Technical enablement · 05

Implementation flow

The order of operations to bring a client engagement onto {{product.name}}, and which steps unlock which capabilities. An engagement moves through seven checkpoints, G1 to G7. Each gate is a decision a person records in the checkpoint console, and each depends on the ones before it: the console refuses a checkpoint that skips the gates ahead of it.

ⓘ

Where this comes from. The gate sequence and bounded options are defined in app-build-brief/08_MODULE_CHECKPOINTS.md and implemented in lib/domain/checkpoint-options.ts and lib/domain/gate-progress.ts. Target durations are design targets marked as inferred in the brief; they measure, they do not auto-decide.

The unlock graph

The spine on the left is mandatory and in order. Two tracks branch from it: attribution (outcomes and unit economics) and the gateway (live monitoring of agent sessions). Optimise needs findings from the baseline and, for measured savings, the attribution track. One example of an order that cannot be broken: no cost per outcome can be computed for a workload until its outcome definition is signed (G4).

Figure 1 — Unlock graph for an engagement. The left spine is mandatory and ordered; attribution and the gateway are parallel tracks; G6 draws on the baseline findings and the gateway's observed usage.

The seven gates

Gate	Question	Bounded options	Escalates when
G1 Scope and connection	Are the connections and workload scope enough for a defensible baseline?	Accept scope; accept with named gaps; remediate before baselining	Coverage is low, or a client-side prerequisite is unresolved
G2 Baseline acceptance	Is the cost baseline accepted as the reported starting position?	Accept; accept with caveats; reject and reconcile	Unexplained discrepancies above tolerance, or a provider restatement mid-period
G3 Governance framework	Are risk tiers, oversight policy and thresholds agreed?	Accept proposed tiering; accept with named changes; escalate for policy review	The most critical tier is proposed with oversight below its mandatory floor
G4 Outcome definition	Are outcome definitions, verification rules and values ready to sign?	Ready to sign; revise a named definition; insufficient basis, return to client	A value basis that cannot be quantified, or a verification rule with no data source
G5 Attribution acceptance	Is CPVO accepted as the reported unit economic?	Accept; accept with caveats recorded; hold pending reconciliation	Any workload awaiting reconciliation, or blending refused on incompatible units
G6 Optimisation authorisation	Which recommended changes are authorised for execution?	Authorise a named subset; authorise none; escalate for technical review	A recommendation touches a high-risk workload, or changes model routing on a client-facing one
G7 Sustained control review	Is the control state holding, and what is re-rated?	Sustained; sustained with named adjustments; trigger a scope and assumption review	Two consecutive periods outside the agreed band in the same direction

Every gate also has a defined "insufficient data, remediate" path, so a checkpoint never asks for a decision the data cannot support. The same console handles AMBER agentic decisions and variance notices.

Step by step

Each step lists what it delivers and what it unlocks next.

Create the engagement and its users

A platform_admin or capability_owner creates the engagement (tenant) in Engagements, sets its tier and data region, and a platform_admin creates users and assigns roles in Admin. Before real client data is ingested, a data protection impact assessment reference is required by the design. The reference is recorded in the optional DPIA reference field on the engagement form (Engagements → manage), stored on the tenant. The app does not block ingest when it is empty, so checking it is a process control for the engagement lead. Involve the Data Privacy Officer.

Unlocks: the engagement workspace and the checkpoint queue.
Connect the billing sources (G1)

On the Connections page add one named connection per provider account, using the least-privilege, read-only credential described under provider connections. Resolve client-side prerequisites early: AWS Bedrock needs model invocation logging enabled (without it there are cost figures but no token counts), and Google Vertex AI needs a detailed billing export to BigQuery, which has no retroactive backfill. Then run G1.

Unlocks: billing sync, the Govern cost baseline and the coverage badge.
Accept the baseline (G2)

Let at least one sync run (or make cron-run), review the cost baseline, drivers and open discrepancies, and record the decision. Accepting with caveats records them on the decision.

Unlocks: the Diagnostic opportunity list and export.
Agree the governance framework (G3)

Assign each workload a risk tier and agree the oversight policy for each tier. Workloads can be created in the app, so no database access is needed.

Unlocks: oversight economics and the attribution track.
Define and sign outcomes (G4)

For each workload, define what one outcome is, how it is verified and what it is worth with a stated basis, then have both the client owner and the Capgemini engagement lead sign. Start outcome event ingestion (POST /api/ingest/outcomes), or record sampled entries for pilots.

Unlocks: CPVO for the workload. Until signed, its state is NOT_ATTRIBUTABLE.
Connect the gateway (parallel)

Point the LiteLLM gateway's callback at the platform with the shared callback secret, and give each application's key a project identifier and workload identifier. Traffic with no workload is stored and can be named later. See gateway.

Unlocks: observed usage per workload, shared-scope apportionment, Gate A enforcement and the live monitor.
Accept attribution (G5)

Review CPVO by workload with states and caveats intact, movement against the prior period and AYCI. Record a baseline per workload if net advantage is wanted.

Unlocks: verified yield and the attribution report.
Authorise optimisation (G6) and measure

Raise recommendations from the opportunity list, authorise a named subset, execute and then measure. Savings carry the window they were estimated over.

Unlocks: measured savings and the savings export.
Review sustained control (G7)

At each review period confirm the control state is holding, or trigger a scope and assumption review. Two consecutive periods outside the agreed band in the same direction mandate one.

Unlocks: re-rating of workloads and a recalibration request to the capability owner if needed.

💡

The demo tenants (make seed-demo) are a quick way to see every state before touching a real engagement. They are excluded from billing sync and carry no real data.

ⓘ

Next: integrations for the credential detail and operations for running the jobs.