Prospect research before the dial: how good data turns an AI voice agent from a robot into a rep

Why 'good enough' data is the difference between 3% and 12%

We've watched hundreds of outbound campaigns launch on Callable, and the pattern is remarkably consistent. Teams that upload a CSV with just phone numbers convert at 2-4%. Teams that upload phone numbers plus a first name and a company convert at 5-7%. Teams that add a single personalized data point — a recent job change, a neighborhood, a rough income band, a life event — clear 10-12% on the same voice, same script, same agent.

The agent is doing the same work in all three cases. The difference is entirely upstream: the prospect on the other end can tell within the first sentence whether they're being blasted or addressed. AI voice agents amplify that difference because they can actually use the data in real time — dropping the right first name, referencing the right city, adjusting the offer to the right income tier — without a human rep having to remember to check the CRM tab.

The three tiers of pre-call data

Not all enrichment is equal. Think of your pre-call data in three tiers, and try to hit at least tier two on every list you upload.

• Tier 1 — Identity: full name, direct phone number, current city or region. Non-negotiable. Without these your agent is calling a stranger with no way to sound like it isn't.
• Tier 2 — Context: employer, job title, approximate age band, homeowner vs renter, rough income or net-worth tier. This is where personalization starts to pay off — the agent can pick the right pitch, the right price tier, the right social proof.
• Tier 3 — Signal: recent life events (new home, new job, new baby, business registration), public affiliations, verified email. Signal turns a cold call into a warm one; a prospect who just moved into a new home is a very different lead for a solar or insurance pitch than one who's been in the same address for 15 years.

Where each tier actually comes from

For B2B lists, the standard stack is well-known: Apollo, Clay, ZoomInfo and LinkedIn Sales Navigator for firmographics and titles; Hunter or Findymail for verified emails; BuiltWith or Wappalyzer for tech stack signals. Any competent RevOps team can assemble a tier-two B2B list in an afternoon.

For B2C — which is where most AI voice deployments actually live, because consumer verticals like solar, insurance, home services, mortgage, real estate, financial planning and healthcare are where phone conversion still dominates — the tooling is thinner and more fragmented. This is the tier that most teams get wrong, because the B2B tools don't cover it and the classic consumer-data brokers gate everything behind six-figure minimums.

The public-records tier for B2C outbound

For consumer campaigns we recommend layering a public-records people-search tool on top of your list. These tools aggregate publicly available records — property assessor filings, business registrations, court records, voter rolls, professional licenses — and combine them into a single lookup by name, phone or address.

The one we've been pointing customers at recently is PersonPages. It's a global people-search index that returns identity, addresses, approximate net-worth and income banding, employer, and public affiliations for a name or phone number, in a lightweight interface that a non-technical ops person can actually use. For a solar or home-services team enriching a list of homeowners, or an insurance agent qualifying a leads file against approximate income tiers, it fills exactly the tier-two gap the B2B tools don't touch.

Using public-records data legally and ethically

Public-records enrichment is powerful, but it stacks with everything in our compliance guide. A quick reminder of the rules that apply the moment you enrich a list:

• Enrichment doesn't create a lawful basis to call — you still need consent (US mobile) or a documented legitimate interest (EU B2B). Enrichment sharpens the call; it doesn't authorize it.
• Never read enriched data back to the recipient verbatim. Saying 'I see you live at 42 Elm Street and earn about $180k' is legal but instantly kills the call. Use the data to pick the right pitch, not to prove you have it.
• GDPR data-subject access requests cover enriched fields too. If you store a net-worth band against a phone number, that field is exportable and deletable on request.
• Retention: purge enrichment fields on the same schedule as your transcripts. Old enrichment is worse than no enrichment.

Wiring enrichment into your Callable agent

Once you've enriched your list, the mechanics inside Callable are straightforward. Every column in your uploaded CSV becomes a variable the agent can reference in its prompt and first message. A prompt line like 'The prospect is {first_name}, a homeowner in {city} in the {income_band} bracket. Lead with the {tier}-tier offer' takes you from a generic pitch to a segmented one without touching the underlying agent.

The pattern we recommend: keep one master agent per use case, and let the data do the segmentation. Uploading three lists with different {tier} values will produce three noticeably different conversations from the same agent — with none of the version-management overhead of maintaining three separate scripts.

What to measure

Enrichment quality is one of the few outbound levers where the ROI shows up in the first 200 calls. Track connect rate, conversation length past the 30-second mark, and booking rate per data tier — not just per campaign. If tier-two enriched rows aren't outperforming tier-one rows by at least 2x on booking rate, your enrichment is stale or your agent isn't actually using the fields.

The teams that get this right treat the enrichment layer as a product surface, not a one-time list upload. Refresh monthly, retire fields that don't correlate with bookings, and add new signal fields when a vertical demands them. The agent gets smarter every cycle without a single line of prompt change.

The takeaway

AI voice agents are a force multiplier on your data, in both directions. Great data becomes great conversations at a scale a human team could never sustain. Weak data becomes weak conversations at exactly the same scale. Spend the afternoon on your enrichment stack before you spend the week on your prompt — the return is almost always higher.