AI data readiness before model shopping

Garbage-in-garbage-out remains the dominant AI failure mode. Invoice extraction fails when vendors change PDF layouts monthly. Forecast models lie when category hierarchies disagree between ERP and planning tools. HR screening inherits historical bias when records are incomplete.

Data readiness prevents wasted pilot spend. Teams discover integration blockers week ten instead of week one.

Regulators and auditors ask where training and inference data originated. Lineage and lawful basis documentation are part of readiness, not optional paperwork.

Ready data accelerates iteration. When samples are accessible and labelled, teams test prompts and models quickly; when data requires manual hunts, progress stalls.

Weak data produces confident wrong answers that look authoritative in UI. Users trust formatted outputs; readiness work prevents automating false certainty at scale.

Audit and risk committees increasingly ask for evidence, not aspirations. Documenting why this topic matters in your context speeds approvals and reduces last-minute governance fire drills before go-live. Tie this section to named owners, review dates, and links in your intranet or GRC tool so it remains operational after the steering deck is filed.

Key components: (1) Inventory — sources, formats, owners, refresh cadence. (2) Quality metrics — completeness, consistency, timeliness for priority fields. (3) Access — SSO, role-based paths for models and humans. (4) Lineage and consent — lawful use, retention, cross-border rules. (5) Labelling and feedback loops for supervised workflows. (6) Integration endpoints — APIs, webhooks, batch exports to systems of record.

Prioritise by use case, not enterprise-wide perfection. Invoice AI needs AP samples; triage needs ticket exports.

Connect remediation to domain delivery — /solutions/ai-in-finance for AP, /solutions/ai-in-it-helpdesk for tickets — so data work ties to measurable workflow outcomes.

Define minimum viable datasets per use case — sometimes hundreds of labelled examples suffice for classification; sometimes you need six months of transactional history for seasonality. Scope prevents boiling the ocean.

Translate components into a RACI snippet: who owns each element, who approves exceptions, and which forum reviews metrics. Without names and dates, components remain abstract bullets nobody executes.

Waiting for a perfect lakehouse before any AI work — while high-value document workflows could start with bounded AP or contract corpora.

Assuming SaaS vendors "handle data" without reviewing subprocessors and export rights.

Ignoring master data ownership. AI cannot fix conflicting customer IDs without steward governance.

Training on historical data without bias review for people-impacting decisions — readiness includes ethical sampling, not only technical pipes.

Indexing entire file shares for RAG without access control inheritance — models leak titles and snippets from folders employees should never query through chat.

Teams often repeat these mistakes after reorgs or vendor changes — keep a short incident log so new managers inherit lessons instead of rediscovering the same failure modes.

We scope data readiness to sponsor use cases, producing a remediation backlog with owners and sequencing — not a multi-year data platform science project unless required.

Samples are tested early in discovery: can we extract, can we integrate write-back, what error rate appears on real documents? Findings feed build-vs-buy and vendor selection.

When data is not ready, we say so and recommend preparatory work before production promises — linking to governance and responsible AI guides where people-impacting data is involved.

Remediation backlogs name stewards and dates, sequenced before model spend. We validate samples in discovery so data work purchases learning, not hope.

Engagements exit with a handover checklist tied to this guide — owners, dashboards, and policy links — so your team can operate without consultant dependency after hypercare ends.

Sample data pulls during readiness should include edge cases: malformed PDFs, legacy encodings, missing fields — averages hide the exceptions that break pilots.

Run this per target use case before model or vendor spend. Inventory: list sources, formats, refresh cadence, and named stewards — no anonymous data team. Access: confirm SSO paths, role boundaries, and whether inference systems inherit folder permissions correctly. Quality: sample fifty to two hundred real records including edge cases; measure completeness and consistency on fields the workflow actually uses.

Lawful use: map data classes to policy tiers, document retention and deletion rules, and verify RAG indexes will not immortalise records HR or finance must purge. Integration: test API or export paths to systems of record; a successful read without write-back still blocks straight-through processing.

Decision gate: green-light pilot only when stewards accept remediation dates for blocking gaps; yellow-light when gaps are bounded and monitored; red-light when master-data conflicts or consent gaps make automation unsafe.

Re-score at pilot midpoint. Data drift — new vendor PDF layouts, org restructures, policy changes — is normal; readiness is a living state, not a one-time sign-off deck. Publish steward names and review dates in the same GRC or intranet space where audit teams already look for control evidence.