Pyramid gates (Data Maturity Pyramid™)

Work top-down only after the level below is honest. Tick what is true today, not aspirational.

Level 1 · Data collection

• Critical business events are captured somewhere, even if fragmented
• You know which teams hold the master copies of customer, order and finance data
• Shadow spreadsheets are acknowledged, not ignored

Level 2 · Data quality

• Core entities have agreed definitions (customer, order, SKU, margin)
• Different reports for the same metric are reconciled or retired
• Someone owns quality issues when numbers disagree

Level 3 · Data integration

• Key systems connect without weekly manual extract theatre
• A central store or warehouse exists for the metrics leadership actually uses
• Join keys (customer ID, order ID) are stable enough for analysis

Level 4 · Data analytics

• Repeatable reporting exists on a defined cadence
• Teams can answer “what happened?” without a hero analyst every time
• Dashboards have owners; unused reports are culled

Level 5 · AI-driven insights

• Leadership has named commercial outcomes for AI, not tool FOMO
• Classified data and access controls exist before models touch personal or confidential data
• Human review paths exist for material automated decisions

Enterprise diagnostic dimensions

Aligned to the six dimensions in the Enterprise Data Maturity tool.

Data quality

• Definitions documented for board and operating KPIs
• Automated checks exist for at least the top revenue and margin feeds
• Known bad fields have owners and fix dates

Governance

• Systems register and classification scheme in use
• DPAs and Article 30 match reality for processors in the AI path
• Access is named-user with MFA on systems that would feed a model

Infrastructure

• Data can be reached without copying to personal laptops
• Environments separate enough that experiments cannot overwrite production
• Logging and backup cover the stores AI would read

Analytics capability

• Analysts spend more time on insight than on stitching extracts
• Metric dictionary stops “two versions of truth” in meetings
• Prioritised backlog of questions exists, not only a tool wishlist

AI readiness

• Use cases ranked by commercial value and data feasibility
• No LLM or agent has blanket access to unclassified customer data
• Vendor AI features are treated as processors, with DPIA where risk is high
• Evaluation criteria exist before pilot spend (accuracy, latency, cost, risk)

Commercial alignment

• Named executive owner and budget for data / AI work
• Success metrics are commercial (margin, cycle time, conversion), not vanity model scores
• Change management for the teams who must use the output is planned

Hard no-gos before connecting an LLM

• No systems register or classification on the data the model would see
• Shared admin credentials on the source systems
• Unresolved dual sources of truth for the metrics the model would cite
• No lawful basis / DPA coverage for processors in the path
• No human escalation for outputs that affect customers, pricing or employment