Skip to content

Case study · AI Agents & Automation

From scattered prospect research to a sales intelligence system

Prospecting runs on reps with browser tabs open: find a company, skim the site, guess at fit, move on. The research is shallow and inconsistent, and the buying signals that should change an account's priority go unseen. This is an agentic system that discovers businesses, researches them, detects buying signals, and produces sales-ready intelligence a team can act on.

The approach
  1. 01Business discovery
  2. 02Web research
  3. 03AI classification
  4. 04Signal detection
  5. 05Enrichment & scoring
  6. 06CRM-ready output
Client
a B2B services company
Problem
Prospect research is manual, shallow, and inconsistent, so reps act on thin signals.
Approach
An agentic pipeline: discover, research, enrich, classify, score, and prepare CRM-ready output.

The problem

Most prospecting runs on a rep with browser tabs open. They find a company, skim the site, guess at fit, and move on. The research is shallow because it is manual, and it is inconsistent because every rep does it differently. Buying signals that would change the priority of an account go unseen because nobody has time to look for them at scale.

The result is a pipeline ranked by whoever was researched most recently rather than by who is actually worth a conversation. The work becomes reliable only when discovery, research, and scoring run as one automated, consistent process.

The reality behind the problem

The obvious fix is to hire more reps or buy a bigger contact list. Neither addresses the problem. More reps multiply the same manual, inconsistent research, and a larger list adds contacts without the context that tells a rep which accounts are worth a conversation or when to start one. You end up with more activity on the same thin basis.

The harder part is that research only becomes rankable once it is structured. A rep's notes in a document cannot be compared across a thousand accounts, so the pipeline order reflects who had time to look, not who fits. And signals decay: a hire that mattered last month is stale now. A pipeline that reads each account once and never refreshes produces the same thin ranking the manual process did, just faster.

What was assumed

Better outreach needs more reps.

What Stallwart asked

What if research and scoring ran as a system, and reps spent their time on the accounts it surfaced?

The desired outcome

The target is not more activity. It is a pipeline ranked by fit and timing, where every account carries the research and signals behind its score, and a rep can trust that ranking enough to work it top-down.

  • Every target account researched on the same basis, so two prospects are comparable rather than judged by whoever happened to look at them.
  • Unstructured web research turned into structured attributes an engine can score.
  • Buying signals detected and dated, so timing is part of the ranking rather than an afterthought.
  • A fit-and-timing score a rep can act on, with the evidence behind it visible.
  • CRM-ready output: the context and a drafted opener attached to each account, not a bare contact.
  • A person approving anything that leaves the system, so outreach stays accountable.

The system we designed

This system discovers businesses that match a target profile, then researches each one's public digital presence: what it does, how it positions itself, and what it has changed recently. AI classification turns that unstructured research into structured attributes, and signal detection flags the events that matter, such as hiring, launches, or shifts in positioning.

Prospects are enriched and scored against the profile, so the pipeline is ranked by fit and timing rather than by recency. The output is prepared as CRM-ready intelligence with the context a rep needs to open a relevant conversation, including a first draft of personalized outreach for a human to approve.

How it works

  1. 01

    Make research structured, not anecdotal

    Unstructured web research is only useful at scale once it is turned into structured attributes an engine can rank. AI classification does that consistently across every account, so two prospects are compared on the same basis.

  2. 02

    Score on fit and timing, not recency

    A signal like a recent hire or launch changes an account's priority. Combining fit with detected signals ranks the pipeline by who is worth a conversation now, rather than who happened to be researched last.

  3. 03

    Keep a human on anything that is sent

    The system prepares outreach; it does not send it. A rep approves or edits, so personalization stays accountable and the brand's voice stays intact.

The pipeline

Discovery to CRM-ready intelligence.

  1. 01

    Business discovery

    Companies matching the target profile

  2. 02

    Web research

    Public digital presence, read per account

  3. 03

    AI classification

    Unstructured research into structured attributes

  4. 04

    Signal detection

    Hiring, launches, positioning changes

  5. 05

    Enrichment & scoring

    Ranked by fit and timing

  6. 06

    CRM-ready output

    Context + drafted outreach for human approval

Each stage is automated; a human approves outreach before anything is sent.

Architecture

What sits underneath.

Underneath the pipeline are four layers, each responsible for one part of turning a raw company into a scored, actionable account.

01

Discovery and source collection

Agents find companies matching the target profile and gather each one's public digital presence: the site, how it positions itself, recent changes, and the public signals that indicate activity.

02

Classification and enrichment

AI turns unstructured research into structured attributes, such as industry, model, size indicators, and positioning, and merges it with enrichment data so every account is described in the same fields.

03

Signal detection and scoring

Detected events such as hiring, launches, or positioning shifts are dated and weighted, then combined with fit to produce a score that reflects both who the account is and whether now is the moment.

04

Output and routing

Scored accounts are written to the CRM with their evidence and a drafted opener, and routed to the right owner, so a rep opens a queue of prepared accounts rather than a blank research task.

The hard parts

Where the engineering judgment was.

01

Agentic research with tool use, not a single prompt

Discovering and researching a company is a sequence of steps and lookups, which suits an agent that calls tools, not one model call. It makes the research auditable and each step replaceable.

The trade-off

More moving parts to run and observe than a single call, which is why evaluation and logging are part of the build.

02

Human-in-the-loop on outreach

Automated sending at scale is how a domain's reputation gets burned. Preparing drafts for approval keeps the throughput while a person owns what goes out.

The trade-off

A rep stays in the loop, so this augments outreach rather than fully replacing it.

Reliability & guardrails

How it avoids the wrong call.

The guardrails keep the system from ranking on bad data or acting without a person, because a confident score on a wrong fact is worse than no score.

Scores carry their evidence

Every score links back to the research and signals that produced it, so a rep can see why an account ranks where it does and catch a classification that got it wrong.

Signals are dated, not just present

A detected event is stored with when it happened, so a stale signal decays in the ranking instead of keeping an account high long after the moment has passed.

Source checks before classification

Research is validated against the right company before it is classified, so a near-name match or a parked page does not enter the pipeline as a real account.

Human approval on anything sent

The system prepares outreach and stops. A person approves or edits before a message leaves, so automated volume never puts the sending domain's reputation on autopilot.

Productionization

What turns it from a demo into software.

What turns the pipeline from a script that ran once into software a team relies on every day.

Agent runs are logged and replayable

Each research run records the steps, tool calls, and sources it used, so a surprising score can be traced and a failed run re-run from where it broke rather than from scratch.

Evaluation on classification and scoring

Classification and scoring are checked against a labeled set, so a prompt or model change that quietly degrades accuracy is caught before it reaches the pipeline.

Refresh on a schedule

Accounts and signals are re-researched on a cadence, so the ranking reflects the current state of the market instead of a snapshot that ages out.

Rate limits and cost controls

Discovery and research respect source rate limits and a per-run budget, so the pipeline scales without getting blocked or spending unpredictably.

What changed

The change is in what a rep opens each morning. Instead of a blank research task and a list ordered by recency, they get a queue of accounts already researched, scored on fit and timing, with the evidence and a drafted opener attached. The judgment moves from finding and skimming to deciding and reaching out.

Because every account is researched on the same basis, the pipeline becomes something a team can trust and reason about, rather than a reflection of who had time to look at what. The work scales without multiplying the inconsistency that manual research creates.

  • Research runs as a consistent system, not a rep with browser tabs.
  • The pipeline is ranked by fit and timing, not by whoever was looked at last.
  • Every account carries the evidence and signals behind its score.
  • A person approves outreach, so volume never runs unattended.

Ownership & handover

What you receive, and keep.

At handover, the pipeline and everything that runs it are yours.

  • The discovery, research, and classification agents, with their prompts and tool definitions.
  • The scoring model: the fit criteria, the signal weights, and how timing enters the rank.
  • The CRM integration and the output format reps work from.
  • Runbooks for adjusting the target profile, retuning scores, and re-running the pipeline.
  • The evaluation set and logging, so accuracy stays measurable as the system changes.

What we learned

The insight that generalizes: lead intelligence is a structuring problem before it is a scoring one. The hard part is not ranking accounts, it is turning inconsistent, unstructured research into the same fields for every company, with signals dated and evidence attached. Once the research is structured and current, scoring is straightforward and a rep can trust it. Get the structure and the refresh right, and the ranking takes care of itself.

This work connects to

  • AI Agents & Automation
  • AI sales automation
  • Lead enrichment and scoring
  • Agentic AI
  • Workflow automation

Frequently asked

What is AI lead intelligence?

AI lead intelligence is the use of AI agents to discover businesses, research their public digital presence, detect buying signals, and enrich and score prospects, turning manual prospect research into a consistent, ranked, CRM-ready process.

How is this different from buying a contact list?

A list gives you contacts. A lead-intelligence system gives you researched, scored accounts with the context and signals behind the ranking, so reps know why an account matters and when to reach out, not just who to email.

Does the system send outreach automatically?

No. It prepares personalized outreach and the supporting context, but a person approves or edits before anything is sent. That keeps throughput high without putting the brand's reputation on autopilot.

How are prospects scored?

Scoring combines fit with timing. Fit measures how closely an account matches the target profile across structured attributes such as industry, model, and size indicators. Timing comes from dated buying signals such as recent hiring, launches, or positioning changes. An account that fits well and shows a fresh signal ranks above one that fits but shows no activity, so reps work the accounts worth a conversation now.

Want prospect research to run as a system, not a browser tab?

Tell us your target profile and your CRM. We will scope the pipeline that discovers, scores, and prepares the accounts worth a conversation.

Last updated: October 13, 2026