AI research belongs in a CRM field, not in the email

Most AI outreach tools research a prospect at send time and spend the finding on one sentence in one email. Write the same finding into a structured CRM property and every campaign after it can read it: BMG LABTECH went from roughly 200 contacts carrying a usable field of research to 59% of a 70,000-contact database, and over a year of that kind of data ops their newsletter open rate moved from 15% to 30%. Same research, different destination, and only one of those destinations still exists next quarter.

What should an AI research agent produce? A structured field.

A research agent should write its finding to a property on the record, not into the draft of an email. A property can be filtered, segmented, routed, reported on, corrected by a human, and read by the next agent. A personalized first line is consumed on send and leaves nothing behind.

BMG LABTECH makes instruments for laboratories, sells to scientists, and collects about 40 new form submissions a day against a database of 70,000 contacts. Markus Hartmann runs their entire digital marketing operation alone. The segmentation dimension that mattered to them, the researcher's field of research, was empty for almost everyone: about 200 contacts out of 70,000 had it.

We built an agent that takes a contact's name, searches four years of scientific publications, classifies the researcher's field, and writes the result to a HubSpot dropdown property. Coverage went from roughly nothing to 59% of the database. Each field-of-research segment went from 100 or 200 addresses, too small to bother with, to 2,000 or 3,000. Over a year, with that agent as one of seven pipelines running continuously, the open rate doubled from 15% to 30%.

Nothing in that result required writing a single personalized sentence. The oncology white paper simply stopped going to microbiologists.

A field compounds. A prompt is spent.

The same research costs the same to produce either way, so the destination is where the return is decided. Once a finding sits in a property, it does work you never paid for again:

• Every future campaign segments on it, not just the one you were writing.
• Lead prioritization and routing can use it, because workflows read properties and cannot read the inside of an email.
• It shows up in reporting, so you can see how much of your market is actually addressable.
• A human can look at it, disagree, and correct it. Nobody audits a sentence that was sent three weeks ago.
• The next agent reads it as input, which is why the second agent you build costs less than the first.
• It survives the rep who leaves. The knowledge is in the system rather than in someone's browser tabs.

This is the practical version of the argument that agents fail on data rather than on models: an agent is only as useful as the structured context underneath it. Research written to fields is that context being built, one property at a time.

Which fields? The ones no data vendor sells you.

In a niche technical market the question that decides fit is almost never firmographic, which is exactly why standard enrichment does not answer it. Apollo, ZoomInfo and Cognism sell you industry, headcount and revenue. HubSpot discontinued its own automatic company enrichment. None of them know which materials a research group works with. That information is public, but it is scattered across publications, grant databases, lab pages and consortium records rather than packaged as a dataset, and assembling it per contact used to cost an afternoon of a rep's time.

Market	The question that decides fit	Where the answer actually lives
Life-science instruments	What field does this researcher work in?	Four years of their publications
Large research institutions	Which department inside "University of X" is the buying unit?	Institutional and group pages
Semiconductor R&D equipment	Which materials, devices and processes does the group work on?	Papers and lab pages
Semiconductor and deeptech accounts	Where on the value chain does this organization sit?	What the organization publicly does
Academic buyers	Does this lab hold active grants, and for how much?	Public grant databases (SBIR, NIH, EU)
Venture-stage buyers	Who funds them, at what typical check size?	Investor profiles

Every row is a field we have had to build for someone, because their market qualifies on a technical dimension no provider carries. The method is the same each time: pick the dimension that decides whether you are relevant at all, then build one narrow agent that answers it and writes to one property.

Name the decision before you build the field.

Build a research field only when you can name the decision it changes, otherwise you are paying machine time for trivia. Field of research existed to decide which content goes to whom. The grant field exists so a rep knows which funded labs to call first: a funding picture like "5 active grants, $15M total" turns an anonymous academic contact into a prioritized one. Both had a decision waiting for them before the agent was written.

The inverse of that discipline is the enrich-everything reflex, and I would argue against it. A filled property that no campaign, workflow or report reads is a cost with a maintenance bill attached. Two questions settle it: which decision changes when this field is populated, and who makes that decision.

Coverage is the other place teams get stuck. Nobody reaches 100%, and waiting for it is worse than shipping at 59%, because 59% of 70,000 contacts is 2,000 to 3,000 addressable people per segment where there were 200. Our position on this has not changed: aim for the coverage that makes the decision possible, not for perfect.

One caveat that costs real money if you miss it. A field is an asset only while it is true. When BMG LABTECH ran their December 2025 Christmas list through job monitoring, 2,368 contacts went in, 1,290 matched to a LinkedIn profile, and 214 of those, 17% of everyone verified, had left the company they were recorded at. Those emails were not bouncing yet. They were just not being read. Research that is written once and never re-checked decays at the same rate as the rest of your CRM.

In a thousand-account market, only the stored research accumulates.

If your total market is around a thousand buyers, you will eventually touch all of them with your own hands, and most of them are scientists who remember who wasted their time. Volume tactics run out of market. What accumulates instead is the structured knowledge you kept: which researcher works on what, which lab holds funding, which department buys, which value-chain position the account occupies. Sends are spent. Fields stay, and they make the next send better without you doing the research again.

That is the operational half of the case for depth rather than volume in deeptech and life-science markets. The strategy is knowing more before you send anything. The implementation is a column in your CRM, filled by an agent, read by everything downstream.

If you sell into a market like this and cannot name the three fields that would decide your prioritization, that is the conversation to have. Ours starts with Diagnostics and Scoping: $2,500, credited toward the build, and you leave with the field list whether or not we build it.

FAQ

Should AI-generated prospect research go into the email or into the CRM?

Into the CRM, as a structured property, and into the email only as a byproduct. Research stored on the record can be segmented, routed, reported on, corrected and reused by every later campaign and agent, while research spent inside one email draft is gone on send. The production cost is identical either way, so the destination decides whether the work compounds.

What fields do niche B2B markets need that standard enrichment does not provide?

The dimension that decides technical fit, which is never firmographic. Fields we have built: a life-science researcher's field of research from their publications, the department inside a large research institution that holds the budget, the materials and processes a semiconductor research group works with, an organization's position on the semiconductor value chain, active research grants and their totals, and a venture-stage account's investor profile and typical check size. Apollo, ZoomInfo, Cognism and HubSpot's own enrichment carry none of these, though the underlying information is public in publications, grant databases and lab pages.

How much field coverage is enough for segmentation to work?

Enough that the decision the field exists for becomes possible, which is usually far below 100%. Field-of-research coverage of 59% across 70,000 contacts produced segments of 2,000 to 3,000 people each, where the previous 200 populated records supported no segmentation at all. Set the threshold from the decision, ship when you clear it, and keep the enrichment running rather than treating it as a one-time backfill.

Does an enrichment field expire?

Yes, and the rate is faster than most teams assume. Employment is the worst offender: on one 2,368-contact campaign list, 1,290 contacts could be verified against LinkedIn and 214 of them, 17%, no longer worked where the CRM said they did. Scientific fields of research are more stable than job titles, and grant funding changes on its own cycle, so each research field needs a refresh interval matched to how fast the underlying fact moves.