What a Marketing AI Agent Costs Per Contact, and When It Beats Manual Work

Selling finance on agents comes down to three numbers: a cost per contact, a cost for the manual process the agent replaces, and the volume at which one crosses the other. Everything below is built to produce those three, with every input visible so someone can challenge it.

If you need background on what separates an agent from a scripted workflow, start with our explainer on what agentic AI is. This piece assumes you already have a workflow in mind and want to price it.

What does one agent run cost?

An AI marketing agent's cost per run, priced on LLM tokens alone, is the sum of input and output tokens across every model call in the run, multiplied by the provider's per-million rates. It seems simple enough until you count the calls. A lead-research agent might classify the contact, pull and summarise a company page, check fit against your ICP, draft an email, critique the draft and revise it. That's six calls, and each one reloads instructions and context.

Anthropic's engineering team is blunt about the trade. In its guide to building effective agents, Anthropic notes that agentic systems often trade latency and cost for better task performance, and recommends the simplest solution that works. Its evaluator-optimizer pattern, where one call drafts and another critiques in a loop, is exactly the kind of design that quietly doubles token spend, so price the loop before you build it.

Current list prices supply the inputs. According to Datafloq, as of June 2026 OpenAI's GPT-5.4 API costs $2.50 per million input tokens and $15 per million output tokens, GPT-5.4 Nano costs $0.20 and $1.25, and batch processing cuts standard pricing by 50% for asynchronous work. The same guide reports prompt caching reduces cached-input cost by 90% on Anthropic's models. Rates move often, so treat these as placeholders and check the provider's own pricing page on the day you model.

What goes into a cost-per-contact formula?

A cost model for AI marketing agents per contact starts with tokens and LLM API pricing for 2026, but in total it needs five variable terms and one fixed term. Leave any of them out and the pilot will look cheaper than production.

Cost per contact = (LLM + Tools + Platform + Review) + Fixed ÷ monthly volume

• LLM = Σ over calls of (input tokens × input rate + output tokens × output rate) × (1 + retry rate). Retries include malformed outputs, failed validation gates and extra evaluator loops.
• Tools = paid enrichment lookups, search API calls, verification services, each at your contracted unit price.
• Platform = orchestration fees or credits per action. On credit-metered platforms this term can dwarf the model bill.
• Review = share of outputs a human checks × minutes per check × loaded cost per minute, plus rework on rejected outputs.
• Fixed = prompt maintenance, evals, logging, monitoring and subscriptions, spread across monthly volume.

The platform term deserves its own warning. Aissist, in its September 2026 pricing benchmark, works through Salesforce's Flex Credits at $0.005 each with 20 credits per Agentforce action, which makes every action $0.10. It finds that the same resolved conversation costs between $0.49 and $4.19 depending only on how many actions the agent takes, which leaves the price in the agent's hands rather than the buyer's. Aissist also shows how per-session pricing at a 25% success rate becomes roughly four times the sticker price, and you should apply the same logic to drafts that fail review.

What happens to cost at scale once humans review the work?

The cost of running AI agents at scale is mostly token cost per task until you add human-in-the-loop oversight, at which point review minutes take over. At low volume you review everything because you have no quality evidence yet. Once acceptance rates stabilise, you can move to sampling. That shift is the single biggest driver of falling unit cost, and it should be earned with data.

The approval points themselves are a governance decision. Our piece on the governance operating model for marketing agents covers how teams set permission tiers and decide when an agent may ship without a human. If the agent writes to your CRM, the review burden also depends on what it is allowed to change; see what marketing agents can write back to your CRM.

How do the numbers work out at 1k, 10k and 100k contacts?

Hypothetical example. All volumes, vendor rates and labour costs below are illustrative assumptions.

The workflow researches an inbound or list contact, checks ICP fit, drafts a personalised first-touch email, self-critiques and revises. That's six model calls per contact, averaging 4,000 input tokens and 600 output tokens each, with a 15% retry rate.

Token cost, two designs. Running all six calls on GPT-5.4 at the Datafloq-reported rates costs about $0.114 per contact, or $0.131 with retries. Routing four simpler calls (classification, extraction, fit check, critique) to GPT-5.4 Nano and keeping drafting and revision on GPT-5.4 drops that to about $0.044, or $0.051 with retries. Anthropic describes this routing pattern directly: send easy requests to cheaper models and hard ones to more capable models. The example uses the routed design.

Other assumptions: enrichment $0.15 per contact plus two search calls at $0.01; orchestration $0.03 per run; review at $60 loaded hourly cost ($1 per minute), 1.5 minutes per check; fixed costs of $2,000 a month for maintenance, evals and logging. Review share falls from 100% at 1k to a 30% sample at 10k and 10% at 100k. The manual comparison is an SDR spending 8 minutes per contact at the same $1 per minute.

The model bill is the smallest line at every volume. At 1,000 contacts, review and fixed costs make up over 90% of the agent's cost, which is why pilots so often look disappointing on a per-unit basis.

Rework changes the picture less than people fear. If 20% of reviewed drafts are rejected and rewritten by hand at 6 minutes, add $1.20 per contact at 1k (to $4.95) and proportionally less at higher volumes as the reviewed share drops. Platform pricing is a bigger risk: on a credit-metered platform charging $0.10 per action, with the agent taking five actions per contact, the orchestration line would jump from $0.03 to $0.50, which matters far more at 100k than any model choice.

At what volume does the agent break even?

Break-even volume = Fixed monthly cost ÷ (manual cost per contact − agent variable cost per contact).

With full review, the agent's variable cost is $0.251 plus $1.50, so $1.75. Against the $8 manual cost, each contact saves about $6.25, and $2,000 of fixed cost is recovered at roughly 320 contacts a month. Below that volume, keep doing it by hand. If your real manual time is 3 minutes instead of 8, the saving shrinks to $1.25 and break-even climbs to about 1,600 contacts, which is why the manual figure has to come from a timed sample of real work.

Google's guidance on agentic AI in marketing makes a point finance will appreciate: measure how agents help the business. Cost per contact is the denominator. The numerator is reply rate, meetings booked or qualified pipeline per thousand contacts, compared against the manual baseline over the same period. An agent at $0.42 that books half as many meetings may lose to a human at $8. Our review of which B2B agent use cases have production evidence is a useful check on which outcomes are realistic to claim.

Where do cost models usually go wrong?

The cost to run an AI agent per user comes down to token cost plus tooling, and confusing that with per-contact cost is one of several mistakes that throw models off.

• Context creep. Over months, teams add examples and history to prompts. Input tokens per call double and nobody updates the model. Cache stable instructions and audit token logs monthly.
• Unbounded loops. Evaluator-optimizer cycles without a hard cap can run five or six passes on difficult contacts. Set a maximum and route failures to a human.
• Retries hidden by the framework. Anthropic warns that frameworks can obscure the underlying prompts and responses, so count retries from raw API logs.
• Review that never steps down. If review share stays at 100%, the 100k column never materialises. Define in advance the acceptance rate that triggers sampling.
• Per-user tooling mistaken for per-contact cost. Seat-based licences behave like fixed costs, so put them in the fixed line.

What should you give finance?

One sheet. It should hold the formula, your logged token counts from a week of real runs, contracted tool and platform rates, a timed manual baseline, the review-share schedule with the acceptance thresholds that justify each step, and the break-even volume under best and worst retry assumptions. Our AI ROI calculator is one place to structure those inputs. If you'd rather have the workflow scoped, priced and built with the approval points in place, that is the work our agentic AI automation team does. Either way, the decision should rest on the table, and the table should survive someone else plugging in their own numbers.

Sources

• Anthropic, Building effective agents
• Google, Infinite marketing in the age of agentic AI
• Aissist, AI Agent Pricing Benchmark 2026
• Datafloq, The Complete AI Model Guide 2026