Worked example: reviewable client onboarding with AI agents

Worked-example boundary

This article is a hypothetical architecture exercise. "S.C." is a fictional placeholder, not an identified Samsarix customer, and the workflow below has no retained production or customer evidence. Use it to reason about boundaries and acceptance checks—not as proof of achieved business results.

The scenario

Assume a consulting firm has documented an onboarding process that involves:

1. Collecting client information via intake form
2. Running background checks
3. Assigning to the right practice area
4. Creating a custom onboarding plan
5. Scheduling kickoff meetings
6. Preparing engagement letters
7. Setting up project management workspaces
8. Coordinating with the finance team for billing setup

Before automation, the team would first capture a measured baseline: step count, owners, elapsed time, corrections, regulated decisions, and failure recovery. Without that packet, improvement percentages would be invented.

A proposed Samsarix design

One proposed Spiral uses six bounded phases:

Phase 1: Intake (Echo) Echo extracts named fields into a reviewable draft and proposes a practice-area label. Missing or ambiguous fields route to a human; a confidence value is only useful after calibration on representative examples.

Phase 2: Verification (Kavach) Kavach validates the response from an authorized, scoped verification provider and records policy flags. Eligibility and high-impact decisions remain with an authorized human reviewer.

Phase 3: Planning (Vega) After approved inputs exist, Vega proposes an onboarding plan with timeline, milestones, and resource requirements for review.

Phase 4: Coordination (Mitra) Mitra proposes available kickoff times from authorized calendars. Sending invitations or changing an accepted meeting follows the configured approval policy.

Phase 5: Documentation (Sage) Sage drafts an engagement letter from approved fields and routes it to the partner for legal review and e-signature.

Phase 6: Activation (Praxis) Praxis prepares the workspace and billing setup actions. Irreversible or financial actions remain behind explicit owner approval.

Acceptance checks before a pilot

• Baseline integrity: Compare the same start/end definition and case mix before claiming a time reduction.
• Handoff quality: Count required human decisions separately from avoidable copy-and-paste transfers.
• Correction rate: Define an error rubric, sample size, and reviewer before comparing workflows.
• Customer experience: Ask participants directly; meeting speed is not a substitute for satisfaction evidence.
• Safety: Exercise ambiguous intake, provider failure, duplicate submission, approval rejection, and rollback.

Why it worked

The design is useful only if it replaces opaque transfers with reviewable coordination. Each phase should emit a bounded receipt that the next phase can consume without granting it broader authority.

Approved verification fields can inform the planning draft, and an accepted kickoff time can inform the document draft. Reuse should be explicit, attributable, and reversible—not invisible shared memory.

Use the pattern carefully. Start free → Begin with sample data, explicit approvals, and acceptance checks before any production pilot.