Miguel Marengo Canales

blog/Silodisa AI log

weekly log · issue 7 · six buttons without AI

26,621 pallets for US$0: six questions AI doesn’t need to answer

Six buttons above the chat answer, with no model and for US$0, the questions that always get the same answer. What does use AI has a cap: US$2.00 per person per day.

Miguel Marengo Canales
 |5 min |comments
a question in the chat question fixed? yes no button · US$0same answer model · ≈US$0.04cap US$2.00/day repeats + fixed rule = button
The path of a questionIllustrative. If the answer comes from a fixed rule, it goes to a button; if it changes every time, it goes to the model, with a cap.
in 30 seconds
  • If a question repeats and its answer comes from a fixed rule, it doesn’t need a model: it needs a button.
  • Six buttons above the Planner’s chat answer with no AI, for US$0. With September 30 data they read 26,621 pallets.
  • What does go through the model is capped at US$2.00 per person per day, about 50 questions, checked before the model is called.
26,621pallets read by one button, US$0
6buttons with no model
15 / 15matches with the independent auditor
US$2.00AI cap per person per day

When you put an AI chat in a warehouse, everyone asks it the same things. How many open locations are there? Which lots are about to expire? Which fast-moving pallets are far from the door? Each question costs the model a few cents, takes a few seconds, and may come back worded differently the second time.

So we did something that sounds odd in an AI project: we took questions away from the AI.

The rule

If a question repeats and its answer comes from a fixed rule, it doesn’t need a model: it needs a button. If the question changes every time (“plan ten class A pallets, but not to level 4”), then yes, let the model read it.

Above the Planner’s chat we put six buttons: Warehouse status, Reclassify ABCD, Misplaced A’s, Free up low levels, Verify plans and Lots about to expire. Each one runs a hand-written calculation, with no model, and its answer carries a footer: “No AI · US$0.” It also carries the signature of the agent accountable for it, like everything else in the app.

six buttons · Sep 30 status by zone, levelUS$0 reclassify 0 of 133 changeUS$0 misplaced A’s 4,703 of 13,970US$0 low levels 953 C/D on L2US$0 verify plans done or not?US$0 expiring 232 in 90 daysUS$0 0 model calls · No AI · US$0
Six buttons, zero modelWhat each one answered with real data from September 30, 2026. Button layout is illustrative.

What they answer

With data from September 30:

  • Lots about to expire read 26,621 pallets across all zones: 232 expire in the next 90 days, 1,047 have already expired and must come out (still in the red), 3 have a doubtful date, 3 have no date, and 25,336 are outside the window. The five figures add up exactly to what was read; if they didn’t, the watchdog flags it.
  • Misplaced A’s: 4,703 of 13,970 class A pallets sit in the rack outside levels 2 and 3, where they belong. They are just 10 SKUs.
  • Free up low levels: 953 class C and D pallets occupy the front of level 2, the handiest spot. Those are the possible gaps.
  • Reclassify ABCD: today, 0 of 133 SKUs would change class. “Nothing to do” is also an answer, and it cost nothing.
each pallet, one box only no date3 doubtful date3 expired1,047 expiring · 90 days232 outside window25,336 sum = pallets read26,621 if they don’t add up, the watchdog flags it
Lots about to expire: five boxes that add upReal data from September 30, 2026. Each pallet falls in exactly one box, checked in this order; the total must equal the pallets read.

The first version of “Free up low levels” also counted the back of level 2 and level 3 as “A’s waiting,” and came out at 11,529 pallets against 953 gaps. A number that said nothing. Now it’s exactly the same 4,703 misplaced A’s.

Why not leave it all to the model

  1. It costs zero. The tests of the six buttons made 0 calls to the model.
  2. It always answers the same. An independent auditor that read only the design calculated the same thing on its own: 15 matches, 0 differences. With a model, the same question may come back in different words.
  3. It leaves the model for the hard part. What’s worth asking it is what changes: a plan with conditions, a question that doesn’t fit in a button.

After each answer, the app suggests up to three next questions, also from fixed rules. If you just looked at “Misplaced A’s,” it offers “Plan moving the A’s to levels 2–3.” That plan doesn’t use a model either.

“Nothing to do” is also an answer, and it cost nothing.

And what does use AI, with a cap

For what goes through the model, each person has a cap of US$2.00 a day, enough for about 50 questions of roughly US$0.04. It’s checked before calling the model: when you hit it, the app says so and sends nothing. It applies to everyone, admins included. The bar at the top of every screen shows how much you’ve spent today.

one person, one day US$0US$2.00 ≈ 50 questions of ≈ US$0.04 cap: nothingis sent checked BEFOREcalling the model applies to all, admins too
A cap checked before, not afterIllustrative: the bar’s progress. The US$2.00 cap and the roughly US$0.04 per question are the app’s; the cap is marked as an assumption.

Testing costs too. Every time we change the Planner we run a battery of 54 questions; one full run cost US$2.47. That’s the price of knowing we didn’t break anything.

What we don’t know

  • How many chat questions moved to the buttons. Every button press is logged; we still need a month of use to count it.
  • Whether US$2.00 is the right cap. It’s marked as an assumption.
  • Whether 90 days is the right window for “about to expire.” Also an assumption, adjustable in Parameters.

What comes next

AI isn’t the chat. It’s knowing which question deserves a model and which deserves a button. BIEN y a Tiempo (done right and on time) also means not paying twice for the same answer.

How many of the questions you ask your system always get the same answer? Write to me.

Miguel Marengo Canales · Silodisa

weekly loglogisticsPlannerAI cost

Comments

comments · I answer every one. If you disagree, even better: say it with numbers.

  • loading…