Data Audit Worksheet: Score Every Data Source Before You Roll Out AI

List every data source the AI will need to touch — CRM, billing system, support tickets, contracts folder, email inbox, spreadsheets, scanned PDFs, the founder's notebook. For each source, score 1–5 across the five dimensions below. Total possible per source: 25. Sources scoring under 12 are unlikely to support an AI deployment without remediation. Sources scoring 20+ are AI-ready. The middle band (12–19) is where most production data lives — workable, but the gaps shape what AI use cases are realistic.

**Score 1:** Data is in a closed system with no API, no export, and no documented schema. (e.g., on-prem legacy ERP behind a vendor lock.) **Score 2:** Data exists but requires manual export (CSV download by a human on demand). **Score 3:** Data is exportable on a schedule (nightly batch) but not real-time. API exists but is rate-limited or partial. **Score 4:** Stable API or database connection with documented schema. Some authentication friction. **Score 5:** Fully API-accessible, well-documented, and the API is supported (vendor will actually fix it when it breaks). **Common failure pattern:** Score 1 source is the 'critical' source. The path forward is rarely 'replace the system' — it's usually 'sync into a secondary store the AI can read.'

**Score 1:** Less than 50% of records have the fields the AI needs. Mostly nulls and blanks. **Score 2:** 50–70% completeness. Critical fields frequently empty. **Score 3:** 70–85% completeness. Some required fields are reliably present, others aren't. **Score 4:** 85–95% completeness. Most records have most fields. Gaps are predictable (specific channels, specific time periods). **Score 5:** 95%+ completeness for all fields the AI use case requires. **Common failure pattern:** Sales CRM with 95% completeness on the deal record but 30% completeness on the contact record. The AI use case needs both. Score the limiting source, not the average.

**Score 1:** Free-text everywhere, inconsistent spellings, duplicate records, no validation. The same customer appears 5 times under 4 different IDs. **Score 2:** Some structure, many free-text exceptions. Duplicates and bad data common. **Score 3:** Mostly structured, but field meanings drift over time (the 'Status' field meant something different in 2023 vs 2025). **Score 4:** Clean schema, occasional anomalies, drift handled when noticed. **Score 5:** Actively governed — schema documented, validation rules enforced, anomalies tracked and resolved. **Common failure pattern:** Free-text status fields that humans interpret correctly but AI cannot. Either standardize at the source or build a normalization layer before the AI touches the data.

**Score 1:** Last update unknown. Could be days, could be months. No SLA on freshness. **Score 2:** Updated manually on a 'when someone gets to it' cadence. Often days behind. **Score 3:** Updated nightly. Acceptable for most analytical use cases, not for real-time AI. **Score 4:** Updated within minutes to hours. Suitable for most operational AI. **Score 5:** Real-time or near-real-time. Required for AI use cases that affect in-progress customer interactions. **Common failure pattern:** Pilot is scoped against a snapshot dataset that's already 30 days old. Production needs real-time. The architecture changes — budget for the difference.

**Score 1:** Unclear ownership. Access is whoever asks and gets to it first. No retention or deletion policy. **Score 2:** Owner identified, but access is informal and broad. No documented retention. **Score 3:** Owner identified, access is role-based, basic retention policy exists. **Score 4:** Strong governance — RBAC, audit logging, documented retention, regular access reviews. **Score 5:** Mature data governance — owner, RBAC, audit, retention, deletion-on-request, classification labels, compliance attestation. **Common failure pattern:** Customer data scoring 1–2 here. AI deployment forces the governance question — which is good, but adds 30–90 days to the project. Score honestly so the timeline is honest.

**Total per source = sum of all 5 dimensions (max 25)** • **20–25 (AI-Ready):** Greenlight for AI use cases. Minor remediation may be needed for specific use cases but the foundation is solid. • **15–19 (Workable with Targeted Fixes):** Identify the lowest-scoring dimension and fix that first. Most sources land here. Plan for 30–60 days of data work before the AI pilot starts on this source. • **12–14 (Heavy Lift Before AI):** The source can support AI eventually, but 60–120 days of cleanup is realistic. Consider whether AI on a sibling source can deliver value while this one is being remediated. • **Below 12 (Off-Limits for Now):** Do not point AI at this source. Either replace the system, build a clean intermediate data store, or scope around it. AI on bad data is worse than no AI — it scales the badness.

**Pattern A: One outlier source scoring 5–7, everything else 18+.** Don't let the outlier delay the project — scope AI around the working sources and remediate the outlier in parallel. **Pattern B: Everything scores 12–16.** Your data is workable but unloved across the board. Get a 30-day data engineer engagement to lift everything by 3–5 points before serious AI investment. ROI on data cleanup is usually higher than ROI on AI features. **Pattern C: Strong on Dimensions 1–4, weak on Dimension 5 (governance).** Common in fast-growing companies. AI deployment will force the governance question and that's fine — but get the governance plan in place BEFORE production, not after. **Pattern D: Strong on Dimension 5, weak on Dimensions 2–3 (completeness and quality).** Common in regulated industries. The data is governed but not loved. Often a sign that the systems were chosen for compliance, not usability. Plan for normalization layers. **Pattern E: Strong on Dimensions 1, 4, 5 — weak on 2 and 3.** Data systems are technically sound but the underlying processes don't capture clean data at source. Fix the process before fixing the data; otherwise cleanup is a treadmill.

Use this table format for each source. Fill in one row per source, total at the right. | Source | Accessibility | Completeness | Quality | Freshness | Governance | Total | Band | |--------|---------------|--------------|---------|-----------|------------|-------|------| | [Salesforce CRM] | 5 | 4 | 4 | 4 | 4 | 21 | AI-Ready | | [Help desk tickets] | 4 | 3 | 2 | 4 | 3 | 16 | Targeted fixes | | [Contracts folder] | 2 | 3 | 2 | 2 | 3 | 12 | Heavy lift | | [Email inbox] | 2 | 4 | 1 | 5 | 2 | 14 | Heavy lift | | [Founder's notebook] | 1 | 5 | 4 | 3 | 1 | 14 | Heavy lift | **Cross-source view (after scoring all):** • Sources scoring 20+: ______ • Sources scoring 15–19: ______ • Sources scoring 12–14: ______ • Sources scoring below 12: ______ The distribution tells you where to put the data engineering dollar.