Designing a lead-routing agent in your CRM: rules, overrides and what to log
Let the agent read messy inbound leads and keep assignment in plain rules. Add a confidence threshold, an override with a reason, and a log sales can audit.
The short answer
Let the agent do what rules cannot: read free-text enquiries, enrich them and classify fit and intent. Keep assignment in explicit CRM rules. Send anything below a confidence threshold to a person, give reps an override with a reason code, and log every input, score, rule and override so sales leaders can audit why each lead went where it did.
Key takeaways
- Split the job: the agent classifies and enriches, while written CRM rules decide who owns the lead.
- Three confidence bands work for most teams: assign automatically, assign with a flag, or send to a person.
- Every override needs a reason code, because overrides are the cheapest training data a routing agent will ever get.
- In KPMG's survey of 204 US leaders at $1bn+ firms (April–May 2026), 26% had full, real-time visibility of what their AI systems cost to run.
- Gartner predicted in July 2026 that AI agents will outnumber sellers 10 to 1 by 2028, yet fewer than 40% of sellers will say agents improved their productivity.
In this article
What should a routing agent decide, and what should rules decide?
The agent should read and classify; written rules should assign. Most CRMs already route well when a lead arrives with a clean country, company size and product field. They route badly when the lead is a free-text enquiry, a forwarded email or a form with half the fields guessed. That messy middle is the agent's job.
Keep ownership in rules because rules are easy to audit. When a rep asks why a lead went to a colleague, "territory rule 4, matched on country and segment" ends the argument. "The model thought so" starts one.
Our view: an agent that both interprets the lead and picks the owner is harder to debug and harder for sales to trust. Split the two, and each half can be tested on its own. This post covers the inputs, the thresholds, the override and the log. It applies whether you build on your CRM vendor's agent or your own.
What inputs should the agent read?
Only the fields you would let a sales coordinator use, and nothing it cannot cite back. Each input should be stored with the decision, so a routing call can be replayed later with the same facts.
| Input | Source | What the agent does with it |
|---|---|---|
| Enquiry text | Form, email, chat transcript | Classifies product interest, urgency and buyer role |
| Company and domain | Form, email address | Matches to an existing account before anything else |
| Firmographics | Your enrichment provider | Fills size, industry and country when the form did not |
| Account ownership | CRM | Hands existing customers to the account owner, always |
| Open opportunities | CRM | Flags duplicates and active deals |
| Territory and capacity | Routing rules, rep calendar | Passed to the rules, not decided by the agent |
Our input list. Illustrative: add or remove rows to match your CRM.
The account-match step deserves its own test. A lead from an existing customer routed to a new-business rep is the most visible failure a routing agent can make. A domain lookup before any model runs prevents it.
Confidence thresholds: three bands
Give the agent three outcomes, not two. A clear case is assigned automatically. A probable case is assigned with a flag the rep can see. Anything uncertain goes to a person in a review queue, with the agent's reading attached.
One lead through the routing agent
- Lead arrives
- Account match
- Agent classifies and enriches
- Confidence band
- CRM rule assigns
- Rep accepts or overrides
Set the bands from data, not from the vendor's default. Run the agent in shadow mode for two to four weeks: it classifies every lead, a person routes as usual, and you compare. Then pick the threshold where the agent's automatic assignments match the human call often enough for your sales leader to sign off.
Our view: start strict, with a large review queue, and loosen it as the override log proves the agent right. A queue that shrinks over a quarter is a better story for sales than a launch that reroutes a key account on day two.
Overrides: make them easy and make them explain
Reps will override the agent; design for it. Put a one-click reassign in the CRM record with a short, required reason list: wrong territory, existing relationship, wrong product, not a real lead, other. Free text is allowed but not enough on its own.
Overrides do two jobs. They fix the lead in front of the rep, and they tell you where the agent or the rules are wrong. A cluster of "existing relationship" overrides points to a gap in account matching. A cluster of "wrong product" points to the classifier.
Seller trust is the measure that decides whether any of this survives. Gartner predicted in July 2026 that AI agents will outnumber sellers 10 to 1 by 2028, yet fewer than 40% of sellers will say agents improved their productivity.1 That is a forecast, not a finding, and our note on it covers what sales operations should take from it. For routing, the lesson is simple: an agent reps cannot overrule is an agent they will route around.
What should the log record?
Enough to replay any routing decision months later. Store the log as a record linked to the lead in the CRM, so sales operations can report on it with the tools they already use.
Fields in every routing log entry
- What came in
Lead ID, raw enquiry, the enrichment returned and the time it arrived.
- What the agent read
Classification, confidence score, model and prompt version.
- What the rules did
The rule that fired, the owner assigned and the time to assignment.
- What people changed
Override, who made it, the reason code and when.
- What it cost and produced
Cost of the agent run, and the lead's stage at 30 and 90 days.
The cost field is the one most teams skip. In KPMG's AI Quarterly Pulse survey of 204 US leaders at firms with revenue of $1bn or more (28 April–25 May 2026), 53% were deploying AI agents.2 Only 26% had full, real-time visibility of what their AI systems cost to operate.2 A cost per routed lead, logged at the source, answers that question for at least one workflow.
The model and prompt version fields matter for a different reason. When routing quality shifts, you need to know whether the leads changed or the agent did.
Paying per qualified lead: define "qualified" first
Outcome pricing is arriving in CRM agents, and it moves the definition problem onto your desk. No Jitter reported in April 2026 that HubSpot would price its prospecting agent per qualified lead and its customer agent per resolved conversation, from 14 April.3
Per-outcome pricing can be good value, but only if "qualified" means the same thing to the vendor's meter and to your sales team. Write your definition into the CRM as fields and stages before you compare offers. Then check, in the log, how many billed outcomes your reps would have accepted.
Our view: compare per-outcome and per-run pricing on your own logged volumes, not on the vendor's examples. Our breakdown of an agent's monthly running costs lists the other lines, from platform fees to the people who work the exception queue.
Permissions and rollout
The routing agent needs to read leads and accounts and to write a classification and an owner. It does not need to delete records, edit opportunities or send emails. Our least-privilege checklist for CRM agents sets out the role to create.
Roll out in three steps: shadow mode, then automatic assignment for the clearest band, then a review of the override log after a month. If you are still choosing which process to automate, run lead routing through our scoring sheet first. Our AI automation page describes how we scope one workflow like this on the systems you already run.
Sources
- Gartner prediction (Jul 2026): AI agents will outnumber sellers 10 to 1 by 2028; fewer than 40% of sellers will say agents improved productivity
- KPMG, AI Quarterly Pulse Q2 2026: 204 US leaders at $1bn+ firms, 28 Apr–25 May 2026 (Jun 2026)
- No Jitter, HubSpot brings outcome-based AI pricing to customer engagement (Apr 2026)
