Research-backed calculator guide

AI Prospect Research Cost per Account: Price the Research Brief, Not the Seat

A practical method to calculate AI prospect research cost per account by combining model usage, web-search calls, human review, fixed tools, and exception reserves.

What “per account” should include

Define the deliverable before collecting prices. A completed account brief might contain firmographic fit, buying signals, relevant people, recent events, evidence links, a qualification score, a next-step recommendation, and a human approval. If your system only creates a draft paragraph, call the unit a draft. If it creates a cited brief that a person checks and sends to a CRM, call it a completed account brief. Mixing those units makes any ROI comparison unreliable.

A practical cost ledger has five layers:

1. Model work. Count input and output tokens for planning, extraction, synthesis, scoring, and formatting. Include repeated prompts and any long account context.
2. Retrieval and data. Count search calls, page fetches, enrichment lookups, CRM requests, and paid databases. A “free” browser step may still carry engineering, rate-limit, or subscription cost.
3. Exceptions. Add retries, failed identity matches, blocked pages, duplicate accounts, escalations, and re-runs. A clean demo is not a production average.
4. Human review. Price the time needed to verify identity, inspect sources, correct the brief, and approve CRM writeback. This is usually the largest variable.
5. Fixed allocation. Allocate monthly platform, monitoring, proxy, storage, and integration costs across the number of completed briefs, not the number of seats purchased.

The denominator should be accepted briefs, not attempted runs. If 1,000 jobs are started but 80 are rejected or abandoned, divide spend by 920 accepted accounts. That convention exposes quality loss instead of hiding it in a flattering unit cost.

A formula you can audit

Use this spreadsheet-friendly structure:

Cost per accepted account = [(I × Pi) + (O × Po) + (S × Ps) + D] × (1 + r) + (M/60 × H) + (F/N)

Where:

• I and O are input and output tokens per account.
• Pi and Po are the respective prices per token, converted from the provider’s price per million tokens.
• S is the number of paid search or retrieval calls and Ps is price per call.
• D is other variable data or API cost per account.
• r is the retry and exception reserve.
• M is human review minutes per accepted account.
• H is the hourly loaded labor estimate.
• F is monthly fixed workflow cost and N is accepted accounts per month.

OpenAI’s published API page shows separate input, cached-input, output, and web-search meters. On the checked page, web search is listed at $10 per 1,000 calls, while model prices vary by model and processing mode ^[1]. Anthropic likewise lists input and output prices by model, prompt-cache rates, and a 50% batch-processing saving; those are provider terms, not assumptions to silently transfer to another model ^[2]. Google’s Gemini pricing documentation also separates token prices from grounding requests and notes that one submitted request can produce multiple search queries ^[3]. In other words, count the meter that actually fires.

Do not compare a per-seat subscription directly with a per-account API estimate. A seat may include workflow, data, support, storage, permissions, and a usage allowance. Convert it to a monthly fixed allocation only after deciding what portion supports prospect research and how many accepted briefs consume it.

Worked cost example

Assume a small U.S. sales-research workflow produces 50 accepted account briefs per month. The brief uses a lower-cost model for synthesis, eight web-search calls, and a human reviewer. The assumptions below are intentionally visible so you can replace them.

The $0.10 and $0.50 token rates in this example are taken from the checked OpenAI page’s published GPT-6 Luna rates, and the $10 per 1,000 web-search-call rate is also from that page ^[1]. They are dated inputs, not a promise that every model or provider will have the same price. The $45 fixed allocation is a scenario assumption, not a vendor quote. It represents the portion of a hypothetical monthly workflow subscription assigned to this job.

For labor, the U.S. Bureau of Labor Statistics reports a 2025 median of $78,760 per year, or $37.87 per hour, for market research analysts ^[4]. That occupation is only a proxy for account research; it is not a quote for your employee, contractor, or agency. I apply a 1.30 loading factor for payroll burden, benefits, management, and paid non-working time, giving a modeled $49.231 per hour. Replace it with your actual loaded rate or regional benchmark.

At 50 accepted briefs, the estimated monthly spend is $295.72. A manual baseline of 25 minutes per account at the same $49.231 loaded hourly rate is $20.51 per account. The modeled workflow therefore saves about $14.60 per accepted brief, or 71.2% against that labor-only baseline. That is not a forecast of revenue, conversion, or payback. It is a controlled comparison of stated assumptions. If review expands from six to 15 minutes, the unit cost rises by about $7.38 before changing any token or search rate.

Why the seat is the wrong denominator

A seat-based question asks, “What does this software cost per user?” A workflow question asks, “What does one accepted account brief consume?” The second question is better for capacity planning because two teams with the same subscription can produce very different unit economics.

Suppose Team A completes 50 briefs and Team B completes 500 while both allocate the same $45 fixed cost. The allocation is $0.90 per brief for Team A but only $0.09 for Team B. That tenfold difference appears before considering different review times or data depth. Under-utilization is therefore a cost driver, even if the vendor’s invoice does not change.

The same logic applies to a research package with unused credits. If a plan allows 300 reports but you accept 80, divide the attributable monthly fee by 80 accepted reports, not by 300 included credits. You may still retain the unused capacity for growth, but it should not be mistaken for current efficiency.

For a more complete ownership model, pair this guide with the AI agent budget guardrail and P95 calculator. Averages can look safe while a small number of large accounts create the month’s worst spend. Use p95 token, search, and review minutes when setting a budget ceiling. If the workflow needs custom engineering before it can produce an accepted brief, estimate that work with the AI agent development cost estimator. If invoices do not match estimates, reconcile usage by model, tool, and account in the AI API invoice reconciliation calculator.

When AI prospect research is economical

AI tends to be most useful when account volume is high enough to spread fixed costs, the research rubric is repeatable, and a person can review a concise evidence trail. It is less attractive when every account needs bespoke investigation, sensitive data handling, or senior judgment that cannot be reduced to a six-minute review.

Track three operational rates alongside cost:

• Acceptance rate: accepted briefs divided by started jobs.
• Review minutes: median and p95 minutes per accepted brief.
• Useful-action rate: briefs that lead to a qualified next step divided by accepted briefs.

The third measure is not a revenue claim. It is a workflow-quality measure that prevents a cheap but unusable brief from winning the spreadsheet. You can also calculate effective cost per useful action as total monthly research spend divided by the number of briefs that pass your chosen action threshold.

Build a small sample before committing to a target. Log 30–50 accounts across easy, average, and difficult cases. Record tokens, search calls, elapsed time, review minutes, rejection reasons, and any paid enrichment. Then replace the scenario assumptions in the formula with observed medians and p95 values. Keep separate cohorts for named-account research, territory research, and accounts with missing identity data; otherwise an easy cohort can conceal a costly one.

What this estimate excludes

The worked example excludes CRM licensing, premium contact databases, proxy or browser infrastructure, compliance review, data-cleaning projects, sales enablement, outbound sending, and the opportunity cost of a delayed handoff. Add each item only when it is genuinely attributable to the account-research workflow. It also excludes revenue, conversion, pipeline value, and payback claims. Those require your own outcome data and a defined attribution window.

Provider pricing changes, promotions, batch discounts, caching, regional processing, and minimum commitments can change the variable cost. Anthropic documents batch, caching, and regional pricing distinctions ^[2]. Google documents free request allowances, paid grounding, and date-bounded prices on its pricing page ^[3]. Record the provider page, model name, unit, and checked date in your ledger. Never store “AI cost” as one unqualified number.

FAQ

What is a reasonable AI prospect research cost per account?

There is no universal number. The defensible answer is your fully loaded cost per accepted brief after adding model usage, retrieval, retries, human review, and fixed-cost allocation. In the worked scenario here, that estimate is $5.91 per account, but the figure depends heavily on six review minutes, 50 accepted briefs, and the stated provider rates.

Are API tokens the main cost of an account brief?

Often they are not. In the worked example, model tokens cost $0.003 per account and eight search calls cost $0.08 before the retry reserve. Human review contributes $4.92 and fixed allocation contributes $0.90. Deep prompts, premium databases, or long review queues can reverse that mix, so measure rather than assume.

Should I divide a software subscription by seats or accounts?

For this decision, allocate the portion of the subscription that supports prospect research across accepted accounts. Dividing by seats answers a licensing question. Dividing by accepted briefs answers a unit-cost question. Keep both views if finance needs the seat cost, but do not substitute one for the other.

How do I include failed or rejected research?

Use accepted briefs as the denominator and put failed work in the retry or exception reserve. If 100 runs produce 85 accepted briefs, total spend belongs over 85, not 100. Also record rejection reasons. A high rejection rate may call for better identity resolution or a narrower brief rather than a cheaper model.

Can I use vendor report pricing as my cost per account?

Only if the vendor defines one completed report in a way that matches your deliverable and you add excluded costs. A list price can be a useful lower-bound input, but it may omit human review, search overages, implementation, data licenses, or unused-capacity effects. Label it as a vendor quote or published list price and keep your own fully loaded estimate separate.

What should I monitor after launch?

Monitor accepted briefs, review minutes, p95 tokens, p95 search calls, rejection rate, data-provider spend, and effective cost per useful action. Review the model and price assumptions whenever a provider changes its pricing page, when prompt length changes, or when the workflow adds a new retrieval or enrichment step.