Something changed in how AI agents get sold to small businesses this year. The pitch is no longer chat with your data. It is log into your invoicing software, process the backlog, and send the reminders. That shift is dragging the entire pricing model for AI deployments away from subscriptions and toward a fee per agent installed. We think it is the right model, and we are already pricing our own builds this way.
What changed: agents that use your software instead of talking about it
For two years, most AI tools sold to small business were chat wrappers. You typed a question, a model answered, and a person still had to open the actual software and do the work. Computer-use agents close that gap. Anthropic shipped this capability directly into Claude, letting a model see a screen, move a cursor, click buttons, and type into whatever application is already open, rather than needing a custom integration built for every tool (Anthropic's computer use announcement, with the technical detail in the computer use documentation). OpenAI shipped a comparable capability under the name Operator, aimed at the same problem: an agent that operates existing web software the way a person would (OpenAI's Operator announcement).
That matters more for operators than it sounds. Most 10-50 person companies do not have a clean API layer connecting their invoicing tool, their CRM, and their scheduling software. They have a person who logs into three different logins and copies data between them by hand. A computer-use agent does not need an integration built specifically for it. It needs a login, a task, and supervision. That is why the work being automated right now is unglamorous and specific: invoice processing, appointment confirmations, review requests, data entry between systems that were never meant to talk to each other. This is exactly the category of work our automation builds target first, because it is where the admin leak tends to be largest and the fastest to fix.
The pricing model is flipping from seats to agents
Subscription software prices per seat because the cost structure is the same whether one person uses it or a hundred: you are paying for access to a tool. Agent deployments do not have that shape. Each agent is scoped to one workflow, requires setup specific to that client's software stack, and needs ongoing supervision as the underlying tools change. The pricing showing up in the market reflects that: a flat fee to build and install a given agent, plus a smaller recurring fee to monitor, tune, and coach the business owner on how to use it well. Not a per-seat license. A per-agent build.
We run our own AI Concierge work the same way. The assessment costs a flat €999, credited in full against the build if the client moves forward, because we would rather get paid for an outcome than for access to a dashboard. Once a system is live, our target for the first working piece is days to weeks, not a quarter-long rollout, because an agent only needs to be scoped to one workflow to start proving itself. The token costs behind these agents also follow a usage-based model rather than a seat license, the same logic laid out on Anthropic's pricing page: the cost driver is compute per task, not headcount, which is another reason billing per agent makes more sense than billing per user.
| Model | What you pay for | What breaks it |
|---|---|---|
| Per-seat SaaS | Access to a tool | Nobody has time to use the tool well |
| Per-agent build | A finished workflow, running unattended | The underlying software changes and the agent needs retuning |
| Monthly management retainer | Supervision, coaching, and improvement of a live agent | The business owner never gets involved and the agent drifts |
The failure mode of the per-agent model is the same one we flag in every case study we publish: an agent that runs 24/7 still needs an owner who checks it, not because the model is unreliable, but because the business changes underneath it. A price list changes, a new intake question gets added, a vendor gets swapped. The retainer exists to catch that drift before a client finds out the hard way.
One agent managing a team of agents
The more interesting pattern building on top of computer-use agents right now is delegation between agents. Instead of one agent doing one task, an orchestrator agent holds the business context and hands specific jobs to narrower specialist agents, the way a manager assigns work without doing all of it personally. This is the same architecture behind our AI Chief of Staff: one agent that understands the whole operation, backed by a Second Brain of company context, delegating to narrower agents for the repetitive execution work.
The practical benefit is that you do not have to teach every agent everything about the business. The orchestrator carries the context, who the client is, what the pricing rules are, what counts as urgent for this company, and the specialist agents just execute. That is a cleaner mental model for operators than juggling a dozen separate AI tools, and it is closer to how you would actually structure a team.
Vertical agents beat general-purpose platforms
There is a real argument in the market right now for general-purpose agent platforms: build once, point it at anything. We think that argument is weak for most operators of 10-50 person companies, for the same reason a generic ops manager is worse than one who has run your specific type of business before. A vertical agent built for real estate lead intake knows that a lead who does not hear back within five minutes is a lead who is already talking to a competitor. A general-purpose platform does not know your five-minute rule exists unless someone configures it, and most businesses never get around to configuring it well.
We build this way because we run a 40-person real estate group ourselves, so our voice agents and follow-up systems are tuned to our own five-minute speed-to-lead threshold before we ever hand them to a client. That is the difference between a platform that could theoretically do the job and a system that already knows what the job is.
What to build this quarter
If you are running a 10-50 person company and watching this shift from the outside, the practical move is not to wait for a bigger platform to mature. It is to pick the one workflow bleeding the most hours, usually somewhere around 10 or more hours a week of someone's time on data entry, reminders, or reconciliation, and scope a single agent to it.
- Audit which repetitive tasks require logging into software you already own, not tasks that need a brand-new tool.
- Price the build as a fixed fee per agent, not a monthly seat license, and negotiate a smaller retainer for ongoing supervision.
- Start with one orchestrator holding your business context, and add specialist agents underneath it as workflows get automated.
- Resist buying a general-purpose platform before you have named the specific vertical rule, like a speed-to-lead threshold, the agent needs to enforce.
None of this requires ripping out your existing software stack. Clients who build with us keep ownership of 100% of the code and files, specifically so a computer-use agent sitting on top of your existing tools does not become another vendor lock-in problem. If you want a sense of where your own admin leak is worst before committing to a build, our revenue leak heatmap is a faster first step than a full assessment. We cover how this keeps developing on the blog.