An AI assistant for a small team
Running assistants for a handful of people without per-seat pricing: what to centralise, what to keep separate, and what it costs.
Per-seat AI pricing scales badly for small teams. Ten people at $20–30/month is $200–300 before anyone has done anything unusual, and it buys a shared product you cannot inspect.
Self-hosting changes the shape of the cost, and adds work you should be honest about.
One shared box, or one each
One shared instance is cheaper and simpler, but everyone’s context lives in one place. Fine for a team with a common workload; wrong the moment someone handles HR, legal or personal data.
One per person keeps context separate and lets each pick a profile. It costs more hardware and, without fleet management, more administration — which is precisely the problem that makes appliance operating systems awkward past one box.
The middle path most teams land on: shared instances per function (support, engineering, ops), not per person.
What it actually costs
Hardware, one-off: a mini-PC at $150–400 handles several users. Cloud, if you would rather not own hardware: roughly €4–17/month per instance.
Model usage is the variable, and you pay providers directly at token rates rather than per seat. For most teams this is dramatically less than per-seat pricing — the crossover is usually a handful of users.
The cost people forget is time. Someone has to own updates, backups and the occasional failure. An appliance reduces that to near zero, but not to zero.
Fleet management is the difference
With more than two devices this stops being optional. What you need:
- One view of every device, with live status
- Updates pushed to a selection, not machine by machine
- Per-device cost visibility, so one runaway workload is obvious
- An audit trail of who ran what
deputyOS links devices to an account and gives you exactly that, with each device holding an outbound tunnel so nothing needs port forwarding.
Compliance
If you are in a regulated sector, self-hosting resolves the data-residency question that otherwise blocks procurement. Audit batches, CVE tracking enriched from OSV/NVD/CISA KEV/EPSS, and compliance evidence export exist for that conversation.
An SBOM is increasingly a procurement requirement in its own right, and it falls out of the build rather than being assembled later.
Where this is the wrong answer
If your team needs frontier capability on every task, wants zero operational burden, and has no data-residency constraint, a hosted product is the better buy. Self-hosting wins on cost, control and auditability — not on raw capability.
businessfleetuse-case