AI without the theater
How to tell an automation that genuinely takes work off someone's plate from a chat window nobody uses. Six questions worth asking before the first deployment.
The symptom is always the same
A company deploys "AI", someone posts a screenshot on LinkedIn, and three months later nobody on the team uses it. Not because the model is bad. Because a chat window on the site or the intranet doesn't solve any specific person's specific task — it's general, so it belongs to no one.
Automations that outlast a month share one trait: someone specific used to do a specific, boring thing by hand, and now they don't. That's narrow enough to measure and boring enough that nobody misses doing it themselves.
Six questions before the first deployment
-
1
Who does this task by hand today
If you can't name a person and their daily frustration, you don't have a project yet — you have an idea for a demo.
-
2
How long does it take today
Without a baseline number, you can't later show that anything improved. Measure before, not just after.
-
3
What happens if the model gets it wrong
Decides whether a person approves the output before it goes further, or the flow can run unsupervised.
-
4
Where does the input data come from
If it's spread across five systems today, building access to it is part of the project — usually the bigger part, not the model itself.
-
5
What will this cost on a monthly scale
Billing per model request grows with usage. Work it out on real volume, not the demo example.
-
6
What happens if it stops working
The vendor's model changes, prices change. If the whole process halts without this one piece, that's a risk, not a feature.
What usually works, and what usually doesn't
-
Works: classification and routing
A ticket, email or document reaches the right person without anyone manually reviewing each one. Easy to measure, hard to get wrong.
-
Doesn't work: a general-purpose assistant
A chatbot meant to answer "anything" about the company has no owner and no success metric — and usually no users either.
-
Works: a first draft of text
A reply, a description, a summary — a person edits a finished draft instead of starting from a blank page.
-
Doesn't work: unsupervised decisions where a mistake is costly
Automatic refund approval, automatic legal replies, automatic payment approval — these are places for an assistant, not an autopilot.
Questions
- Will this replace employees?
- A well-designed workflow removes one repetitive task, not a whole role. If someone has ten tasks today and an automation handles one, they're left with nine — not nothing.
- Which model should we choose?
- That depends on the task and the sensitivity of the data, not on which model is currently loudest. Simple classification runs cheaply on a small model; a complex summary needs a stronger one. You pick after measuring the task, not before.
- What does it cost to start?
- A first, well-scoped workflow is usually a few weeks of work. Ongoing cost depends on request volume — we give an estimate before starting, based on your real numbers.
Tell us what's most boring and repetitive at your company.
That's where we start — one workflow, a measurable result.
Write to us