Onboarding that fits the customer
A confidential, growing software company onboarded every customer by hand, on one path. We designed tiers so software runs the repeatable steps and people keep the hard decisions.
The challenge.
A signup for a handful of licenses asked for as much human time as a large rollout. The company needed onboarding that could grow with it. That does not mean handing the whole job to an AI agent.
The approach.
Requests split by license count. Under five: zero-touch, automated, with a person only when something fails. Five to twenty-five: automation plus human checkpoints. Above twenty-five: a person leads, and automation supports them. Behind the split: a deployment engine, durable orchestration such as Temporal, an API layer, templates, health checks, and reporting.
A request moves through received, validated, classified, provisioning, configuration, verification, and completion. Failure, approval, and manual work are states of their own. A run can pause and resume, a repeated step does not apply twice, and the trail can be audited. Software executes. AI only interprets — requirements, a suggested tier, an exception summary, a customer draft. A person approves what is hard to undo. A six-phase roadmap sets the build order. Most of this workflow should stay ordinary software.
How it moves.
Scroll to follow it through. Drag to turn the model, and select any part to read what it does. The drawing is the design, not a product screenshot.
Select a step or part to read what it does.
Onboarding requests arrive at a hub and are switched by license count onto three tracks that share the same seven states, from received to done. The zero-touch track runs on its own, records a failed step as an exception and resumes from that step. The assisted track stops at a gate until a person approves. On the complex track a person leads the whole way.
01
Each request is weighed first.
A new customer is classified by license count, then moves through the same states: received, checked, classed, provisioned, configured, verified, done.
02
Small requests run on their own.
Under five licenses, software runs the whole path. No person is needed unless something fails.
03
A failure is recorded, not repeated.
When a step fails, the exception is written down and the run resumes from that step. A repeated step does not apply twice, and the trail can be audited.
04
Mid-sized requests wait for a person.
Five to twenty-five licenses: automation runs, then pauses at a checkpoint until someone approves what is hard to undo.
05
Large rollouts are led by people.
Above twenty-five licenses, a person leads and automation supports them. AI only interprets: requirements, a suggested tier, an exception summary, a customer draft.
Illustrative flow. The platform is not built.
Where it stands.
A strategy and an architecture to build against. Deterministic steps, model judgment, and human approval stay in separate roles. The platform itself is not claimed as built.
Project status: analysis, strategy, and architecture are complete. The platform is not claimed as built.
A useful next step.
Let’s talk.
Bring the workflow that’s stuck. We’ll help you find a first step — even if that step is not a model.
Based in Toronto. Meetings in the GTA or by video.