AI email triage for business: why first-time-right routing matters most

AI email triage for business means a model reads each incoming message, decides what it is and sends it where it belongs. That destination might be a team queue, a support ticket, a standard response or the archive. Compare the Cloud describes triage as sorting incoming mail into urgent, actionable and archive before anyone reads it in full.

Most products in this space are sold as inbox tidying for individuals. A business has a different problem. You have shared mailboxes, several teams and requests that carry deadlines. For you, the value of triage comes from getting each message to the right people on the first attempt.

That is the figure to design around: first-time-right routing. A system that summarises beautifully but sends a quarter of messages to the wrong team creates more work than it removes. A system that routes accurately lets each team start on its own work straight away.

The sections below compare the main approaches, put numbers on volume and licence costs, look at how HM Treasury built its correspondence triage tool, and show a live Claude service desk in front of Zoho Desk. They finish with a checklist and practical next steps.

Every misrouted message gets handled at least twice. Someone in the wrong team opens it, works out it is not theirs, and passes it on. The sender waits longer, and the receiving team often has to re-read the thread to understand it. The time saved by automation leaks away in that second touch.

The stakes depend on where the mistake lands. Compare the Cloud notes that when Microsoft's Copilot inbox feature misjudges an email, the message is marked with the wrong priority but is not moved or deleted. In a personal inbox that is a minor annoyance, because the email is still in front of you.

Queue routing is less forgiving. If a triage system files a customer complaint in the sales queue, the service team never sees it until someone notices. If it files a subject access request as general enquiries, a legal clock keeps running while nobody owns it.

There is also a data protection angle. The ICO lists information sent to the wrong person as a common type of personal data breach. It also says that sending an email internally to the wrong department is lower risk than sending it to an unknown person outside the business. Lower risk is not zero risk, and a triage system that misroutes at scale multiplies the exposure.

The numbers explain why businesses want this automated. According to Compare the Cloud, the average business user received 121 emails and sent 40 each day in 2025. The same analysis reports that 81 per cent of UK small and medium businesses rely on email as their primary client communication channel.

Manual triage is expensive at that volume. The article estimates that it takes around 28 per cent of the average knowledge worker's day. It also reports that over 25 per cent of business inboxes now use AI to categorise or prioritise email, so adoption is already well under way.

Now apply accuracy to volume. Take a team of five, each receiving 121 emails a day: that is 605 messages. The illustration below uses round rates, not measured ones.

• At 99 per cent first-time-right, about six messages a day need re-routing.
• At 90 per cent, about 60 messages a day need re-routing.

Sixty misroutes a day is not a rounding error. It is a steady stream of rework, delayed responses and confused customers, and it is enough for staff to stop trusting the system and start checking every decision by hand. Once that happens, most of the saving has gone. The gap between 90 and 99 per cent matters far more than which model or vendor you pick.

Triage tools fall into three broad groups. They differ less in how clever the model is and more in what they decide and what happens when they get it wrong.

Copilot's feature ranks messages using the sender's relationship to you, their job title, the content and whether action is required. It reached general availability in April 2025 for new Outlook for Windows and Outlook on the web. Microsoft pulled it from Outlook for iOS and Android in February 2026 after user feedback, which is a reminder that per-user features can change under you.

For a business, the second and third approaches are where routing accuracy counts, because they move work between people. They also work best when you build AI into the systems you already use rather than adding another inbox to watch.

Licence costs are easy to compare. Compare the Cloud gives these UK prices per user per month:

• Microsoft 365 Business Premium: £16.90, which does not include Copilot.
• Microsoft 365 Copilot add-on: £16.10 at full price.
• Power Automate Premium: £11.60.

The structure matters as much as the price. Copilot is licensed per user, so ten users cost £161 a month for the add-on alone. Power Automate is different when the flow runs on a shared mailbox: only the flow owner needs a Premium licence. The article estimates a basic classification flow takes about two hours to build and test, and the AI Builder prompt approach needs no model training or labelled data.

What none of these prices cover is the cost of wrong decisions. A cheap flow that misroutes one message in ten can cost more in staff time than a pricier set-up that gets almost everything right. Set that against the estimate that manual triage takes around 28 per cent of a knowledge worker's day. The saving comes from removing that work, and misroutes put some of it back.

So when you compare quotes, ask for the expected first-time-right rate and how it will be measured. A price without an accuracy figure tells you only half of what you are buying.

The public sector publishes more detail than most vendors. HM Treasury's algorithmic transparency record, dated 17 December 2024, describes a tool that helps its Correspondence and Information Rights team triage incoming ministerial and official correspondence. Several design choices carry over to any business.

Split the job into parts

The tool predicts eight fields per case, including team, subject, summary and a high-risk flag. It does not use one model for everything. A Support Vector Machine classifies the correspondence team, GPT-4 prompting produces summaries and flags, and retrieval-augmented generation matches letters to standard lines. The record says this method was chosen because it needed minimal retraining and allowed custom models for each subproblem.

Test on history, keep humans in charge

The training data was 1,085,710 historical documents, 252.07 GiB in total, tested with an 80/20 split to establish the classifier's accuracy. The assigned team member monitors predictions and can overwrite any field. The tool may suggest standard lines, but a Treasury employee still prepares and signs off every response.

Results, and one gap

Components including subject, high-risk flag and summary were evaluated as performing as well as or better than existing processes. The tool is expected to save several hours of manual work per day for multiple team members. One gap stands out: the record states that no DPIA had been undertaken at that stage.

The same principle holds in a commercial service desk. Asset Services Group (Message Direct) runs an IT service desk on Claude inside Microsoft Teams, fronting Zoho Desk, and in the published Teams IT assistant case study their Head of Technology reports 99.7% first-time-right routing and a 40% efficiency gain on the system Svennis built.

Two things about that set-up are worth copying. First, the headline measure is routing accuracy, not the number of chats handled or responses drafted. That is the figure that tells you whether the helpdesk team can trust what arrives in their queue.

Second, the assistant sits in front of the helpdesk rather than replacing it. Staff ask for help in the tool they already use every day, and the request becomes work in the system the service team already runs. Tickets, ownership and reporting stay where they were. The AI's job is to get each request to the right place, correctly described, on the first attempt.

This is the same pattern as HM Treasury's tool: the model makes the routing decision and prepares the ground, and people do the work that needs judgement. It suits any business where email or chat feeds a queue, whether that is IT support, customer service or accounts payable.

Triage means a model reads your mail, and much of that mail contains personal information. Compare the Cloud's analysis states that when an AI model sorts incoming email, that is processing of personal data by automated means under UK GDPR. That is not the same as the stricter rules on automated decisions about people, which apply only in narrower cases.

Treat the Compare the Cloud view as a starting assumption, and check it against your own legal advice. Our overview of AI law in the UK covers the wider picture.

The ICO defines personal information as anything that could identify someone, such as photographs, comments or other records. Information about a company itself, such as a generic address like info@company.co.uk, is not personal information. A named work address like john.smith@company.co.uk is, and so is anything in the message about that person. Most customer and staff mail will contain some personal data, so plan on that basis.

Two routing rules follow from ICO guidance:

• Recognise subject access requests. Even if you refuse a request, you must tell the person within one calendar month. A triage system should flag these and send them straight to whoever handles them.
• Plan for misroutes as incidents. Information sent to the wrong person is a common breach type. The ICO says you should begin a risk assessment even before you have all the facts, and continue it even when a breach is not reportable.

Both ICO guides note they are under review following the Data (Use and Access) Act, so check the current versions before you finalise your process. HM Treasury's record from December 2024 said no DPIA had been done at that point. Have your own assessment written down before go-live.

These steps draw on the approaches above. They apply whether you use a Microsoft flow, a helpdesk integration or a custom build, and they fit within wider AI automation work.

1. Define queues from real mail. List the teams that actually handle work, with a one-line rule for each. Keep the list short. Overlapping categories are the main source of misroutes.
2. Add a fallback queue. When the model is unsure, it should send the message to a queue a person checks, rather than guess.
3. Split the decisions. Like HM Treasury, treat team, priority, risk flags and suggested responses as separate outputs. You can then measure and fix each one on its own.
4. Test on history. Run a sample of past messages, with known correct destinations, through the system before it touches live mail.
5. Keep override easy. Staff must be able to reassign in one step, and every reassignment should be logged.
6. Measure first-time-right weekly. Divide messages that stayed in their first queue by total messages routed. Review the misroutes by category and tighten the rules where they cluster.
7. Route legal deadlines explicitly. Subject access requests and complaints get their own rules and alerts.

The reassignment log is the most useful asset here. It shows exactly where the rules are weak and gives you evidence when you report results to the board.

You can start without buying anything. The aim of the first fortnight is a baseline, so you know what good looks like before you automate.

1. Count a week of mail. For one shared mailbox, record how many messages arrive, which team finally handled each one and how many were passed on at least once. That gives you today's first-time-right rate.
2. Draft the queue list. Write the categories and rules from that week's real messages, including a fallback and explicit rules for subject access requests.
3. Choose the approach by where mistakes land. Personal prioritisation suits individuals. Shared mailbox classification or helpdesk triage suits teams that pass work between each other.
4. Pilot on one mailbox. Run the system alongside your current process for two weeks and compare its routing with what staff actually did.
5. Set a target before go-live. Agree the first-time-right rate you need, and do not switch off manual checking until you reach it.
6. Write down the data protection assessment. Cover what the model reads, where data is processed and how you will handle a misroute.

If email triage is one of several tasks you are weighing up, our page on AI by business task sets out common options. The guide to how any company can use AI helps you rank them by effort and return.