Judgment is moving from people to machines, so everything right of that line needs governing
Unchanged wrap-around: data, architecture, DevOps, and cloud governance.
New: model and evaluation management, evals in CI.
Agents do the work.
People own the outcome.
Every agent has an identity, a named human owner, and limited rights; no agent accepts its own work.
Action controls include authorization limits, human-in-the-loop thresholds, blast-radius limits, reversibility.
Agents prove themselves in sandboxes before promotion; multi-agent orchestration only after single-agent reliability.
Every accepted output carries a receipt
Who accepted it, against which criteria.
One reference model, built once and reused by every operating model
Nine components | Reference model |
|---|---|
Set once and inherited | Strategy & Governance, Technology, Environments, Performance, Partners & Suppliers |
Set per operating model | Organization, People, Demand, Service Delivery |
Shared AI platform service, a business within a business, organized into workstreams.
Team: platform engineering, ML and GenAI engineering, data and knowledge engineering, evaluation and assurance, forward-deployed engineers.
Leadership habits: standards, architecture reviews, mentoring, redesigned roles where AI drafts and people accept.
Find the value, prove it, then capture it, with a decision at every gate
Months 1–4
Find
Assess seven dimensions, publish reference architecture, set cost baseline, pick 2–3 workflows.
Gate 1
Months 5–9
Prove
Pilot workflow releases, evals and governance live, value net of AI spend.
Gate 2
Months 10+
Capture
Copy proven patterns, stand up second operating model.
Scorecard: leading (readiness, eval pass rates, adoption), lagging (cost per accepted outcome, lead time, quality).
Readiness L1 AI-Assisted → L5 Collective AI; pilot areas reach L4 in year one.
Brian lambert strategy thoughts on
Transforming
while performing
An AI platform is an operating model with a technology stack inside it,
not the other way round.
One shared engine of people and AI producing measurable business outcomes, while the business keeps running.
Six jobs, Capture, Resolve, Compute, Condition, Explain, Steer: Only Steer creates value; the rest is cost .
An answer after the decision is worth a lot less, however accurate.
Measure contributions the business accepts, not models trained, pilots launched, or tools bought.