Data Readiness for RAG and AI Agents in B2B

Why data readiness beats model choice

Buyers and operators judge AI on whether answers and actions match reality. A stronger model cannot fix an index built from last year's export or a connector that ignores site-level permissions. Data readiness is a product and ops problem: who owns truth, how updates flow, what must never be retrieved or written. Treat it as a gate before spending on embeddings, agents, or vendor demos. If you cannot name the system of record for each field an agent will use, you are not ready for side effects. Start with Q&A over a clean, owned corpus—or pause the pilot.

• Model quality is secondary to ACL and freshness
• Wrong data in tools becomes wrong ERP posts
• Discovery time here costs less than incident cleanup later
• Pass a readiness checklist before contractor kickoff

Ownership and systems of record

Assign an owner per corpus and per transactional domain: policy manuals, CRM accounts, ERP inventory, ticket macros. Owner means who approves content changes and who is accountable when AI cites garbage. Separate systems of record from convenience copies. Vector indexes, data lakes, and nightly CSVs are derivatives. Agents that write must target the SoR; RAG should prefer SoR-backed exports with provenance. Document conflict rules: if CRM and ERP disagree on customer address, which wins? Agents and retrieval both need that rule encoded, not debated in Slack during the pilot.

• Name SoR, owner, and update cadence per domain
• Ban 'shadow' Spreadsheets as production knowledge without a sunset plan
• Record who can revoke a document or freeze a connector
• Link ownership to support escalation paths

ACL and tenancy before you index

Map who may see what today in the source systems. If SharePoint groups, ERP roles, and SaaS RBAC disagree, AI will not magically unify them—it will leak or refuse incorrectly. Encode tenant, site, role, and partner boundaries in metadata before the first embed. Align with multi-tenant architecture and SSO and identity so the retrieval principal matches the logged-in user. Test cross-tenant and cross-role leakage with real fixtures, not demo accounts that see everything.

PII and sensitive fields in corpora and tools

Free-text tickets, emails, and notes are PII mines. Decide what may enter embeddings, logs, and model prompts. Redact or exclude national IDs, health data, payment details, and secrets by policy—not by hope. Tool layers that call CRM or HR systems need field-level allowlists. An agent with 'read customer' that returns full SSN is a compliance incident waiting for a questionnaire. Persist AI traces under the same retention and access rules as other audit logs. Never dump raw PII into debug buckets 'just for the pilot'.

• Classify fields: public / internal / restricted / forbidden for AI
• Strip secrets and credentials from ingested docs
• Separate training/eval datasets from production PII where required
• Document DPA and subprocessors before vendor pilots

Freshness, lag, and ingestion design

Define an SLO for index lag: how stale may a policy or price list be before the answer is unsafe? Event-driven or publish-time reindex beats weekly batch scrapes for operational knowledge. Tombstone deleted and revoked documents immediately. Stale chunks that outlive the SoR are a common cause of confident wrong answers. For agents, freshness applies to tool reads too: cache TTLs on inventory or entitlement checks must match business risk. Pair ingestion design with ERP integration patterns when transactional truth must stay live.

Cleanup ROI: what to fix before the pilot

Not every folder needs a month of taxonomy work. Prioritize cleanup by blast radius: high-traffic operator questions, money-moving fields, regulated policies, and anything an agent may write. Cheap wins: remove duplicates, tag document owners, drop obsolete versions, fix broken ACL on the pilot corpus only. Expensive work: full enterprise taxonomy—defer until the pilot proves value. Budget cleanup explicitly in project estimates and MVP scope. Ignoring it makes AI look expensive when the real cost was dirty inputs.

• Pilot corpus: one domain, one persona, owned content
• Measure unsupported-claim rate after hygiene, not before
• Kill orphaned drives from the retrieval path
• Track cleanup hours vs. incident risk avoided

Discovery checklist before an AI pilot

Run this in technical discovery before vendor selection or contractor kickoff. Fail any critical item and you get a data project, not an AI project.

• SoR named for every domain the pilot will read or write
• Owner and update process documented for the pilot corpus
• ACL model mapped to authenticated user roles
• PII / forbidden fields classified and excluded
• Freshness SLO and reindex/tombstone path defined
• Ten golden questions with expected sources identified
• Write actions (if any) gated and idempotent
• Audit retention and access for AI traces agreed

Different bars for RAG vs agents

RAG needs citation-worthy, ACL-filtered documents. Agents also need reliable APIs, stable IDs, and conflict rules across systems. Do not promote a chat pilot to tool-calling until transactional readiness matches. Tool connectors to ERP, CRM, and tickets inherit MCP and tool-layer design: narrow schemas, permissions, and observability—not god-mode API keys. Human approval does not excuse bad master data; it only slows the damage. Fix IDs and ownership first.

Acceptance criteria for contractors

Require a data readiness report as a deliverable: SoR map, ACL tests, PII exclusions, lag metrics, and a go/no-go recommendation. A flashy demo over a shared folder is not acceptance. Tie payment milestones to leakage tests and golden-set retrieval quality per contractor hiring practices and AI contractor evaluation. Include runbooks for reindex, revoke, and rollback before production traffic.

Next steps

Pick one workflow and list every field and document the AI would touch. Mark SoR, owner, ACL, PII class, and max staleness. Gaps on that list are your real backlog. Then read RAG in production, agentic AI, other resources, case studies, book a call, or contact for a readiness review before the pilot budget lands.

FAQ