The First 90 Days
A staged rollout plan — what to enable when, and the order that earns organisational trust rather than spending it.
Varcio is deliberately capable of acting on your infrastructure autonomously. That capability is only valuable if your organisation trusts it, and trust is earned in a specific order. This page is the order.
The governing principle
Every stage below is observe → understand → control → automate. Skipping straight to automation is the single most common way a cost programme loses its mandate.
Phase 1 — See (days 1–21)
Goal: one number everyone believes.
Connect every account, not a sample
Partial connection is worse than no connection, because it produces a confident total that is wrong. See Connecting your cloud.
Get ingestion green and keep it green
Check Overview daily for the first fortnight. You are building confidence that the pipeline is reliable before anyone makes a decision on it.
Spot-check findings against your own dashboards
Take five findings and validate them independently. Where a finding carries a measured utilisation sparkline, check it against what your team already believes.
Establish the baseline
Run FinOps Maturity and record the score. This is your programme's before-picture and you will want it later.
Phase 2 — Understand (days 14–60)
Goal: every dollar has an owner.
Run the tag advisor before authoring any tag policy
Tag Governance proposes policies from tags you already use. Codifying existing practice reaches compliance far faster than imposing a new scheme nobody follows.
Define the minimum viable tag set
Owner, environment, cost centre. Longer lists reduce compliance — resist the urge to be comprehensive on the first pass.
Run ownership inference sweeps
Ownership assigns untagged resources by naming convention, account structure, and network placement. This reaches a scale hand-tagging never will.
Allocate the genuinely shared remainder
Cost Allocation distributes shared cost — clusters, egress, logging, monitoring — so nothing accumulates in an unowned bucket.
Socialise the allocation methodology with receiving teams before the first chargeback. Disputed methodology is the usual reason chargeback programmes fail.
Start unit economics from suggestions
Unit Economics can propose metrics derived from telemetry you already collect. That is far faster than defining them by hand.
Phase 3 — Control (days 30–75)
Goal: stop the next expensive change before it ships.
Author every policy in advisory mode first
Regardless of confidence. Policies has a test lab and a simulation endpoint — use them to see what a policy would flag against real data.
Watch advisory findings for a full cycle
Correct the policies that produce false positives. A policy that cries wolf will be disabled by the first team it inconveniences.
Establish the exceptions workflow before blocking anything
Teams need a governed route around a policy, or they will find an ungoverned one.
Promote proven policies to blocking, selectively
Start with tag requirements — they are the least disruptive. Blocking requires Pro or above.
Shift left with PR Cost Review
PR Cost Review prices Terraform changes before merge. Seeing a $15,000/month increase in code review costs nothing; discovering it two weeks after it shipped costs two weeks plus an engineer to unwind it.
Phase 4 — Automate (days 45–90)
Goal: continuous control rather than periodic cleanup.
Define execution windows first
In Integrations, before enabling any automation. This is what prevents an automated action firing during a change freeze or peak trading period — and it is what makes automation acceptable to operations teams.
Add the execution role
A distinct, verified step in Cloud Accounts.
Start with approval-first, and dry-run everything
Optimization in approval-first mode. Dry run simulates the whole action without calling your provider.
Target scheduled shutdown early
Resource Parking is usually the highest-certainty saving available — non-production infrastructure running 24/7 for a 40-hour working week. It is also unusually safe, because nothing is destroyed.
Graduate Autopilot one trust level at a time
Autopilot has three modes: approval-led, then Co-Pilot (agent drafts, human confirms), then always-on. Move a rule class up only once it has a demonstrated track record.
Autopilot never terminates compute. Underutilised instances are stopped, not destroyed, and
anything tagged finops:protected=true is unconditionally skipped.
Bring approvals to where people already are
Route them into Slack or Teams. Approval queues that live only in the dashboard are not checked often enough, and a backlog is the most common cause of optimisation programmes stalling.
Reviewing at day 90
Recompute maturity
Compare against your Phase 1 baseline. Direction matters more than absolute value.
Realised vs identified
A persistent gap means an execution bottleneck, not a detection problem.
Check approval queue depth
A standing backlog is the clearest signal that trust settings need revisiting.
Report upward
CEO View is built to be understood in about thirty seconds — use it to keep sponsorship alive.