Skip to content
George
ITENRO
Book a call
Book a call15–30 min
ITENRO

Journal / AI · Flows

When AI isn’t the answer: how I work it out in a Sprint

George18 September 20262 min read

GDM România: the site on a laptop
The GDM România social agent in the Telegram chat, on a phone

The request nearly always arrives in the same words: “can AI do this?”. Nearly always, yes. The useful question is a different one: which step actually improves, at what cost, under whose control.

To answer it I don’t open a slide deck. I open the real flow, with the real documents, and try one single step. That is the definition Sprint: a map of the flow, the essential specs, an estimate and a release plan.

Two numbers interest me: how much time that step costs today, and how much checking it would cost if a machine did it. If the second isn’t clearly smaller, the project doesn’t start.

Diagram: one step tried on real data leads to two outcomes, and AI only goes in when it improves thingsOne stepTried onreal dataIt improves: it goes in,with human approvalIt doesn’t: I say soand I stopDiagram: one step tried on real data leads to two outcomes, and AI only goes in when it improves thingsOne stepTried onreal dataIt improves: it goes in,with human approvalIt doesn’t: I say soand I stop
The Sprint in one picture: one step, two possible outcomes.

One step, on real data

Inside the Sprint I pick one concrete step and try it on the real documents, measuring before and after. No demos on made-up data.

At GDM România that step was running the social channels: writing, laying out, scheduling, checking again. Hours that appear on no invoice and vanish every week.

I counted the before on the hours the team reported in an ordinary week, and the after on the flow already in service. Same period, same channels, same posting rhythm.

Weekly time on the social channels compared: from 10–12 hours by hand to 15–20 minutes with the agentBy hand10–12 hours a weekWith the agent15–20 minutes a weekWeekly time on the social channels compared: from 10–12 hours by hand to 15–20 minutes with the agentBy hand10–12 hours a weekWith the agent15–20 minutes a week
Indicative estimate of the time spent running the social channels, not a guaranteed measurement: it depends on how many channels you keep and how often you post.
Sometimes the correct answer is that AI isn’t needed. When that happens I say so, and I don’t build.

It happens more often than you would think. If the blockage is a field filled in badly upstream, or a permission nobody granted, a model only adds one more layer to check. There you fix the flow, and that’s it.

Saying it costs me a job and saves months. That is why the Sprint is paid for on its own: it exists to decide, not to sell the next phase.

Always a person who approves

When AI does belong, it comes in with written limits. The social agent beside this line prepares everything on its own, but it only publishes after someone taps “Approve”.

The Telegram chat of the GDM România social agent: the proposed draft and the approval button

The agent lives on Telegram, where the team already is every day: no new tool to learn. It proposes the text and the image, gets a yes or a correction, and only then publishes.

I write the limits before I build: what it may do alone, what has to pass through a person, what it must never touch. Three lines in a document, worth more than any setting buried in a panel.

The result on a phone. Videos in my articles loop without sound, and stop whenever you want.

The rule that holds it all together fits in a few lines, and it’s the first thing I write:

# human approval rule
if confidence < threshold or touches(sensitive_data):
    hand_over_to(owner)   # no automatic decision
else:
    carry_on(flow)

A threshold, a limit, a person. When the system isn’t sure it doesn’t decide: it asks.