Demos aren't desks
An agent that dazzles in a slide deck still has to survive your exceptions, approvals, and Tuesday afternoon volume.
Most AI buying starts with a demo. The model answers a clean question. The room nods. Someone asks about timeline. Nobody asks who owns the failure when the agent is wrong with confidence.
Desks are different. They have tone rules, refunds that need a human, CRM fields that lie, and three tools that disagree about the customer. An agent that cannot escalate cleanly is not autonomy. It is a new ticket source.
What the demo hides
Demos are trained on the happy path. Real work is the exception path: the customer who already complained twice, the invoice that almost matches, the lead that looks qualified until you open the notes. If the vendor cannot talk about those without changing the subject to model names, you are still in demo land.
A desk also has volume. One clever reply is not a queue. Ask what happens at 4pm when the same eight questions arrive at once, and what happens when the approved answer is missing.
Demand a map of the job
Before you buy, ask for a map — not a feature list. The map is boring on purpose:
- What comes in, and from where.
- What the system is allowed to read.
- What it is allowed to write back — and to which tool.
- When a person has to take over, and what they get besides a blank ticket.
- How you will know the work actually got done.
If that map cannot be drawn on one page, the agent is not ready for a desk. Customer support and sales leads are two versions of that map we already write down.
Escalation is the product
The interesting part of an agent is not the reply. It is the brake. Who can approve a refund. What must never be sent. What gets logged. People stay in the work for judgement; the system should arrive with the history attached.
What we build instead
Spirality builds agents for desks. Context first, brakes on actions, and someone still answering when the system misbehaves. Impressive is optional. Owned is not. If aftercare is part of the job, say so early — that is a depth decision, not a go-live surprise.
Keep reading.
Start with the work. Not the AI.
The useful first question is not which model to buy. It is which process still depends on memory, retyping, and whoever happens to be at the desk.
Read the note
The aftercare question
Consulting, build, or build-and-run — the wrong depth wastes more money than the wrong model.
Read the note
Organisational memory beats another model
Swapping models is easy. Teaching the organisation to remember how it works is the hard part.
Read the note