Risk that’s provable. Language that reads human.
A triage agent for customer success. Every week it reads the account and hands the CSM a call they can defend to a customer — no black-box score, no model inventing things. It plugs into the existing stack and leaves it alone.
Weekly health review
4.0→1.0
hrs/wk
At-risk flagged ≥60d out
40%→90%
Reconcile→Split→Gate→Act
The fix is four moves, and this agent is all four. Reconcile the customer to one ID — five systems, five names for the same account. Split the risk math into code and let the model only narrate. Gate every draft behind grounding and a human. Close the action gap so a flag becomes a tracked intervention.
Salesforce, Gainsight, Totango, Catalyst already hold the account, log the activity, compute a health score, and fire playbooks. What none of them do is turn the signals into a call a CSM can defend.
The platforms
A number with no receipts
- Hold the account, log the activity
- Compute a health score — hand-set or weighted
- Fire playbooks
This agent
A claim with a trail
- Reads the raw signals
- Makes a risk call the CSM can defend
- Every claim traces to a source
“usage down 22%, tickets up 58% — and here’s the email where the champion went quiet.”
One line governs the whole system. It’s the thing that separates this from every AI-for-CSM tool that lets a model decide who’s at risk.
The risk math is code. The language is the model.
Deterministic · code
Code decides
- Risk scoring and every threshold
- Which flags fire, and how severe
- Escalation routing
- Action routing + playbooks (owner, SLA, steps)
- The grounding check
Reproducible. Auditable. A flag means the same thing every run.
Probabilistic · model
The model writes
- The health summary
- The single next action
- Renewal talking points, and their tone
It only ever narrates flags the code already produced.
It reads from the stack, read-only, and adds the reasoning layer on top.
→signals in
read-only
The agent
Risk engine
code · scoring & thresholds
Language
model · summary & draft
Grounding
every claim traces to a signal
Eval & drift
weekly self-scoring
Action gap
route → playbook → track · the flag becomes a tracked intervention
The agent never sends. A human does.
→snapshot
flags · routed actions
CSM workflow
confirm flags · run playbook · log outcome · send
no writeback · access-scoped · read-only
Tool-agnostic by design. The agent is a reasoning layer over a swappable stack — the deterministic engine and the LLM are isolated, tagged-in components. Swap Gainsight for Catalyst, Snowflake for BigQuery, or the model itself, and the risk math and the language layer don’t move. The eval loop re-scores every iteration, so the architecture survives changes to the external stack.
Three guardrails — built in.
✓
Grounded
Every flag cites the exact signal and threshold that fired it. Click through to the usage line, the ticket log, the email where the champion went quiet. Nothing gets asserted that can’t be traced to a sentence.
H
Human in the loop
A CSM confirms or dismisses every flag and edits every draft. The agent never sends anything on its own. Customer-facing output stays draft-only, always.
⚭
Watches itself
Weekly it grades its own work — grounding, flag precision, draft quality. When any dimension drifts, it says so. An agent that can’t be measured is one that can’t be put in front of a customer.
The long tail
The low-ARR book no human has time to watch. Under the old process, erosion goes unnoticed until the renewal is already lost. The agent monitors every one of these, weekly, forever.