The half of support that is the same question again.

The problem

Open any support inbox on a Monday and sort it by subject. A third of it is the same five questions. Where is my order. How do I change the booking. Does this fit the older model. What are your hours over Easter. Can I have a copy of the invoice. Each takes a competent person two or three minutes, which is exactly why nobody has ever treated it as a problem worth solving. The problem is not the two minutes. It is that the two minutes arrive sixty times, in no useful order, in between the things that actually needed a human.

So the hard tickets queue behind the easy ones. The customer with a real fault, who is the one at risk of leaving, sits in the same line as somebody who wants a VAT receipt, and the line is ordered by when things showed up. Everyone who has worked in support knows this. Nobody can fix it from inside the queue, because triaging it properly would itself take all morning.

The standard answer is to build a help centre and point people at it. The evidence there is bleak: Gartner found only 14% of service issues are fully resolved in self-service, and an earlier survey put the share of customers who felt they had completely solved their own problem at 9%. People genuinely do try. They search, they find three articles, none of which quite describes their situation, and then they email you anyway, slightly crosser than when they started.

It fails because a help centre hands people documents when what they wanted was an answer. Take does this fit the older model. The real answer depends on which model they bought, which is in your order system, and on a compatibility note living in a spreadsheet. No article covers that. Something that can read both and write one sentence does, and that is a fundamentally different machine from a search box.

How it works

Six things happen between a message arriving and somebody being satisfied, and the sixth is what stops the system quietly getting worse.

1. 01Every message is read and classified before anything elseWhat is this about, how urgent is it, how annoyed is the person, and is it a question, a fault, a complaint or a cancellation. Classification happens first because the routing decisions all depend on it, and because a cancellation that gets handled as a general enquiry is the most expensive misfile available.
2. 02The answerable ones are answered, with the source attachedDrawn from your own material: the policies, the product notes, the previous tickets that were resolved well, and live data from the order or booking system where the question is about this customer rather than about the world. Every answer carries where it came from, so a colleague reviewing it can check in a click rather than taking it on trust.
3. 03Confidence decides whether it sends or draftsHigh confidence on a routine question sends. Anything lower arrives as a draft on an agent's screen with the sources shown, which they send, edit or discard in a few seconds. This threshold is the single most important setting in the system, you own it, and it is normal to start it high and lower it as the evidence comes in rather than the other way round.
4. 04Everything else is routed with its context already gatheredTo the right person or team, with the account, the order history, the previous tickets, the entitlements and a two-line summary attached. The work being removed here is not the reply, it is the four minutes of looking things up before the reply, which is most of what handling a ticket actually consists of.
5. 05Anger and risk skip the queueA message that reads as a cancellation, a legal threat, a safety issue or a customer who has already written three times is escalated to a person immediately rather than taking its turn. Nothing automated replies to those. Queue position by arrival time is the thing that loses accounts, and this is the step that fixes it.
6. 06What it could not answer is reported every weekThe questions it declined, the drafts agents rewrote heavily and the topics that keep coming back, as a short list with volumes against them. That list is the product. It tells you which page of your website is unclear, which product note is missing, and where the next percentage point of automation is, and a system without it stops improving the day it ships.

What it's built from

Four pieces, and your helpdesk stays exactly where it is. This sits in front of it rather than replacing it.

• Claude, for the reading, the judgement and the draftingClassifies each message, decides whether it has enough to answer, writes the reply in your tone against retrieved material rather than from general knowledge, and reports its own confidence honestly. Abstaining is configured as a correct outcome, which is the setting that separates a system agents trust from one they start ignoring.
• n8n, the part that decidesSelf-hosted on European infrastructure, and the actual product. Routing rules, confidence thresholds, escalation triggers, working-hours behaviour and the weekly report all live here as something you can open and read. When you decide that billing questions never auto-send, that is one rule changed in one place.
• Notion, or wherever the answers actually liveThe policies, product notes and internal guidance, indexed in place rather than migrated into a new system somebody has to maintain. Resolved tickets are part of the corpus too, and are often the better part: the way your best agent explained something in March is more useful than the document nobody has updated since 2023.
• Slack, for the handoffs and the exceptionsEscalations arrive where the team already works, with the context attached and a link into the helpdesk. Anything urgent pages whoever is on duty. Teams works the same way. A handoff that needs somebody to go and look somewhere else has not handed anything off.

What it's worth

Do this with your own numbers rather than ours. It takes a minute, and it is the only version of it worth believing.

• You handle 800 tickets a month and half of them are routine questions with a known answer.400 a month at four minutes each, including the looking-up, is around 27 hours. That is most of a person, spent entirely on things the company already knows.
• Three quarters of those routine ones can be answered confidently and the rest become drafts.300 answered outright and 100 arriving pre-written for a few seconds of checking. The remaining 400 tickets still need people, and now they get them sooner.
• Your first response time on the hard tickets is currently a working day because of the queue in front of them.Clearing the routine half moves that to within the hour without hiring anybody. The customers who notice are the ones with real problems, which are the ones worth keeping.
• Two customers a month leave after a slow or badly handled exchange, at 3,000 euros a year each.72,000 euros of annual revenue walking out through a queue ordering problem. This is the number that justifies the build, not the 27 hours.

These are illustrations using round numbers and your own volumes, not results we are reporting. Ticket mix varies enormously by business, and the share that can be answered confidently is the single thing we measure first, because it decides whether this is worth building for you at all.

The research behind those assumptions

Published figures rather than ours. The first two explain why the obvious fix, a help centre, does not work as well as everybody hopes.

• 14%of customer service issues are fully resolved in self-service, on a survey of customers who tried. Gartner, 2024
• 9%of customers report solving their issue completely through self-service, on an earlier survey. The gap has barely moved in five years. Gartner, 2019
• 28%of the working week goes on reading and answering email, which is the shape of a shared support inbox. McKinsey Global Institute, The Social Economy
• 19%more goes on searching for information, which is the four minutes before the reply rather than the reply itself. McKinsey Global Institute, The Social Economy

What it costs

This is sold as one system at the automation price on the services deck, and it sits in the middle of that range at three to four weeks. What moves it is how much of your material is written down and how many systems an answer has to reach into. A company with current product documentation and one order system is quick. A company whose answers live in the heads of four people, with a one-off system behind the order lookup, is the top of the range, and part of the build is writing some of that down.

The running cost is a per-message cost in fractions of a cent plus a small hosting line, so it scales with ticket volume rather than with headcount. That shape matters: the months when support is busiest are the months the system is worth most, and a per-seat licence prices it backwards.

Outside the build price is somebody watching it, which is the care retainer. Here that largely means reading the weekly gap report with you and doing something about it. A triage system nobody tunes answers exactly the same proportion in month twelve as it did in month one, which does not mean it is stable. It means the business moved and the system did not notice.

What goes wrong, and what happens then

Every one of these is in the build because it happens, not because somebody raised it once in a workshop.

• It answers a question it should not haveThe threshold exists for this, and it starts high. Anything touching refunds, legal matters, safety, cancellations or money is set to never auto-send regardless of confidence. Those categories are a hard list rather than a judgement the model makes each time.
• A customer realises they are talking to a machine and asks for a personThey get one, immediately, without having to ask twice or guess a magic word. Every automated reply says what it is and how to reach somebody. The quickest way to make this system hated is to make the exit hard to find.
• The same customer writes for the fourth timeRepeat contact on the same issue is an escalation trigger by itself, regardless of what the message says. Somebody writing again is telling you the previous answer did not work, and sending them a fifth good answer is worse than sending them a person.
• The answer was right last month and is wrong nowThe material is indexed in place and changes are picked up, and answers carry their source so a stale one is visibly stale. The weekly report flags topics where agents are heavily rewriting drafts, which is usually the first sign that something underneath has changed.
• A message arrives in a language nobody on the team readsHandled in the language it arrived in where your material supports it, and routed with a translation where it does not. In the Nordics this is routine rather than an edge case, and pretending otherwise produces a queue of messages nobody opens.
• The helpdesk or the order system is downRetries with backoff, and a message is never lost: it is recorded in the pipeline before any downstream system is asked to take it. If context cannot be gathered, the ticket is routed without it and labelled, rather than held back waiting for a system that is not coming.
• Agents stop trusting the draftsThe thing to watch for, and it is measured rather than sensed: the share of drafts sent unedited is on the weekly report. A falling number means the thresholds or the material need work, and it shows up there weeks before anybody complains about it.