The agent design, the boundary list and the integrations are the same problem in each; what changes is where the configuration lives and which of the three the platform will let us do natively rather than through the API.
Can the agent really change a subscription, or does it just send a link?
It can change it, and that is most of the point. Wired into Recharge, Skio or Smartrr, the agent can skip the next box, shift a delivery date, swap a flavour or variant, and update a shipping address inside the conversation. We gate it deliberately: discount codes, refunds above your threshold and anything that reads as a second cancellation attempt go to a human. A skip that takes one message is very often a cancellation that did not happen, which is why this sits closer to retention than to support.
What does the AI itself cost to run?
Gorgias currently bills roughly $1 per AI resolution on its published pricing. Across the platforms the working range is $0.40–$1.20 per resolution, which for most brands in this category lands somewhere between $750 and $3,000+ a month, paid to your vendor rather than to us. It is worth modelling before you commit, because below a certain ticket volume the arithmetic simply does not work — and if that is where you are, we will tell you in the audit rather than after you have signed a build.
What happens when it gets something wrong?
It will, which is why nothing launches without shadow mode first — the agent drafts, your team sends, and every draft is reviewed against real tickets before a customer sees an automated reply. After launch, the boundary list is the safety net: intents outside it are handed to a person with the conversation history attached, not answered badly. Anything that slips through gets read weekly, and the fix is usually a boundary or a knowledge gap rather than the model.
We’re fresh or frozen. Is cold-chain WISMO actually automatable?
The bulk of it is. Most of that volume is a customer with a delivery window, a tracking number and an anxiety — the agent can pull the carrier scan, the dispatch record and the delivery window and answer in one message, including the awkward case where the scan has stalled over a weekend. What we route to a human is the arrived-warm ticket, anything involving a replacement of perishable stock, and any conversation with a health question in it. See the subscription brands page for how the cold-chain layer sits on top of the normal subscription mechanics.
Find out what you’re losing.
Before you commit to anything, we tell you exactly what you’re losing and what it costs to stop it. Two weeks. Fixed fee. Credited in full against any build you go ahead with.
- Fee
- $1,500–$3,000, fixed
- Duration
- Two weeks
- Credited
- In full, against any build
- You supply
- Read access + one 45-minute call