Two people are now most of the way to full-time on the inbox, replies are slipping past a day, and someone has started forwarding things to whoever is least busy. The suggestion that arrives at this point is always the same: get an AI assistant.
Before buying anything, spend an hour counting what is actually in there. Most support volume is not a support problem. It is a symptom of something upstream, and the cheapest fixes remove the question rather than answering it faster.
Count the customer support emails first
Take a hundred consecutive enquiries, not a sample of the memorable ones, and sort them into five piles. An hour, and it changes what you buy.
| Pile | What it means | The fix |
|---|---|---|
| Where is my thing? | Almost always the largest pile, and almost never a support problem. The customer cannot see a status you already know. | Publish the status. A tracking page, an automatic update at each stage. This removes the question rather than answering it. |
| How do I do this? | Something in your product or process is unclear at the moment people meet it. | Fix it where they meet it. A sentence on the page beats a help article, which beats an assistant explaining it. |
| You got something wrong | A real failure. These should be rare and, if they are not, no amount of support tooling helps. | Fix the cause. This pile is a defect report wearing a support costume. |
| I want to buy something | Not support at all, and frequently buried under the rest. | Route these separately and answer them fastest. Every hour these wait is money. |
| Genuinely one-off | The unusual, the judgement calls, the human situations. | This is what your team should be doing. Everything above is stealing time from it. |
In most inboxes we have looked at, the first two piles are the majority, and both are cheaper to remove than to answer.
The upstream fixes, cheapest first
- Publish status automatically. Order confirmed, dispatched, delayed, whatever your stages are. One email at each stage removes most of the largest pile and costs almost nothing.
- Answer the top five questions where people ask them. Not in a help centre nobody visits. On the checkout page, next to the field, in the confirmation email. Placement matters more than content.
- Fix the form that generates the follow-up. If half your enquiries need a reply asking for an order number, the form should have asked for it. This is the highest-return change on the list and it takes an afternoon.
- Set expectations honestly. A stated reply time reduces chase emails dramatically, and chase emails are a large hidden share of volume. Under-promise and the second email never arrives.
- Write eight saved replies for the eight questions you answer most. Not automation, just not retyping. Immediate and free.
- Then look at a help centre, once you know from your counting which articles it should contain.
Work down that list and a large share of inboxes stop being a crisis without any software being bought. The ones that do not are the ones with genuine volume, and those are worth automating.
When an assistant genuinely helps
We build these, and the conditions where they pay back are specific rather than general.
- Real volume, enough that repetition is measurably eating someone's week rather than annoying them.
- A large share of the same twenty questions, which your counting will have told you.
- Written answers that already exist and are true. If they do not, you are commissioning documentation with a chat window attached, which is a legitimate project priced honestly.
- A named owner with half an hour a week to read transcripts and correct what is behind them.
- A cheap failure mode, where a wrong answer mildly irritates rather than creating an obligation.
- Round-the-clock demand, which is where an assistant does something your team genuinely cannot.
Five or six of those and it will pay for itself. Two or three and you are buying a project rather than a solution. When you should not build a chatbot works through the negative cases, and can you trust what an AI assistant tells your customers covers designing for the day it is wrong.
Fix the routing before the volume
A separate problem that masquerades as volume, and it is worth checking because the fix costs nothing.
In most drowning inboxes, a meaningful share of the pain is not the number of emails but the fact that everything arrives in one undifferentiated pile. A sales enquiry worth four figures sits between a delivery question and a supplier invoice, and whoever opens the inbox works top to bottom. The result is that the valuable messages get the same attention as the trivial ones, and the urgent ones wait behind the routine.
- Separate sales from support at the point of entry, with different addresses or different forms. Anything that might be revenue should not queue behind a password reset.
- Give the genuinely urgent a different route, and say plainly on your site what qualifies. Without that, everything is marked urgent and nothing is.
- Stop things that are not customer enquiries reaching the same place. Supplier invoices, notifications, newsletters. A surprising share of an overwhelming inbox is not customers at all.
- Assign ownership per conversation, not per shift. The most expensive pattern we see is three people each partially handling the same thread and each assuming another has it.
None of that reduces the number of messages. It changes which ones get attention first, and in businesses where the inbox is genuinely full it often does more for the outcome than a reduction in volume would.
The option people skip
Between doing nothing and letting an assistant talk to customers there is a middle that is safer, cheaper and produces most of the benefit: let it draft, and have a person send.
Your team stops composing the same reply for the fortieth time and starts reviewing a draft, which takes seconds instead of minutes. The customer still gets a human answer. Nothing incorrect reaches anyone without being seen. And you find out how good the answers actually are before trusting them, using real questions, at no risk.
For most businesses this is the right first step and it is rarely offered, because it is a smaller sale than a customer-facing assistant. If it works well for three months, letting it answer the safest category directly becomes an easy decision rather than a leap.
Measuring whether any of it worked
Support gets measured badly almost everywhere, and the wrong measure quietly rewards the wrong behaviour.
- Enquiries per hundred orders, not enquiries per week. Volume rising while the business grows is not a problem. The ratio is the number that tells you whether anything improved.
- Contacts per resolved issue. How many messages it took to finish, in both directions. This catches the reply that answers a question nobody asked, which looks fast and creates a second email.
- First-reply time, honestly measured, including out of hours rather than only during the working day. Customers do not pause the clock overnight and neither should the metric.
- The share you could have prevented. Re-run the five-pile count quarterly. If the first two piles are shrinking, the upstream work is landing.
- Not deflection rate. The measure most support tooling reports, and the one most likely to mislead: a customer who gave up and a customer who was helped look identical in it.
That last point matters if you do buy an assistant, because deflection is what the dashboard will show you and it is not what you want. Ask instead what share of conversations ended with the customer saying thank you rather than asking for a person, and read a sample yourself each week.
What it costs
| Route | Cost | Removes |
|---|---|---|
| Status updates and better forms | A few days of work, $2,000 to $8,000 | Often the largest pile |
| Saved replies and expectations | An afternoon of your own time | The chase emails |
| A help centre worth having | $3,000 to $12,000 including writing it | The how-do-I pile, partly |
| An assistant that drafts for a human | $8,000 to $25,000 | The typing, not the reading |
| An assistant answering customers directly | $15,000 to $60,000 | The volume, if the conditions above hold |
Illustrative ranges from the kind of work we quote, not a price list. Note the order: the cheapest rows remove questions, the expensive rows answer them faster.
Not for you if
One number worth knowing before you spend anything: what share of last month's enquiries could have been prevented rather than answered. If it is most of them, the inbox was never the problem, and the software you were about to buy would have made you faster at something you should have stopped doing.