Customer support automation: 61% of tickets resolved without a human.
A support agent is only useful if it knows when to stop. Veyra's answers 61% of tickets and hands the rest over with the account already pulled.
The same work, twice
- 1Read the ticket and work out the intentSupport · 2m
- 2Open the account in the admin toolSupport · 2m
- 3Check the billing record and planSupport · 3m
- 4Search the help centre for the policySupport · 4m
- 5Ask a colleague whether the exception appliesSupport · 6m
- 6Write and send the replySupport · 5m
- 7Tag the ticket and update the macroSupport · 2m
The shape of the run, drawn from the Veyra build. Your intents, thresholds and volumes come out of the audit week.
Why this queue costs what it costs
Veyra's support queue grew with every new customer. Eighty per cent of tickets were the same six questions, but each one still needed a person to read the account, check a billing record and write a reply. The plan on the table was to double the support team.
We built a triage agent inside the existing helpdesk instead. It reads the ticket, pulls the account and billing context, and answers directly when it can cite a policy or a record. Anything ambiguous routes to a human with the context already attached, and anything below the confidence threshold gets no answer at all.
Ten weeks from the scoping call, 61% of tickets closed without a person, median first response fell from four hours to eight minutes, cost per ticket dropped 34%, and CSAT moved up from 4.4 to 4.5 out of five.
Deflection without the part customers hate
Everyone has met the support bot that answers a specific question with a link to a general article. It fails because it is optimising for containment rather than for resolution, and because it has no access to the account in front of it.
The difference in a working system is evidence. The agent answers when it can point at a policy line or a record in the account, and it says nothing when it cannot. Containment is an outcome of that rule, not the goal it is tuned against.
The second effect matters as much as the first: the tickets that still reach a person arrive with the account, plan, billing history and a draft attached, so they take less time than they used to.
The threshold is a dial your team owns
Every routing decision is logged with its confidence and the evidence behind it. That gives the support lead a real instrument: they can see the tickets that sat just under the line, read what the agent would have said, and move the threshold with knowledge rather than instinct.
It also makes the honest trade explicit. Push the threshold up and deflection falls while accuracy rises; push it down and the reverse. That decision belongs to whoever owns CSAT, and after handover they can change it without us.
What an agent must never do
Refunds, cancellations, plan changes, data deletion and anything else that is hard to reverse always require a person, regardless of confidence. The agent may prepare the action and pre-fill the form; it does not press the button.
That rule costs a few points of deflection and buys the thing that keeps the project alive: no customer has an irreversible action taken by a system that was 91% sure.
It lives inside the helpdesk you already use
Veyra's ran inside Zendesk. The pattern works with Intercom, Front, HubSpot and Freshdesk, and with a homegrown desk that exposes an API. What matters is that agents keep their existing tool, their macros and their reporting.
A separate console that support staff have to check is the most common way these projects quietly die. Everything the agent does is visible in the ticket itself, in the same place a colleague's note would be.
Measuring it honestly
Deflection rate alone is a vanity metric, because a system can close tickets by frustrating people into giving up. The set we report is deflection, reopen rate, CSAT on automated versus human replies, median first response and cost per ticket.
Veyra's CSAT went up slightly rather than down, which is the number we would look at first in any review. If reopen rate climbs while deflection climbs, the threshold is wrong and the fix is a dial, not a rebuild.
Cost, timeline and payback
A production support agent with evaluations, logging and rollout is a $50,000 to $150,000 project over eight to twelve weeks. A narrower first version, covering the two or three highest-volume intents inside your existing desk, can start as an automation pilot from $4,000.
Veyra passed payback in month five on ticket cost alone, before counting the six hires they did not make. The arithmetic to check it yourself is tickets per month, minutes per ticket and loaded hourly cost.
How the build runs
Read the queue
A few hundred real tickets, clustered by intent, with handling times and the policies each one relies on.
Build the agent and the evaluation set
Retrieval over your policies and account data, plus a graded set of historical tickets with known-correct answers.
Shadow, then release by intent
Drafts go to agents first. Intents graduate to automatic replies one at a time, each with its own threshold.
Hand over the dials
Thresholds, intents and reporting owned by your support lead, with documentation and 30 days of support.
Before and after, in numbers
Veyra, 90 days in production against the previous quarter, exported from their helpdesk.
Where we have built this
Related reading
AI automation
For the operational half of the queue: refunds prep, order status, account admin.
Questions before a pilot
It depends on how much of it is repeatable and how much of your policy is written down. Veyra reached 61% because 80% of their tickets were six questions. A queue full of bespoke technical investigation lands far lower, and we would tell you that after reading a sample rather than after the build.
It only answers when it can cite a policy line or a record from the account, and below-threshold tickets get no answer at all. Every reply is logged with the evidence it used, so a wrong answer is traceable to a specific source rather than to a mood.
Yes. Automated replies are identified as automated, with an obvious route to a person in the same message. Hiding it is both a trust problem and, in several jurisdictions, a compliance one.
At Veyra nobody was let go and six planned hires were not needed. The team spends its time on the tickets that need judgement, which is also the part of the job people prefer. If your plan is headcount reduction rather than absorbing growth, say so at the scoping call so the business case is honest.
Yes, and the practical limit is your policy documentation rather than the model. If your help centre only exists in English, replies in other languages are translations of an English source, which is usually acceptable for support and rarely acceptable for legal or billing terms.
By intent, not all at once. Drafts go to agents first, then the highest-volume intent graduates to automatic replies with its own threshold, then the next. A bad week comes from switching the whole queue over on a Monday.
Want to see this run on your own customer support?
Bring twenty real examples to the scoping call. The audit week counts the hours, writes the rules and comes back with a fixed price, and you keep the process map whatever you decide.