Kurt FischmanFounder, Marshal
Kurt is the CEO of Marshal, the Managed Agent Operations company.
Six kinds of provider will sell you an AI agent, and only some of them are still responsible for it in month six. Sort the market by operating model rather than by software: who builds it, who watches it daily, who fixes it when it breaks, and who absorbs the cost of a mistake. Get those four answers in writing before you sign anything.
Picture the ordinary version of the problem. It's a Friday evening. An AI agent your business bought in March, the one that answers inbound calls and books appointments, has just booked three jobs into a slot that closed two weeks ago when you changed your service area. Nobody notices until Monday, when two customers call to ask where the technician is.
Now the useful question. Not "which agent should we have bought," but: who was supposed to catch that, and what do they owe you?
For most small businesses the honest answer is nobody, because nobody was ever assigned. The purchase was structured as a build, and the failure happened during the run. Those are two different commitments, sold by different kinds of company, and the gap between them is where most of this spending quietly goes to die.
An AI agent is a hire, not a purchase
An AI agent is software that takes a job end to end: it reads the incoming email, decides what the customer wants, checks the calendar, books the slot, sends the confirmation. That's the difference between an agent and the chatbot you already ignored. A chatbot answers. An agent acts, which means it can be wrong in ways that cost you money rather than just annoying somebody.
Anything that acts on your behalf needs supervision, correction, and maintenance, for the same reason a new hire does. Your prices change. A supplier renames a product category. Your calendar tool ships an update that breaks a connection. There's nothing exotic in any of it; it's just Tuesday. The agent doesn't notice any of it, because judgment isn't what it does. Somebody has to notice on its behalf, and that somebody is either on your payroll, on a contract, or imaginary.
So the real selection question isn't about capability. Every provider in this market will demo something impressive. The question is what I'd call run responsibility: the standing duty to watch the thing, catch the failure, and fix it. Six operating models divide up that duty differently.
The six operating models
The market presents itself as a list of products. It behaves like six operating models, which differ mainly in who is on the hook after launch.
| Operating model | What you get | Who watches it after launch | Sensible when |
|---|---|---|---|
| Software subscription | A self-serve platform with agent templates you configure yourself | You. The vendor covers its product, not your workflow logic | The job is simple, cheap to get wrong, and somebody in-house enjoys this work |
| Freelance automation builder | One or two narrow automations built quickly in a workflow tool | You, once the delivery window closes, unless you retain them separately | A contained task with a clear input, a clear output, and low cost of failure |
| Agency, build only | A defined project: custom workflows, integrations, and handover documentation | You, from the moment of handover | You have real operational capacity and want to own the asset outright |
| Agency with a support retainer | The same build plus a monthly window of hours for tuning and fixes | Shared, within the scope and hours the contract names | You want help on call but accept that exceptions reach your team first |
| Managed service provider | Monitoring, incident response, access management, and vendor coordination | The provider, for the systems the service agreement lists | You want a named party accountable for catching and fixing failures |
| In-house owner | An employee who builds and runs the system as part of their job | Your company, entirely | Agent work is becoming core to operations and you can fund the seat |
Read the third column first. It's the only one that describes what happens after the invoice is paid.
A software subscription and a managed service can look identical in a sales conversation and behave nothing alike in month six. The difference isn't the technology. It's the operating model, and there are six worth knowing.
Two of these get mistaken for each other constantly, so they're worth separating. A software subscription sold with the phrase "it runs itself" is describing the software, not your business process. The vendor is genuinely responsible for platform uptime and product defects, and genuinely not responsible for the fact that your intake form asks a question your agent misreads. Similarly, a support retainer is not the same thing as somebody watching. A retainer usually buys you a defined number of hours and a response window for problems you report. If nobody on your side is looking, an unreported problem is a problem nobody fixes, and you'll pay the retainer all quarter for the privilege.
The in-house option deserves more respect than it usually gets from people selling the alternatives. If agent work is becoming central to how you operate, an employee who owns it accumulates knowledge about your business that no outside party will match. It's also the most expensive and slowest option, and it fails badly if that person leaves.
The category is genuinely hard to shop
Some of your confusion is not your fault. Gartner has a name for a real practice in this market: "agent washing," the rebranding of existing products such as AI assistants, robotic process automation, and chatbots without substantial agentic capabilities. In its June 2025 assessment, Gartner estimated that only about 130 of the thousands of vendors claiming agentic AI are real. Treat that as one analyst firm's estimate rather than a census, but the direction is worth taking seriously: most of what's marketed to you as an agent is something older wearing a new label.
Gartner also forecasts that over 40% of agentic AI projects will be canceled by the end of 2027, attributing that to escalating costs, unclear business value, and inadequate risk controls. That's a prediction, not a measured outcome, and it was made looking mostly at larger organizations. Read it as a caution rather than a statistic about businesses your size. What makes it relevant is the shape of the three named causes: none of them is "the model wasn't smart enough." All three are management failures, and two of them are decided by how you structure the deal.
Five questions that reveal the operating model
Ask these before you get to price. The answers sort providers faster than any feature comparison, and the discomfort a question produces is itself informative.
- When the agent does the wrong thing on a Saturday, who finds out, and how? You're listening for a specific mechanism: an alert, a queue somebody reviews, a person with a name. "You'll see it in the dashboard" means you're the monitor.
- What's the first month after launch supposed to look like? A provider who has actually operated agents will describe corrections, edge cases, and tuning. One who describes only training and handover has sold builds, not runs.
- What's explicitly outside your scope? Good providers answer this quickly and specifically. Vagueness here shows up later as an invoice or an argument.
- Which decisions does the agent make alone, and which come to us for approval? You want the expensive and irreversible actions routed to a human on your side. Issuing a refund, promising a delivery date, and emailing your largest customer belong in different risk categories.
- If we end this in six months, what do we keep? Ask about your data, your workflow logic, and your integrations. Some answers are fine. No answer is not.
What it costs, and what the price hides
Price guides published by providers themselves through 2026 cluster in a recognizable range: roughly $2,500 to $7,500 one time for a single workflow, roughly $7,500 to $25,000 for a connected set of them, hourly rates for specialists somewhere between $75 and $300, and monthly support running from a few hundred dollars to around $5,000 depending on how much of the watching the provider takes on.
Use those as calibration, not as market rates. They're sellers describing their own pricing, which makes them useful for spotting a quote that's wildly out of family and useless as a benchmark. The more important point is structural: when a provider splits the number into a setup fee plus a monthly fee, the monthly fee is where the run responsibility lives, and its size tells you how much of it they've actually accepted. A build-only quote isn't cheaper than a managed one. It just moves an unpriced operating cost onto your team, where it will be paid in your attention instead of your budget.
Where we sit, and why we'd say this
Full disclosure, because you should discount this accordingly: Marshal is the Managed Agent Operations company that designs, deploys, and operates AI agents as a service for small businesses. We're the last row of that comparison. We benefit if you decide the run matters more than the build, so weigh the argument on its merits rather than on our say-so.
The merits, briefly. Every operating model here is the right answer for somebody. If the job is contained and cheap to get wrong, a subscription and an afternoon of your own configuration is a perfectly good decision, and paying anyone a retainer for it is waste. If the job touches revenue, customers, or a promise you have to keep, the question stops being which software and becomes which party is accountable when it fails. Answer that one deliberately, in the contract, with a name attached.
The pattern worth avoiding is the one that catches most founders: buying a build because it's a legible one-time number, then discovering you've hired an operating obligation you never staffed. You can decide to carry the run yourself. Just decide it on purpose, before Monday morning brings the phone call about the technician who never came.