AI Research Agent Tiering: Which Accounts Get the Full Pass

AI research agent tiering matches research depth to an account's fit and signal strength: a full pass, a light pass or template fields only.

ON THIS PAGE · 8 SECTIONS
  1. What AI research agent account tiering actually is
  2. Which accounts should get the full research pass?
  3. The three tiers, side by side
  4. The cutoff rule: fit score, signal strength and deal size
  5. How much does a full AI research agent pass actually cost?
  6. Building the tier router in four steps
  7. Where tiering breaks down
  8. Questions about AI research agent account tiering

THE SHORT VERSION

  • AI research agent account tiering means matching how many steps, sources and minutes an agent spends per account to that account's fit and signal strength, not running every account through the same expensive pass.
  • A full deep research pass costs $1 to $7 in model and search spend alone, before anyone reviews the output. Run that on a 2,000-account list and you have spent real budget on accounts that were never going to reply.
  • Three tiers cover most teams: a full agent pass with committee mapping for tier 1, one or two targeted lookups for tier 2, template fields only for tier 3.
  • The cutoff is fit score plus signal strength, not headcount or revenue alone.
  • Every full pass still needs a human skim before it reaches a rep. The failure mode is a confident wrong fact, not a missing one.

AI research agent account tiering is the rule set that decides how deep an agent digs on a given account before a rep ever sees the brief. A tier 1 account gets a full multi-step pass: buying committee, recent triggers, stack changes, a drafted opener. A tier 3 account gets fields pulled from a template and nothing more. The point is to spend agent time and review time where a deal is actually likely, not spread it evenly across a list where most names will never open an email.

What AI research agent account tiering actually is

Most teams that add an AI research agent make the same first mistake: they point it at the whole account list and let it run the same workflow on every row. The agent does not know that a thirty-person company with no signal and a six hundred-person company mid hiring spree for the exact role your product replaces deserve different amounts of attention. Left alone, it treats them the same, because nobody told it otherwise.

Tiering fixes that by attaching a research depth to each account before the agent runs, not after. The tier decides the prompt, the number of sources the agent is allowed to check and whether a human reviews the output before it reaches a rep. This is the same idea account tiering has always covered in B2B sales (see our account tiering primer), applied one layer deeper: tiering the research itself, not just the outreach cadence that follows it.

Which accounts should get the full research pass?

An account earns the full pass when two conditions hold at once: it scores well against your ICP and it is showing a live signal right now, a funding round, a hiring spike in the right role, a stack change, a champion who moved jobs. Fit alone is not enough. Teams that skip the signal half end up running expensive passes on accounts that match on paper but have given no real reason to buy this quarter. A perfect-fit account with no signal this week is a tier 2 account that waits for one. A weaker-fit account with a strong signal can still earn a light tier 2 pass, but rarely the full one.

If you already run an ICP scoring model, the fit half of this decision is already built. The signal half comes from whatever you track in your signal system: job posts, funding events, technographic changes, champion moves. Multiply the two and you get a short list, usually under a fifth of the full account base, that justifies the cost of a deep pass.

The three tiers, side by side

TierResearch depthTypical time per accountWho reviews it
Tier 1Full agent pass: committee mapping, trigger detail, stack check, drafted opener20 to 30 minutesHuman skim before send
Tier 2One or two targeted lookups: confirm fit, pull one recent fact5 to 10 minutesSpot check on a sample
Tier 3Template fields only, no agent research step1 to 3 minutesNone, caught by reply data instead

Research depth and review load drop fast as you move down the tiers. Time estimates per firstsales.io, 2026.

Fit score and signal strength feed a router that decides which research prompt an account gets, before the agent runs.

The cutoff rule: fit score, signal strength and deal size

Pick numbers before you need them, not while a rep is arguing for an exception. A workable starting rule: tier 1 requires an ICP fit score in your top band (commonly the top 20 to 30 percent of your scoring model) plus at least one active signal less than two weeks old. Tier 2 is everything with a mid-band fit score or a top-band account with no current signal. Tier 3 is the rest, plus any account that already failed a deal in the last twelve months without a material change. Write the three bands down with the exact score and date cutoffs, then run new accounts against the written rule instead of a manager's gut call each week.

  • Deal size changes the thresholds, not the structure. A $50,000 deal can justify a full pass at a wider fit band than a $7,500 deal.
  • Write the rule down with the exact score cutoffs, not "accounts that feel important." Reps will argue every account up a tier if the line is not explicit.
  • Re-score on a schedule, weekly for signals, monthly for fit. A tier is a snapshot, not a permanent label.

How much does a full AI research agent pass actually cost?

$1 to $7cost of one deep research agent task, standard to max depth
20 to 30 minhuman time a tier 1 account earns for research and review
$0.65 to $1.20cost to enrich one contact on a standard data waterfall

Sources: tokencost.app on Google's April 2026 Gemini Deep Research pricing, firstsales.io, 2026, landbase.com on Clay's March 2026 pricing update.

Those numbers compound fast. Worked example: a 2,000-account list split 10 percent tier 1, 35 percent tier 2, 55 percent tier 3 runs roughly 200 full passes, 700 light lookups and 1,100 template fills. At the low end of the ranges above, that is about $400 in agent compute for tier 1 plus around 80 hours of review time, versus the same $400 and 80 hours spent flat across all 2,000 accounts if nobody tiered at all, most of it on accounts with no fit and no signal. The difference is not the total spend. It is where the spend lands.

Building the tier router in four steps

  1. Score fit first. Run your ICP model before the agent touches the account. Tiering on top of a weak fit score just automates a bad call faster.
  2. Attach a signal feed. Pull the active signal, if any, onto the account record so the router can see it at the same time as the fit score.
  3. Write the prompt per tier. A tier 1 prompt asks for committee mapping and a drafted opener, built the way we lay out in our research agent prompt template. A tier 2 prompt asks for one confirmed fact. Do not run one prompt and trim the output after, the agent still spends the tokens either way.
  4. Route the review step with the tier. Tier 1 output goes to a human before it reaches a sequence. Tier 2 gets sampled. Tier 3 skips review and relies on reply data to flag bad fields over time.

FROM THE YARD

The first time we ran a full agent pass across an entire account list instead of just the top tier, we found the real cost two weeks in, not on day one. The agent was confidently wrong about a company's funding stage or recent leadership change in roughly one account out of twenty, enough that a prospect corrected us in a reply before we caught it ourselves. Since then every tier 1 brief gets a two-minute human skim before it reaches a rep. Tier 2 and tier 3 skip that skim because the cost of being wrong there is low enough that reply data catches it instead.

Where tiering breaks down

Tiering is not free of failure modes of its own. The most common one is tier drift: an account earns tier 1 for a signal that fades, but nobody moves it back down, so it keeps eating full-pass budget on a cold trail. The fix is the re-scoring cadence from the cutoff rule above, not a one-time tag.

The second is over-engineering the split. Four or five tiers sound more careful than three, but they add router complexity and confuse reps who now have to remember five different outreach plans instead of three. Three tiers cover almost every team we have seen run this well and a fourth tier is usually a sign someone wants an exception process more than a real new category.

The third is treating tier 3 as fully unattended forever. A tier 3 account can throw a strong signal tomorrow and the router needs to re-check it against the signal feed on the same schedule as everything else or it stays a permanent dead pile instead of a holding tier. The fourth is letting the agent's confidence stand in for accuracy: a fluent, well-formatted brief reads as trustworthy even when a fact in it is wrong, which is exactly why the review step in the steps above exists for tier 1 and why reply data has to do that job for tier 2 and tier 3 instead of nothing at all.

Questions about AI research agent account tiering

How many tiers should an outbound account list have?

Three works for almost every team: a full research pass, a light lookup and a template-only tier. More tiers add router complexity and give reps more plans to remember without meaningfully improving targeting.

Does every tier 1 account need a human to review the AI brief?

Yes, at least a short skim. Tier 1 is where a confidently wrong fact does the most damage, since it reaches a prospect in a personalized send rather than a template field nobody scrutinizes.

What happens to the tier when a signal fires on a tier 3 account?

It should move up immediately if the fit score supports it. The router needs to re-check every tier against the live signal feed on a schedule, not just score tier 1 and tier 2 and leave tier 3 untouched.

Should tiering be based on company size or fit score?

Fit score, with signal strength as the second input. Company size alone misses the smaller account with a live trigger and the large account that has never matched your ICP.

Does a full AI research pass replace human research entirely?

No. It replaces the manual digging a rep used to do by hand, not the final check. A person still needs to confirm the agent's facts before a tier 1 account gets a personalized send.

Sources: tokencost.app, firstsales.io, landbase.com.

WRITTEN BY

Hlib Storchak

Founder of Shipyard GTM. Builds and runs outbound systems for B2B teams from Riga. 2000+ meetings booked for clients so far.

BUILD IT FOR YOUR MARKET

Want this running for your team?

30 minutes. We map your market on the call and show you how we would build your outbound system.

Book a call