Enrichment turns Companies House, sponsor licence, workspace, and approved provider evidence into reviewed company intelligence. It is not a free-form data append tool. Every value needs a source, evidence, confidence, policy decision, review state, and downstream-use rule before it can power filters, exports, signals, or Direct Mail.
The sources currently registered
| Source | Current role | Use now? |
|---|---|---|
companies_house | Deterministic company classification and canonical context. | Yes, for low-risk company facts. |
uk_sponsor_licence | Home Office sponsor licence evidence and sponsor-aware claims. | Yes, for source-backed sponsor facts; review first. |
official_website | Domain, website, and contact-path discovery. | Only with evidence and review. |
workspace_context | Internal segment, company, prior claim, and workspace facts. | Yes, for governed segment enrichment. |
ai_model | Future AI source for summaries, classifications, and rationales. | No. Keep disabled until evidence and policy are proven. |
End-to-end workflow
- Choose the target. Start with one company or a materialised segment. Do not enrich an unbounded query.
- Check source status. In the app, open
Admin → Enrichment Ops. Disabled and degraded sources fail closed. - Create a run. The engine records source, task type, target, requester, policy version, and counters.
- Build context. The server gathers company profile data, SIC codes, sponsor links, address context, prior evidence, and known conflicts where policy allows.
- Collect evidence. Evidence and observations are stored before interpretation. Raw provider payloads must not be exposed to the browser.
- Apply policy. The policy engine assigns confidence, review requirement, canonical-promotion eligibility, and downstream blocks for exports or Direct Mail.
- Review in the workspace. Pending claims appear in workspace Enrichment Review. Approval makes a claim usable in that workspace only.
- Promote only when eligible. Platform admins decide canonical promotion separately in Canonical Promotions.
- Use downstream. Approved workspace claims can power filters, segments, signals, exports, and Direct Mail readiness.
Do we need an LLM?
Not to prove the trust spine. Companies House, sponsor licence data, and workspace context are enough to test run creation, context capture, evidence, policy, review, approval, and downstream filtering.
LLMs become useful for summaries, research packets, classification explanations, prospecting rationales, and review assistance. AI should run server-side through an approved adapter, never from the browser. AI-only outputs should stay capped at medium confidence and require review unless stronger non-AI evidence supports them.
Downstream acceptance path
A successful enrichment test is not just a completed run. The run must produce workspace-safe context that downstream modules can use without mixing draft, rejected, or platform-only claims into operational files.
- Open the segment detail page and confirm the latest enrichment run is complete with processed inputs and no run-level errors.
- Review pending claims when policy requires review. If there are no pending claims, confirm the run only produced already-approved, no-review, or non-actionable context.
- Generate an export from the same segment. The CSV may include approved enrichment context and sponsor fields, but it must not include rejected claims or raw provider payloads.
- Create a Direct Mail campaign from the segment, materialise recipients, quote, approve, and queue. Recipient merge data should use only approved workspace context.
- As platform admin, inspect workspace pages in read-only mode to verify data visibility. Inspection should show evidence and downstream readiness, but mutation controls should be disabled unless the admin is also a real workspace member.
Recommended first test
- Create or use a segment with 5 to 10 companies.
- Include at least one company matched to a sponsor licence.
- Start segment enrichment.
- Process the run from Enrichment Ops.
- Open the run detail and confirm observations, claims, and evidence.
- Approve one claim and reject one claim in workspace review.
- Confirm approved enrichment appears in Intelligence Studio filters.
- Confirm exports and Direct Mail only use approved claims.
What not to do yet
- Do not connect an LLM before the deterministic test passes.
- Do not enrich thousands of companies until retries, stuck-run handling, and quotas are verified.
- Do not promote workspace claims to canonical unless the source policy says they are eligible.
- Do not expose enriched claims on public SEO pages without a separate allowlist.