Free tool
Enterprise AI Solution Canvas
Turn an AI idea into a proposal someone can assess. Complete the ten connected decisions, then review it for missing evidence, weak boundaries and contradictions.
- How to write each entry
- Label it: fact, decision, assumption, gap or request
- Stuck?
- See the Atlas example
Business problem and outcome
- Which user or stakeholder experiences the problem?
- What happens today?
- What material pain, delay, error, cost or missed outcome results?
- What evidence supports the problem statement?
- What outcome would justify intervention?
User and workflow
- Who is the primary user for the proposed release?
- What triggers the work?
- Which inputs, judgements, handoffs and outcomes exist today?
- Where is the proposed intervention point?
- Which part of the workflow remains unchanged?
AI task, pattern and ambition
- What precise task will the AI-enabled product perform?
- Is the pattern search, retrieval, classification, extraction, drafting, assistance, tool use or delegated multi-step action?
- Is the ambition advisory, augmentation or automation?
- What simpler pattern was considered?
- Which capabilities are explicitly outside scope?
Value and success
- What is the current-state baseline?
- What mechanism connects the product behaviour to the expected outcome?
- Is the benefit cashable saving, capacity released, cost avoided, service improvement or risk reduction?
- Which outcome, product, quality and risk measures matter?
- Which decision will each threshold support?
Knowledge, data and access
- Which sources are authoritative and eligible?
- Who owns authority, quality and change?
- Which status, date, jurisdiction, relationship and permission metadata are required?
- Can content be extracted and retrieved with conditions and exceptions intact?
- How do updates, removals and access changes propagate?
Required behaviour and boundaries
- Which inputs and context may the product use?
- What output object must the user receive?
- Which controls are deterministic?
- Which behaviours require representative evaluation?
- What happens for missing context, no evidence, conflict, prohibited requests and service failure?
Human oversight and exceptions
- Which role reviews routine output?
- Which role owns interpretation or consequential judgement?
- What evidence appears at the moment of review?
- Can the reviewer edit, reject and escalate?
- What happens when the reviewer is unavailable?
- How will rubber stamping, correction effort and escalation capacity be measured?
Risk, governance and ownership
- What are the most credible failure modes and consequences?
- What are the blast radius, detectability and recoverability?
- Which preventive, detective, corrective and recovery controls apply?
- Who owns each control and the residual-risk decision?
- Which legal, regulatory, privacy, security or compliance determinations are required?
- Which changes reopen approval?
Evaluation and monitoring
- Which representative routine, difficult, restricted, prohibited and failure cases are needed?
- How will extraction, eligibility, retrieval, response, task, workflow and outcome quality be evaluated?
- Which failures are hard gates?
- Which graded measures use thresholds?
- Can the complete configuration and result be reproduced?
- Which production signals or sampled reviews continue the evidence after release?
Pilot and delivery roadmap
- What is the smallest real operating boundary that can produce decision-quality evidence?
- Which uncertainties are answered before real use, during supervised use and during bounded operation?
- Who owns product, knowledge, engineering, evaluation, support, incidents and change?
- What are the containment, rollback, recovery and exit routes?
- Which capabilities remain deferred?
- Which of the five decision routes is requested: proceed, sequence, reshape, park or stop? See Appendix C.
Saved on this device only. Your entries stay in this browser. Nothing is sent to us.