Can AI sort incoming work and send it to the right person automatically?

12 October 2026

What does AI work routing actually do?

An AI routing workflow watches one or more agreed intake points, such as a shared email address, website form, helpdesk, messaging channel or document folder. It reads the request, identifies the type of work, extracts useful facts and chooses a destination from rules you have approved. It might turn a new quotation request into a CRM opportunity, send an urgent service fault to the duty engineer, assign an invoice query to accounts, or place an incomplete form into a queue for clarification.

The AI is only one part of the workflow. Reliable systems combine classification with ordinary business rules. AI can understand that a customer saying their boiler has stopped on a freezing morning probably needs urgent attention even if they never use the word urgent. Rules then decide which region, contract level and engineer skill apply. An integration writes the record to the correct system, and monitoring confirms that the handover succeeded.

This is becoming normal operational technology rather than an experimental idea. The Office for National Statistics reported in July 2026 that about 35% of UK businesses with 10 or more employees used at least one AI technology, up from about 12% in late 2023. Improving business operations was the most common purpose, reported by more than 60% of larger businesses using AI. Sorting and routing incoming work is a clear example because it removes delay without asking AI to make the final commercial decision.

Which incoming work is easiest to route?

Start with high-volume work where people already follow recognisable patterns. Good candidates include sales enquiries split by territory or service, support tickets divided by product and severity, job requests assigned by location and skill, accounts messages separated into invoices, credit control and supplier queries, and customer documents checked for missing information before processing.

A practical first project needs four things. First, enough examples: usually 100 to 300 past items gives you a useful test set, while thousands of examples may improve confidence for a complex process. Second, a limited number of destinations: three to eight clear queues are easier to test than 40 individual employees. Third, objective rules: postcode, product, contract, due date and request type are stronger signals than a vague judgement such as who seems best. Fourth, a named exception owner who will handle uncertain cases.

Some requests should not be routed automatically at first. Complaints alleging harm, safeguarding concerns, legal threats, payment changes, job applications and messages containing health or special category data deserve stricter controls. You can still use AI to flag them, but a person should confirm the destination. Customer expectations support that hybrid design. MaxContact surveyed more than 1,000 UK consumers in February 2025 and found that 70% still wanted to speak to a real person when explaining a specific situation, while 55% abandoned calls when waits became too long. The lesson is not to remove people. It is to use automation to reduce avoidable waiting and get unusual cases to a capable person faster.

How should the workflow handle uncertainty?

Every classification should produce both a proposed destination and a confidence signal. You then set thresholds based on consequence. A routine brochure request might be routed automatically at 85% confidence because a mistake is easy to correct. A complaint, payment issue or safety-related job may require 98% confidence plus human approval. There is no universal threshold. The right number comes from testing real examples and measuring the cost of each kind of mistake.

Use three outcomes rather than a simple yes or no. High-confidence items go straight to the approved queue. Medium-confidence items receive a suggested destination that a person confirms. Low-confidence items go to an exception queue with the original message and the reason the system was unsure. Never silently discard an item because it does not fit a category. The fallback should be visible and owned.

The Information Commissioner's Office says meaningful human reviewers need suitable knowledge, experience, authority and independence to challenge AI decisions. Its human review guidance recommends documented testing criteria, target accuracy rates, acceptable tolerances, override logs and a manual fallback. That is sensible operational advice even where Article 22 is not triggered. Log the original item, suggested category, confidence, final route, human correction and outcome. Review errors weekly during a pilot and monthly once the workflow is stable.

What does a sensible technical setup look like?

The simplest architecture has five stages. An intake connector receives the message. A security layer checks attachments, sender details and allowed data. An AI service classifies the work and extracts fields. A rules engine applies business constraints. Finally, an integration creates or assigns the record and sends an acknowledgement. Monitoring records whether every stage completed.

You may not need custom software. Microsoft-first businesses can combine Outlook, Power Automate, AI Builder or Copilot Studio and Dynamics 365. A support team already using Zendesk, Freshdesk, Intercom or HubSpot should test the routing and AI features inside that platform before adding another supplier. Businesses with several systems might use Make, Zapier or n8n to orchestrate the flow. A custom service becomes reasonable when routing depends on proprietary rules, high message volume, strict hosting needs or unusual systems.

Give the workflow the least access it needs. Reading one shared inbox and creating draft records is safer than granting access to every mailbox plus permission to send, delete and edit. Keep credentials in a managed account rather than an employee's login. Separate classification from consequential action. The system may suggest that a request is an urgent refund complaint, but it should not approve the refund. It may identify a sales lead, but it should not promise a delivery date or price. Make each step reversible wherever possible, and keep the original request attached to the new record so staff can verify the interpretation.

How much does AI routing cost and how long does it take?

For a UK small business, expect roughly £2,000 to £8,000 for a focused workflow connecting one intake channel to one destination system. That normally covers process mapping, a labelled test set, configuration, integration, testing, staff guidance and a short support period. A more complex project spanning several inboxes, documents, CRM records and job systems commonly costs £8,000 to £25,000. Bespoke or regulated deployments can exceed that.

Software may cost from tens to several hundred pounds per month, depending on message volume, platform licences, AI usage, document processing and monitoring. The hidden cost is staff time. Someone must label examples, agree routing rules, test edge cases and review exceptions. Budget 15 to 40 hours of internal input for a focused pilot. If a supplier quotes only for building the automation and assumes your process is already clear, the final bill will probably rise.

A small pilot usually takes two to six weeks. Week one maps the process and gathers examples. Weeks two and three build and test. The remaining time runs the workflow in shadow mode, where it recommends routes without moving live work, then releases only the safest categories. Measure first-response time, percentage routed correctly, percentage needing human correction, number of lost or duplicated requests and time spent triaging. A worthwhile pilot should reduce manual sorting by at least 30% without increasing missed or misdirected work. If it merely moves effort from the inbox to an exception queue, stop and redesign it.

What usually goes wrong?

The first failure is automating disagreement. If sales, operations and customer service use different definitions of urgent or cannot agree who owns a request, AI will expose the conflict rather than solve it. Write the routing policy first. The second failure is measuring technical accuracy while ignoring operational outcomes. A classifier can be 95% accurate overall and still misroute half the rare high-risk cases. Break results down by category and consequence.

The third failure is letting the system act on missing information. If a job request lacks postcode, contract number or access details, the correct response may be to ask for clarification, not guess a destination. The fourth is training or configuring the system with old, inconsistent examples. Historical inboxes often reflect workarounds, staff absences and outdated territories. Treat them as evidence to review, not perfect truth.

The fifth failure is having no manual fallback. An API outage, changed email format or expired connection can quietly stop routing. Set alerts for unprocessed items, define how staff switch back to the shared inbox and test that fallback quarterly. Finally, do not confuse faster routing with better service. If the right team receives work immediately but still has no capacity, customers gain little. Routing is valuable when it removes avoidable delay, makes ownership visible and provides data about where demand is building. It cannot compensate for inadequate staffing or a broken fulfilment process.

When this does not apply

Do not make this your first AI project when incoming volume is low. If one person sorts ten straightforward messages a day in 15 minutes, the saving will not justify a £5,000 implementation. Ordinary mailbox rules, form fields or helpdesk categories may solve the problem for almost nothing. Use AI when language is variable, useful information is buried in free text or attachments, and manual triage is creating a measurable delay.

It is also the wrong project when routing itself has legal or similarly significant effects on a person. Recruitment screening, credit decisions, insurance outcomes and access to essential services need specialist data protection and legal review. AI may assist administration, but a superficial human click is not meaningful oversight. The reviewer must understand the case, have time to challenge the system and possess authority to change the result.

Finally, wait if your records, permissions or system ownership are weak. Connecting AI to several poorly controlled systems expands the problem. Start with a narrow read-only or draft-only workflow, use synthetic or redacted test data where practical, and prove reliability before adding new actions. The best first version often looks modest: one inbox, five categories, one destination system, one exception owner and a weekly accuracy review. That may sound less impressive than an autonomous agent, but it is far more likely to produce a dependable business result.

Is This Right For You?

This is a good fit if work arrives through a shared inbox, website forms, support tickets or uploaded documents and somebody spends several hours each week reading, copying and forwarding it. It is especially useful when the destination is predictable from facts such as customer, postcode, product, urgency, skill, contract or job type.

It is not a good fit if every request needs expert judgement, your team cannot agree who owns which work, or the destination data is unreliable. Do not automate a confused process. First write a one-page routing policy, clean up the queues and name an owner for exceptions.

If you want to test the idea, start with 100 to 300 recent requests. Measure how accurately a proposed system would classify them before allowing it to move live work. If you would like an independent view of whether your process is ready, book a free call. There is no pressure to proceed, and the honest answer may be to improve the process before buying any AI.

Frequently Asked Questions

Does the AI need access to every company email?

No. Start with one shared inbox or intake channel and grant only the permissions needed to read agreed messages and create a draft or queue item. Wider mailbox access adds risk without improving a focused routing workflow.

Can it read attachments as well as email text?

Yes, if the workflow includes document extraction for approved file types. Test scans, photographs and unusual layouts separately, block unsafe files, and route unclear or password-protected documents to a person.

How accurate should an AI router be before launch?

There is no single safe percentage. Aim for at least 90% to 95% on low-risk categories in shadow testing, but require stronger controls for safety, financial, complaint or personal-data cases. Measure serious errors separately from harmless ones.

Will customers know AI routed their request?

Usually they do not need a special notice merely because software assigned a queue, but your privacy information should explain relevant automated processing of personal data. Be transparent when AI interacts directly or influences a significant decision.

Can AI route work by urgency?

Yes, but define urgency with observable criteria such as safety words, service-level commitments, outage scope or due dates. Never rely only on emotional tone, and require human confirmation for high-consequence escalation.

What happens when the AI is unsure?

It should place the item in a visible exception queue, preserve the original message, explain the likely categories and alert a named owner. Uncertainty must never cause a request to disappear.

Should I use the AI features in my helpdesk or build a custom workflow?

Test the features in your existing helpdesk first because they are quicker to deploy and easier to support. Build a custom workflow only when several systems, proprietary routing rules, unusual data controls or high volumes make the standard features inadequate.