Matthew Carlson · October 28, 2026 · 5 min readShort answer: about $4–8 and 0.8–1.4 million tokens per client, measured, not estimated. Metering every run is what turns "AI is expensive" into a unit cost you can put in a proposal.
What a research pull does
When a new client comes into the agency I work with, one prompt runs nine skills in sequence and in parallel: a crawl of the client's site, an external sweep of review platforms, directories and public records, a competitor scan, and then the packaging steps that produce seventeen deliverables, from a structured client profile to a CRM import file, a brand-voice guide and a 50-question FAQ. The whole pipeline runs as a directed graph: a serial build, a validation gate, a checksummed freeze of the source document, four readers running in parallel, and a single join at the end.
That's a lot of model calls and a lot of crawling. Without measurement it would be impossible to say what one client costs, which means impossible to price.
How the metering works
- Every agent run records its tokens and its model. The platform routes work across three Claude tiers depending on the step, so the cheap model handles extraction and the expensive one handles judgment. The telemetry shows which step used which model and what it cost.
- Crawl credits are metered per step. The crawler is the other variable cost; each profile declares its budget. The lightest profile, a seed pull that gates the client's onboarding invite, allows at most six external calls and targets five minutes. The full build allows the whole sweep.
- The cost is reported at the end of every run, next to the artifacts. Nobody has to go looking for it.
Measured across real client pulls, the full pipeline lands at roughly $4–8 and 0.8–1.4 million tokens per client. The lighter webhook-triggered research report, which runs when a lead enters the CRM, is budgeted at about eight page scrapes and four searches and finishes in under twenty minutes.
Why the number matters
Three reasons, in the order I discovered them:
- Pricing. A known unit cost can be marked up, bundled or passed through. An unknown one gets padded, and padding loses deals.
- Design. Once you can see cost per step, you find the expensive steps and ask whether they need the expensive model. Routing extraction to the small model and judgment to the large one came directly from reading the telemetry.
- Trust. When leadership asks "what is this costing us?", the answer is a number with a date on it. The whole 33-agent platform, serving the entire agency, operates for about $32–42 a month. Commercial alternatives in the same category run $49–249 per user per month.
Compute is a budget line
The same discipline applies when AI coding agents do the building. A multi-hour unattended build inside a CRM, at high compute settings, is real money. I run those in a fresh session per phase and use a small skill that warns when the context window grows larger than the task needs. Treating compute as a line item, rather than a surprise on the invoice, is part of the engineering.
What to ask before you buy an AI pipeline
- What does one run cost, measured on real inputs?
- Which steps use which model, and why?
- Is crawl or API usage capped per run?
- Where does the cost show up after each run?
If a vendor can't answer the first question with a number, the pipeline isn't finished.
Quick answers
How much does an AI research pipeline cost per run?
Measured on real clients, about $4–8 and 0.8–1.4 million tokens per full pull, with crawl credits metered per step. The lighter webhook-triggered report is budgeted at about eight page scrapes and four searches.
Why route AI work across different model tiers?
Because cost telemetry shows which steps need judgment and which only need extraction. Sending extraction to the small model and judgment to the large one came directly from reading per-step costs.
What does a multi-agent AI platform cost to operate?
The 33-agent platform serving an entire agency runs for about $32–42 a month in model and infrastructure spend, against $49–249 per user per month for commercial alternatives in the category.
Have a process AI should be running? Describe it and I'll tell you whether it's worth building. Book a call or ask the assistant.