Confidence, sources & selection

That accountability rides on three things attached to each value: a source, a confidence, and a selection_reason. This page explains what each one means, and how an AI agent should act on them.

Account-based marketing for AI agents. The point is that an agent can trust what it reads — or know, precisely, when not to.

The three signals on every value

An enrichment returns eighty-nine canonical fields — forty-three for a person, forty-six for a company — plus forty signals. Each value arrives with three things alongside it:

• Source — which provider the value came from. abm.dev queries ten sources, among them LinkedIn, Companies House, Perplexity, Tavily, and Hunter, and resolves them into one response. The source tells you who said it.
• Confidence — a score from 0 to 1. How sure we are, given everything we looked at.
• selection_reason — why that value was chosen over the alternatives.

No value travels without all three. If we could not cite it, we do not return it.

How to read confidence

Confidence runs from 0 to 1. Higher means more agreement across the sources we consulted. A value several sources concur on scores higher than one a single source reports. Conflicting or thin evidence scores lower.

That is the whole rule. We do not publish fixed bands or cut-off thresholds — confidence is a continuous signal, and the right line to draw depends on what your agent is about to do. Writing to a CRM warrants a higher bar than drafting a first line for review.

There is no 0-to-100 scale and there are no named tiers. One number, between zero and one. Read it as a probability, not a grade.

For the deeper treatment of confidence, see /docs/concepts/confidence-scores.

How to read sources

Ten sources go in. One answer comes out.

abm.dev resolves the ten providers into a single response so your agent reads one clean value, with the winning source named. No per-source bills. No per-field charges. No silent fallbacks. And no fabricated facts — a value is cited or it is not returned.

That last line is the load-bearing one. An empty field is an honest answer. A made-up field is not, so you will never get one.

On the providers and how resolution works, see /docs/concepts/data-sources. On the fields themselves, see /docs/concepts/canonical-fields.

How to read selection_reason

People change jobs. Companies change names. Sources disagree, and one value has to win.

selection_reason is the explanation attached to a value — why that value was chosen over the alternatives. It tells you what tipped the decision, not just what the decision was. Pair selection_reason with confidence before an agent acts: a well-reasoned, high-confidence value is safe to write; a thinly-reasoned or low-confidence one is worth a second look.

Where to see it

Each value’s source, confidence, and selection_reason come back on the enrichment itself. Two endpoints carry it:

POST https://api.abm.dev/api/v2/enrichments
GET  https://api.abm.dev/api/v2/enrichments/{id}

x-api-key: YOUR_API_KEY
Content-Type: application/json

A minimal request:

{ "targets": [ { "type": "person", "email": "[email protected]" } ] }

Each target also accepts a LinkedIn URL or a company domain.

For the exact response shape — the literal key names carrying source, confidence, and selection_reason — read the /api-reference/enrichment. For batch jobs, see /api-reference/jobs.

How agents consume it

• Claude — add the MCP server at https://mcp.abm.dev/mcp (Settings → Connectors → Custom → paste the URL). It exposes abm.dev’s enrichment tools inside Claude. For the tool list, see /for-agents.
• Cursor, Claude Code, Windsurf — tell the agent to read https://abm.dev/llms.txt, request a key, save a .env, run a test enrichment, and report back with citations.
• Scripts and the terminal — call the REST API directly. CLI specifics live at /docs/cli.

For long-running agent loops, results stream in real time over SSE. For batch jobs, a webhook fires when the job finishes. See /api-reference/jobs for the job mechanics.

A simple rule for action

Read all three before you act.

• Source tells you who said it.
• Confidence tells you how sure to be.
• selection_reason tells you why that value was chosen.

High confidence, a clear selection_reason, source you recognise — write it. Anything thinner — verify, or surface it for a human.

Start

• /docs/getting-started — your first enrichment.
• /dashboard/playground — run one free, no key.
• /#pricing — per-enrichment, no subscription, credits never expire, all ten sources included. From about €0.29 each. Free launch credits with code LAUNCHCODES.

abm.dev is operated by Foxley Farm Operations Ltd (UK Company No. 16392009), Oakham, United Kingdom. GDPR-compliant.

Questions about a specific value? Open a support ticket.