AI Agent Pricing

A work-unit is one completed workflow. When one of your Marshal agents takes a defined job and carries it through to completion, that represents a single work-unit. Routing a qualified lead, booking a meeting, resolving a ticket, producing a brief: each finished workflow is one unit. You are billed for work that gets done, not for seats, software licenses, hours, or raw usage like tokens and prompts. Every completed workflow counts the same. There is no weighted schedule to decode, so you can always see exactly what your agents delivered and what you paid for. Each plan includes a monthly minimum, and you can set alerts and caps to keep spend predictable.

Each agent system has a monthly minimum of 35 completed work-units at $15 per unit. If your agent completes more work than the minimum, additional completed work-units are billed at the same unit rate unless you set a cap. You can set alerts and limits so the system does not quietly run up spend in the background.

Yes. You can set budget alerts, monthly caps, and approval rules. Some workflows can run automatically. Others can require human approval before the agent sends, updates, books, publishes, or escalates anything. We define those gates during scoping.

A work-unit counts when the agent carries the defined workflow to its agreed endpoint. If the workflow errors, stalls, or gets routed to an exception queue before completion, it does not count as completed work. Every run is logged so you can see what happened, what was completed, and where human review was used.

No. After we scope a workflow, Marshal builds, tests, and deploys the agent inside your environment on our dime. You get to observe it executing your work in real time. This proof step exists so you can evaluate agentic AI the way it should be evaluated: by watching it perform, not by reading a deck.

You walk. No subscription. No "but you owe us for the build" conversation. We carry the build cost because we are confident in what we ship, and we accept the consequences when we are wrong.

Start with a workflow that happens often, follows a repeatable pattern, and already costs the business time, speed, or revenue. Good first workflows include inbound lead follow-up, customer support triage, CRM cleanup, meeting prep, sales research, intake routing, reporting, and operational handoffs. Bad first workflows are vague, rare, politically sensitive, or dependent on judgment nobody can explain.

You need to help us define what “done correctly” means. That almost always means that we'll need access to the tools involved, examples of the work being done today, the rules your team follows, and one internal owner who can make decisions quickly.

The agent does not guess its way through judgment calls. When a workflow falls outside the rules, it moves to an exception queue or routes to a human for review. The goal is to let machines handle repeatable work while humans stay in control of judgment, edge cases, and customer-sensitive decisions.

Yes. We build on the tools you already own. Standard integrations cover the common revenue, communication, and operations platforms (HubSpot, Gmail, Slack, Calendar, Notion, Salesforce, Apollo, Linear, Stripe, Intercom, Airtable, and the usual suspects). Advanced integrations covering deeper API work, custom middleware, or legacy systems are scoped. No rip and replace, no forced migration.

Your core systems stay where they are. Your CRM, inbox, calendar, and project tools never move. Agents reach into those systems through credentials you control and can revoke at any time. The workflow logic, the agent runtime, and the orchestration layer run inside a secure, hosted environment with encryption in transit and at rest, and full audit logging on every run. Your data is not used to train models, ours or anyone else's. You retain ownership of both your data and the outputs generated by the agent system.

No. It means the work output of the agent systems in Operator is roughly equivalent to what a full-time employee would produce on the same set of repeatable tasks. The point is leverage, not headcount cuts. You still need humans for the judgment calls, the relationships, and the exceptions. The agents handle the volume so your people can focus on higher-order tasks.

Marshal is capable of launching within days. Things move quickly when system access and rules are clear. Messy data, multiple teams, or complex approvals will naturally take longer. We scope all of this before the build.

Marshal is not a fit for businesses where recurring knowledge work doesn't impose a meaningful cost. If human time spent on follow-up, routing, reporting, research, CRM updates, intake, customer replies, or operational handoffs is a minor cost, agentic AI is probably a solution to a problem you don't have.

Yes. We bill monthly and deliver services over recurring 3-month sprints. You are free to cancel at any time and we'll pause your account at the end of that current sprint.