By KYN AI Advisory Team — AI implementation specialists, Singapore
A task routing agent reads each incoming request, labels its type, and assigns it to the right person using a table you maintain. The model does the reading. The table does the deciding. When no rule matches, or the request is unusual, the agent sends it to a named triage person rather than guessing. This split keeps assignment predictable and lets a manager change who gets what without touching code.
What problem does a routing agent solve?
In a team of 5 to 50, requests arrive by email, WhatsApp, a web form and a shared inbox. Someone reads each one, decides who owns it, and forwards it. That someone becomes the bottleneck, and the work stalls when they are away.
A routing agent removes the reading and forwarding. It does not remove ownership. Every task ends up with one named person and a due date, and the log shows how it got there.
How does a task routing agent work?
Five steps, in order.
- Capture. Pull the request from each channel into one queue with sender, time, channel and text.
- Label. The model assigns one type from a fixed list: for example quote request, delivery query, invoice question, complaint, supplier issue, other.
- Look up. A rules table maps type, and where relevant customer or region, to an owner and a backup.
- Create. The agent writes the task to your work tracker with owner, due date and a link to the source message.
- Notify. The owner gets a message with the request and the due date.
Steps 2 and 3 are separate on purpose. The model reads. The table decides.
Why keep the rules in a table, not a prompt?
A prompt that says "send complaints to Dewi" hides policy inside text that nobody reviews. A table is auditable.
| Type | Condition | Owner | Backup | Due | |---|---|---|---|---| | Quote request | Value unknown | Sales desk | Sales manager | 4 working hours | | Quote request | Existing key account | Account owner | Sales manager | 2 working hours | | Invoice question | Any | Finance | Finance lead | 1 working day | | Complaint | Any | Operations manager | Director | 2 working hours | | Other | Any | Triage person | Operations manager | 1 working day |
The numbers in the due column are examples. Set yours from what you promise customers. When someone leaves or a rule changes, you edit one row. The change is logged with who made it.
What should the agent do when it is unsure?
Uncertainty is normal. Handle it with rules.
- Ask the model for the label and a confidence band, and treat low confidence as "other".
- Send "other" to a triage person, not to the least busy person.
- Flag requests that contain a legal term, a regulator, a threat or a refund demand. Route them to a named person and stop any automatic reply.
- If a customer is not in your records, create the task and mark the account as unmatched.
A wrong route that reaches a person who forwards it costs minutes. A silent guess that sits unread costs a customer.
How do you handle workload and absence?
Plain rules beat clever balancing.
- Check a leave calendar before assigning. If the owner is away, use the backup.
- Cap open tasks per person. When a person passes the cap, notify the manager rather than reroute silently.
- Reassign only on a person's request or a manager's. The agent does not move tasks that are already owned.
Round-robin looks fair but ignores skill and account history. Use it only inside a group where everyone can do the work.
How do you test it before it goes live?
Use real history.
- Export 100 recent requests that a person already assigned.
- Run the agent in shadow and compare its owner with the person's choice.
- Review every disagreement. Some will be agent errors. Some will show the table is wrong or that two people were doing the same job.
- Check the cases that matter most: a complaint labelled as a quote request, and a key account routed to the general desk.
- Run in shadow for two weeks before the agent creates real tasks.
Count misroutes by type, not as one total. A good overall rate can hide a poor rate on complaints.
What are the security and privacy limits?
The agent reads customer messages, so give it controls.
- Its own restricted account, with read access to the intake queue and write access to the task table only.
- No ability to send replies to customers. Acknowledgements, if you want them, come from a fixed template that a named person approves.
- Message text can carry instructions aimed at the model. Treat it as data. See prompt injection in business AI agents.
- Row-level security so a team sees only its own accounts.
What does it cost to run?
Labelling one short request is a single model call. KYN's reply desk sorts a reply for about S$0.01, a figure from a live system that varies with the model and message length. The cost sits in the build and in keeping the rules table accurate, not in the calls.
What does a routing log look like?
The log is the record that lets you improve the rules. Each row has the same fields. The example below is illustrative.
| Time | Channel | Label | Rule matched | Owner | Due | Outcome | |---|---|---|---|---|---|---| | 09:02 | Email | Quote request | Quote, key account | Account owner | 11:02 | Done in 1h 10m | | 09:15 | WhatsApp | Complaint | Complaint, any | Operations manager | 11:15 | Reassigned by manager | | 09:40 | Form | Other | No rule | Triage person | Next day | Relabelled as invoice question |
The third row shows the value of the log. A request landed in "other", a person relabelled it, and that relabel is evidence for a new rule. Review rows where the outcome differs from the route every week, and add or change one rule a week at most.
How do you write a good label list?
The label list is the only thing the model decides, so keep it short.
- Use 6 to 12 labels. More than that and neighbouring labels blur.
- Make each label mean one thing a rule can act on. "Urgent" is not a type. "Delivery query" is.
- Include "other" and "not a request", such as out-of-office replies and newsletters.
- Write one sentence and two examples per label in the instructions.
- Review the list every quarter and merge labels that always map to the same owner.
If two labels always route to the same owner, they are one label as far as the rules are concerned.
How do you measure whether routing works?
Three numbers, weekly:
- Reassignment rate. Tasks moved by a person after the agent routed them, split by label.
- Time to first action. Minutes from arrival to the owner's first reply or note.
- Triage queue size. Items waiting in "other". A growing queue means a missing label or a missing rule.
Report them as numbers per label. A single overall rate hides the complaint category, where a wrong route costs the most.
Where KYN fits
KYN builds routing as part of the operations system in a cloud account in your name. Agents use their own restricted accounts, every action is logged with the actor, and anything outbound waits for a named person. Our reply desk for an insurance brokerage uses the same label-then-rules pattern. The first 14 days are a review of your stack and interviews with your team. A full build ships in under 90 days. See integrations and security.
Frequently asked questions
Can the agent assign tasks from WhatsApp messages?
Yes, if the messages reach a system the agent can read, such as the WhatsApp Business Platform inbox. It treats each message like any other request. See managing WhatsApp enquiries across a sales team.
What if two people think a task is theirs?
The table assigns one owner. Anyone can ask a manager to change a rule, and the change is logged. Shared ownership produces tasks nobody does.
Is a routing agent better than a shared inbox rule?
An inbox rule matches keywords. An agent reads meaning, so "my shipment never arrived" and "where is my order" reach the same owner. Start with rules if your volume is low and your categories are few.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
