Onboarding stalls are detectable weeks before anyone notices.

Most onboarding programmes find out a customer is stuck at the QBR, or at renewal. This agent reads usage, ticket, engagement and stakeholder signals daily, classifies why an implementation has stalled, drafts the specific intervention, and routes it to whoever actually has the right to act. The interesting part isn't the drafting — it's the decision rights.

Active implementations

$2.4M

ARR in onboarding

Flagged at risk or stalled

51d

Median time-to-first-value

19d

Earlier stall detection vs. QBR cycle

Account	Segment	ARR	Day / target	Milestones	TTV forecast	Agent verdict

Click any account to open the agent's diagnosis, the signals behind it, and the drafted intervention.

Success-plan milestones

Signals evaluated

Highlighted rows tripped a segment-normalised threshold. Thresholds differ for Enterprise and Mid-market — a 21-day activity gap is normal in one and a red flag in the other.

Stall classification

Drafted intervention

Routing

The part most AI-in-CS demos skip

An agent that can draft a customer email can also damage a customer relationship. So the question isn't what the agent can do — it's what it's allowed to do without a human in the loop. This is set deliberately, and it's a leadership decision, not a technical one.

Agent-owned

Milestone status updates from telemetry
Recalculating TTV forecast
Internal alerts to the CSM and their manager
Fast-path flags on ahead-of-plan accounts
Compressing a plan when milestones land early

Human-reviewed

Any customer-facing draft
Re-scoping a success plan
Changing what counts as "first value"
Stopping or restarting the TTV clock
Adding a named stakeholder to the account record

Human-owned

Exec sponsor escalation
Anything commercial — credits, extensions, concessions
Declaring an onboarding failed
Involving another team's leadership
Any account already in a churn conversation

Why it's drawn here and not elsewhere

Three principles, and the reasoning behind each.

1. The agent owns observation, humans own consequence.
Reading telemetry and updating a forecast has no downside if it's wrong — you correct it tomorrow. Sending an escalation to a customer's CFO has a downside you can't take back. The line follows reversibility, not difficulty.

2. Anything that moves the number is human-reviewed.
If the agent could re-scope a success plan on its own, it could improve the TTV metric by redefining value. Metrics an AI can quietly move are metrics you can't run a business on.

3. Every intervention names an owner and a date.
An AI-drafted action with no human name attached is an action nobody does. The draft is only useful because it lands in a specific person's queue with a deadline.

Onboarding stalls are detectable weeks before anyone notices.

Success-plan milestones

Signals evaluated

Stall classification

Drafted intervention

Routing

What the agent actually does, in order

Suppression rules

Stall patterns it can name

The part most AI-in-CS demos skip

Why it's drawn here and not elsewhere

Where implementations actually fall out

What changed operationally

What it got wrong