By KYN AI Advisory Team — AI implementation specialists, Singapore
A support agent and a sales agent should not share one policy, even when they read the same inbox. A support agent resolves a problem on something the customer already bought, and success is a closed ticket. A sales agent pursues a purchase the customer has not made, and success is a next step. They differ on tone, permissions, escalation and what each may promise. Route each message to one policy first, then let that agent act.
This article sets out the differences, a routing rule for a mixed inbox, and the hard cases where a message belongs to both.
What is the real difference between the two agents?
The difference is the goal and the cost of being wrong.
| | Support agent | Sales agent | |---|---|---| | Goal | Resolve the issue | Move the deal forward | | Success measure | Resolved, time to resolution, reopen rate | Reply, meeting, quote, outcome | | Customer state | Existing customer, usually identified | Prospect or lead, often unknown | | May promise | Fixes, refunds within a stated limit, delivery dates from the system | Meetings, a quote from the approved price list | | May never do | Pitch during an open complaint | Concede price or terms | | Tone | Calm, specific, short | Warm, curious, brief | | Stop condition | Issue closed or handed to a person | Opt-out, reply, or loss |
Vendors build both. Platforms such as Relevance AI list separate agent templates for support, sales and operations. The templates differ because the policies differ.
Why does one inbox need two policies?
Because the same sentence means different things to each.
Take "Can you send me the price again?" From a prospect, it is a buying signal and goes to sales. From a customer on an annual plan, it may be a billing query and goes to support, or it may be a renewal and goes to the account owner. An agent with one blended policy guesses. A guess in a sales context costs a lead. A guess in a support context costs a complaint.
Three risks of a blended policy:
- Upsell during a complaint. The agent sees a customer with an open issue and pitches an add-on. This is the most common way an automated message makes a bad situation worse.
- Support answers as sales. The agent quotes a refund or a discount because it was trained to be helpful. Refunds and discounts are different authority levels.
- Wrong data scope. A sales agent needs lead data. A support agent needs order and ticket data and sometimes personal data. Mixing them widens access for both. Keep each agent on its own restricted account and its own data scope.
How do you route a mixed inbox?
Classify first, act second. The router is a small agent whose only output is a label and a confidence score.
- Match the sender to the customer table. Known paying customer with an open order, invoice or ticket: support. Known customer with nothing open: account owner or sales, depending on the topic.
- Read the intent. Words about a fault, delivery, invoice error or access point to support. Words about price, scope, timing, a demo or "how do I get started" point to sales.
- Check for both. A message that reports a problem and asks about more capacity has two intents. Route to support first. Flag the second intent for a person after the issue is closed.
- Fall back to a person when confidence is low or the sender is unknown and the topic is unclear.
Log the label, the confidence and the rule that fired. When a person re-routes a message, record it. Re-routes are your training set for improving the rules.
What are the hard cases?
| Message | Route | Why | |---|---|---| | "My invoice is wrong and I want to add two users" | Support, then sales flag | Fix the invoice before any sale | | "I'm unhappy, I'm thinking of cancelling" | Person, same day | Retention is a commercial decision | | "Can I get a discount on renewal?" | Account owner | Price authority, not support or SDR | | "Is your system secure?" from an unknown sender | Sales, with a security answer approved by a person | Pre-sale assurance | | "Your agent emailed me three times" | Support and a person | Possible opt-out and complaint | | Legal threat or regulator mention | Person only | Stops all automation |
The last two rows matter most. A complaint about unwanted automation must go to a person and must trigger suppression. See out-of-office, bounces and unsubscribes.
What policy settings differ?
Write these as configuration, not as prompts that can drift.
- Refund and credit limit. Support: up to a stated amount without approval. Sales: none.
- Discount authority. Support: none. Sales: none. Both hand over to a person; see why sales agents need hard limits on discounts.
- Outbound initiation. Support: replies only. Sales: may start a conversation, after approval.
- Escalation triggers. Support: refund over the limit, repeat contact on the same issue, safety, legal. Sales: price pushback, competitor named, legal or procurement terms. See when an AI sales agent should escalate.
- Memory. Support sees the ticket history. Sales sees the deal history. A summary line passes between them, not the full record.
How do you measure each agent separately?
Measure against the goal, not volume.
- Support: first-response time, time to resolution, reopen rate, share handed to a person with a reason.
- Sales: reply rate, meetings booked, quote-to-outcome rate.
- Router: share re-routed by a person. Set a threshold, such as 5 in 100, and review the misses weekly.
If one agent's numbers improve while the other's fall, check the router first.
Where KYN fits
KYN builds sales, support and operations agents as separate agents with their own restricted accounts and their own data scope, in the client's own cloud account. Every action is logged with the actor. Outbound messages, money and deletions wait for a named person. For an insurance brokerage we built a reply desk that sorts each reply at about S$0.01 per reply.
The first 14 days are a review of your stack and interviews with your team. A full build ships in under 90 days. See sales systems, operations systems and security.
Frequently asked questions
Can one AI agent do both support and sales?
It can, but it is safer as two agents behind one router. Each gets its own permissions, data scope and escalation rules. A single agent with a blended policy tends to pitch during complaints or make promises outside its authority.
Who owns a customer who writes in about both a problem and a purchase?
Support owns it until the issue is closed. The sales flag is then passed to the account owner with a one-line summary. This keeps the customer from receiving a pitch while the problem is open.
Do I need separate inboxes?
No. One shared inbox works if the router labels each message first and each agent acts only on its own labels. Separate inboxes help when volumes are high and the teams are separate.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
