“Intent data” is the most overloaded phrase in B2B. It means one thing to Bombora, another thing to G2, another thing to 6sense, and something completely different to a Claude agent calling tools over MCP.
I am George, founder of Leadpipe. Every week I get the same question from a team building an LLM-driven sales agent: “what intent data do I feed it?” The question sounds simple. It is not. Most intent products on the market were built for a human ABM marketer reading a quarterly report. An LLM agent needs a different shape, a different refresh rate, and a different granularity.
The short answer: agents need person-level intent, fresh within 24 hours, delivered as structured JSON over a tool call, with pages, topics, and a score attached. Anything else is a report an agent cannot act on.
Here is what that means in detail.
The four properties an LLM agent needs from intent data
An LLM agent is not a dashboard. It does not look at a leaderboard, nod, and go back to work. It reads a payload, reasons over fields, makes a decision, and calls another tool. That places hard requirements on the shape of the data.
| Property | What the agent needs | What most intent products deliver |
|---|---|---|
| Granularity | Person-level (named, with email) | Account-level surge score |
| Freshness | ≤24 hours from event | Weekly or monthly batch |
| Structure | JSON with typed fields | PDF report or dashboard CSV |
| Context | Pages viewed, topics, source domains | Aggregated score, opaque sources |
If an intent product misses any of these, it is still useful for a human running ABM. It is not useful for an agent writing an email at 3pm.
Granularity: person, not company
This is the biggest mismatch. Bombora and similar account-level intent products tell you “Acme Corp is surging on CRM migration.” Useful for an ABM marketer building a Q2 target list. Not useful for an agent that has to decide whether to email Sarah in Revenue or Mike in IT.
Agents have to address a person. “Acme Corp” does not have an email. Sarah Chen, VP Revenue at Acme Corp, does. And Sarah specifically, not Mike, is the one who was on your pricing page this morning and read two HubSpot comparison pieces across the web this week.
Person-level intent is what Leadpipe’s Orbit product was built for. Orbit reads a cross-site pixel network across 5M+ sites, resolves to individuals, and scores each person’s intent on 20,000+ topics. Daily refresh. Deterministic matching where the signal supports it. See Person-Level Intent Data: How It Works for the mechanics and why company-level intent is not enough for the background.
The test is simple: can your agent call the intent tool with “find VPs of Revenue at 50-500 person SaaS companies researching CRM migration this week” and get back a list of named people with business email, LinkedIn, and a score? If yes, it is person-level intent. If no, it is something else.
Freshness: 24 hours or less
An LLM agent is working in a loop. It runs, calls tools, sends, then runs again hours later. If the intent data is a week old, the agent is acting on signals that have already expired. The prospect already booked a demo with your competitor on Tuesday. It is Friday. The window closed.
Monthly and weekly batch intent is the norm at most incumbent providers. It was the norm because human ABM workflows ran on a monthly cadence. Agents do not. An agent operates at event-level cadence, which means intent data has to be fresh at event-level cadence too.
Leadpipe refreshes within 24 hours across the full graph. Visitor ID data fires via webhook within seconds of a person being identified on your site. Orbit person-level intent refreshes daily across 5M+ sites. If you are building an agent and your intent provider gives you a CSV on the first of every month, that is an ABM report, not an agent input.
Structure: JSON over tool calls
LLM agents read JSON. They do not read PDFs. They do not open dashboards. They call a tool, get a structured response, and reason over fields.
The minimum structured shape for person-level intent looks like this:
{
"person": {
"email": "sarah@acme.com",
"name": "Sarah Chen",
"title": "VP Revenue",
"linkedin": "linkedin.com/in/..."
},
"company": {
"domain": "acme.com",
"size": "200-500",
"industry": "SaaS"
},
"intent": {
"score": 87,
"topics": ["crm migration", "hubspot alternatives"],
"observed_at": "2026-04-22T14:23:11Z",
"sources": ["first-party pixel", "cross-site network"]
}
}That payload is readable by Claude, GPT, Gemini, any model you throw at it. Every field has a use. The agent can filter on score, group by topic, decide priority by freshness. Full schema in the webhook payload reference.
Leadpipe delivers this shape over four interfaces:
- 23 REST endpoints for direct query
- Real-time webhooks (First Match, Every Update) for push
- TypeScript SDK (
npm install @leadpipe/client) for code - 27-tool MCP server for agents talking over MCP
All four hit the same backend, so an agent does not have to pick a lane.
Context: pages, topics, sources
A score by itself is not context. A score plus a topic is slightly better. A score plus a topic plus the pages viewed on your own site and the source domains where the person was researching is the real thing.
An agent that sees “intent score 87” cannot write a good email. An agent that sees “intent score 87, matched topics: CRM migration and HubSpot alternatives, pages on your site: /pricing 3 min and /vs-hubspot 90 sec, return visit” can write a good email. The context turns a score into a narrative the model can riff off.
This is why I keep saying: agents need behavior, not just signal. The death of the lead form piece is the long version of why behavioral context matters more than ever in an agent world. The short version: when the agent writes the email, it needs something to reference. The form filled out six months ago is not it. The page read twenty minutes ago is.
Types of intent an LLM agent should pull
Put together, an LLM agent should pull from three intent sources, weighted differently depending on the motion:
| Intent source | What it tells the agent | Weight for agentic outbound |
|---|---|---|
| First-party visitor ID | This specific person just engaged with your site | Highest |
| Third-party person-level intent (Orbit) | This person is researching your category elsewhere | High |
| Third-party account-level intent (Bombora, G2) | This company is surging on your topics | Medium, for list building |
| CRM behavior | This known contact opened your last email or booked a call | High, for follow-up |
First-party beats third-party in every case because the signal is on your own property and the data ownership is cleaner (see the first-party data glossary for why). Third-party person-level is the next best input. Account-level intent is useful for building the target universe, not for writing the email.
The concrete example
Same agent, same model, same prompts, two different intent feeds.
Agent A feeds on Bombora-class account-level surge data, weekly refresh. Input to the email-writing tool:
Acme Corp is surging on "CRM migration" topic this week.The agent, given that alone, writes a generic pitch to a random VP at Acme. Reply rate on cold sends: 1 to 2%.
Agent B feeds on Leadpipe visitor ID plus Orbit person-level intent, 24-hour refresh. Input to the same tool:
Sarah Chen, VP Revenue at Acme Corp, visited /pricing for 3 minutes and
/vs-hubspot for 90 seconds this morning. Orbit shows she has been reading
CRM migration and HubSpot alternatives content across 4 other sites this
week. Intent score: 87. Return visit: yes.Same agent, same prompts. The output is a referenced, timely email to the right person. Reply rate on the identified segment: 10 to 20%. We see this pattern consistently across the Leadpipe customer base.
The model is the same. The delta is the intent shape.
Where most incumbents fall short for agent use
I want to be honest about the competitive landscape because I think it clarifies the gap.
- Bombora is excellent at what it does: account-level surge signals at scale. It is not person-level and the refresh is not daily. Not an agent input by itself.
- G2 intent is high-quality buyer-in-market signal, tied to G2 category pages. Narrow in scope, narrow in topic coverage. A useful supplement, not a base layer.
- 6sense (
$55K+/yr) and Demandbase ($55K+/yr) are ABM platforms, not agent-facing APIs. Account-level. Built for a marketing team, not a tool-calling agent. See Orbit vs Bombora for the deeper comparison. - Common Room (~$15-50K/yr) reads community signals (GitHub, Slack, Discord). Person-level in its niche, but not web-wide.
- Clearbit (now HubSpot Breeze) is company-level enrichment integrated into HubSpot, not an agent-facing intent API.
None of these are bad tools. They are not designed for an LLM agent calling tools over MCP. That is a different job, and it needs a different product.
Leadpipe as the intent layer for agents
Why we built Leadpipe to serve this specifically:
- Orbit delivers person-level intent across 5M+ sites, 20,000+ topics, daily refresh. See the Orbit launch post.
- Visitor ID covers first-party intent on your own site with 30-40%+ match rate on US B2B traffic, deterministic, 8.7/10 in the independent accuracy test.
- MCP server exposes 27 tools over the Model Context Protocol. Claude, Cursor, Codex, or any MCP client gets tool-level access to topic discovery, audience queries, visitor lookup, and pixel management. Install:
npx -y @leadpipe/mcp. - Webhooks deliver First Match and Every Update in real time. Structured JSON. No polling.
- Suppression and exclusion at the API level, which matters because you do not want your agent messaging customers or churned logos on an intent hit.
The underlying thesis is in Why Every AI Agent Needs an Identity API. The identity primitive and the intent primitive collapse into the same layer in an agent world, which is why we ship them together.
What to check before you wire intent into an agent
A short checklist:
- Does the intent data have a named person attached, not just a company?
- Is it fresh within 24 hours, not a monthly batch?
- Does it come back as structured JSON from an API or webhook, not a dashboard?
- Does it include topics, pages or sources, and a score, not just a score?
- Can you filter by ICP at the API level, not after the fact?
- Can you suppress customers and churned logos before the agent sees the record?
If the answer is no on more than one of these, the intent feed is not agent-ready. Fix the feed before you blame the model.
Every plan ships with the same identity graph, 23 REST endpoints, webhooks, and a 27-tool MCP server. Start in 5 minutes →
