ATLAS™
AI product systemsA framework for designing, delivering and operating products where AI is central to the value. It covers the full lifecycle of probabilistic systems: trust, measurable value, guided adoption and responsible scale.
“AI products do not deliver value instantly. They earn trust, learn with users, and compound value over time.”
When to use ATLAS™
Any system whose behavior is probabilistic and data-driven, and must earn trust over time.RAG bots and copilotsAssistants powered by retrieval-augmented generation.
Intelligent agentsSystems that decide and take actions on their own.
Decision enginesRecommendation and decision support.
Intelligent workflowsProcesses where AI drives execution.
Three principles
What makes AI products different from software.
01Invisible AIThe best AI experience keeps users focused on their problem. If you need an elaborate interface to showcase the AI, the fit may be off.Test: would this be better as a simple autocomplete?
02Metrics are hypothesesReal value metrics emerge from POC observation. If users feel value your metrics miss, change the metrics, not the users.Hypothesis → discovery in POC → validation at scale
03Trust is earnedStart with low autonomy and expand it as the system proves itself. Users should not trust AI on day one, and they won’t.Design for progressive autonomy
InformShows relevant information
SuggestProposes, user decides
AssistDrafts the work, user edits
Act with confirmationExecutes after approval
Act autonomouslyExecutes, user audits
The lifecycle
Eight stages in three phases.
Discovery1Problem framing2Data readiness3Experience design
Adoption and delivery4POC and guided adoption5System delivery
Value and scale6Value validation7Learning and co-evolution8Operations and scale
4 · POC and guided adoptionWhat does the system need to learn, and what does the user need to learn?
6 · Value validationWhich POC metrics actually correlate with value, and which new ones emerged?
7 · Learning and co-evolutionWhat is the system learning, and what are we learning about the customer?
POC Playbook
Guided adoption in four weeks.
A POC is where you discover what users actually value. Come with hypotheses, leave with evidence.
Get the playbook| Week | Focus | Activities |
|---|---|---|
| 1 | Setup and baselineDocument value hypotheses | Success criteria · 5 to 10 pilot users · current metrics · observation plan |
| 2 | Guided introductionObserve first reactions | Training · shadowing · qualitative feedback · subjective value signals |
| 3 | Value discoveryFind what users value | Feedback vs. metrics · collaborative prompt tuning · new value signals |
| 4 | Validation and decisionRecommend go or no-go | Expected vs. actual · “Would you miss this?” interviews · learnings |
Apply ATLAS™