Case study
Autonomous front desk
The email channel of a ground transport marketplace, run end to end on Claude at a fixed monthly cost.
Agents work the inboxoutreach, support, bookings
- Sector
- Ground transport marketplace
- Year
- 2026
- Role
- One engineer
- Stack
- Claude, React, Node.js, SQLite
- Runs on
- Terraform, AWS, Cloudflare
The problem
A marketplace inbox carries three different jobs at once. New suppliers have to be contacted, followed up and set up. Customers ask questions that need answering the same day. And a proportion of booking requests arrive as prose in an email rather than through the booking form. All three are repetitive enough to want automating and too varied to script, and doing them with metered LLM APIs prices the whole approach out before it is useful.
What we built
A set of agents that work the email channel end to end. They run supplier outreach as a pipeline, and once a supplier agrees the details go into the mobility platform so the supplier can log in and trade. They answer customer support on the same channel. And they read booking requests that arrive as email and turn them into bookings on the platform. The agents run on Claude against a fixed-fee subscription rather than metered API calls, so cost is a known monthly number and scaling means adding another subscription rather than watching a usage bill. Orchestration runs on Agentic World, a control plane built for this: a Node.js daemon holding the shared work queue and event log in SQLite, with a React console over it.
The long version
The economics are the point of this one. Agent workflows are easy to justify on a slide and hard to justify on an invoice, because metered API pricing scales with exactly the thing you want more of. Building against fixed-fee subscriptions instead turns cost into a number that can be budgeted, and turns scaling into a decision about how many subscriptions to run rather than a bill nobody can forecast.
The work suits agents better than most: a long tail of similar-but-not-identical conversations, each one low stakes on its own and collectively the constraint on how fast a marketplace can grow. Supplier outreach is the clearest case, because the same conversation happens hundreds of times with different operators. Customer support is the same shape from the other direction. And booking requests that arrive as email are the case where an agent earns its place most obviously, because the alternative is a person reading prose and retyping it into a form.
What makes it hold together is that the handover is into a real system rather than a spreadsheet. A supplier who agrees gets an account on the mobility platform. A booking request becomes a booking. The agent’s job finishes exactly where the platform’s begins, which is also the line past which a mistake would be expensive.