Silodisa AI · weekly log · installment 1
Every extra pallet move costs us $4.75 MXN. And we didn't know it.
$4.75 MXN: what it costs to make one extra pallet move. How we're turning a warehouse that “works” into a warehouse that measures, explains and leaves the final word to the people on the floor.
- If a forklift with its driver costs $95 MXN an hour, moving a pallet that takes three minutes costs $4.75.
- Thirteen AI agents, each with a human owner and an “I disagree” button. Artificial intelligence recommends; people decide.
- A dashboard that never comes out green if a number is red or an alarm is on; whatever can't be measured with confidence stays grey.
A few weeks ago we asked ourselves an uncomfortable question: how much does each extra pallet move cost us, in pesos and in minutes? We didn't know. We knew the warehouse worked, that the trucks went out and that inventory reconciled almost always. But “almost always” and “it works” aren't numbers, and without numbers there's no way to improve.
That question gave birth to Silodisa AI, the intelligence layer we're building on top of our warehouse system. It isn't a new WMS or a pretty dashboard. It's a group of artificial intelligence agents that read what happens on the floor, calculate it with real physics and tell people what's best to do. Our motto sums it up: "BIEN y a Tiempo" (right and on time).
Measure first, give opinions later
The first thing we did wasn't programming. It was capturing the warehouse as it is: the meters of every aisle, rack heights, where there's a tunnel and where there's a closed bridge, how long a Double Reach takes to lift to a fourth level and how long a counterbalance takes to carry a pallet to stage. Every data point carries its source (the system, an assumption or something measured on the floor), and no one uses it in a calculation until an administrator approves it.
With that, the system converts any move into three things: meters, seconds and pesos. The same rule for every move: the actual distance on the floor, the time of the equipment that covers it, and what that time costs at the rate confirmed by an administrator.
If a forklift with its driver costs $95 MXN an hour, moving a pallet that takes three minutes costs $4.75. It sounds like little, until you multiply it by the thousands of pallets moved each month.
We had a closed bridge between two aisles and intuition said “we have to open it.” We ran the numbers.
The detour cost 63.8 meters for every lot that crossed there, but only a handful of lots actually crossed, and opening it meant removing rack. The decision was not to open it, with the math in hand. That's what we're after: warehouse decisions made with data, not hunches.
Who owns each minute of the truck
A truck can spend four hours in our yard while the warehouse is responsible for only forty minutes. We used to argue about it; now we measure it. The Cronos agent splits each truck's time into five parts and grades the warehouse only on the part that actually depends on it.
The five parts add up to the total to the minute. Only the time at the door grades the warehouse; the rest belongs to the carrier or to the appointment.
Agents with a human owner
Today we have thirteen agents working, each with a name and a concrete task. Some of them:
- Gatekeeper · recommends. Suggests the inbound or outbound door that generates the fewest meters, instead of the “usual” one.
- Put-away · organizes. Proposes aisle and level for each pallet according to its ABCD class, calculated from 80 days of real outbound shipments.
- Planner · organizes. Understands questions written the way people talk and builds re-slotting plans with their savings per day and how many days they take to pay back.
- Guardian · evaluates. Watches over safety: shifts with a faulty forklift, rack above 95%, expired product in rack.
- Counter · evaluates. Looks after inventory accuracy with cycle counts and a plan per shift.
- Verifier · audits. Checks every day whether what the agents suggested matched what was actually done.
And here's what matters most to me: every agent has a human owner. Their name appears on every suggestion, every alarm and every plan. Next to the signature there's a button that says “I disagree.” When someone presses it, they write one line with the reason, and the agent's owner has to respond: either explain it, or accept that the agent needs improving. Agents don't adjust themselves; their owner proposes the adjustment. Artificial intelligence recommends; people decide.
Nothing moves on its own
A rule we don't negotiate
No agent writes to the warehouse system on its own. The Planner can build a plan of fifty moves, but for them to become real tasks, someone with authority has to review it, approve it and send it.
Before a task reaches the forklift operator, the system checks again that the pallet is still where the plan saw it; if it isn't, it says so and asks for confirmation. Every dispatch, every approval, every change of rate or parameter is recorded in a log with who, when and what the previous value was. If something goes wrong, we know why.
An honest dashboard
The warehouse is graded on six numbers: inbound and outbound trucks served on time, pallets put away in under 3.5 hours from coming off the truck, inventory accuracy, complete orders and on-time shipments. Each one has a green and a red that an administrator sets in a single place, and the safety, expiration and space alarms are always visible.
Two dashboard rules I really like: the grade never comes out green if a number is red or an alarm is on, and if the system can't measure something with confidence, that number stays grey and says so. We'd rather have a dashboard that admits what it doesn't know than one that looks good.
How we're building it
We work in short blocks of one or two days, and every block closes the same way: test cases written before the code, an automated walkthrough of every screen in three roles, on laptop and on phone, and an independent auditor that recalculates the numbers on its own and has to match what the app shows one by one. The Dashboard, for example, was validated with hundreds of matching figures before anyone was shown it. If the auditor and the app don't match, it doesn't deploy.
Everything runs in the cloud, with secrets in a single place, and the agents automatically choose the cheapest AI model that answers each question well. AI spend is shown at the top of every screen, in pesos, every day.
We're now taking the same approach outside the warehouse's four walls: fleet, operators, in-transit cold chain, deliveries on a service-level clock and replenishment at the point of consumption. Same discipline: measure first, agents with owners, nothing moves on its own.
What we're after
- Time. Fewer meters per pallet, trucks served in 45 minutes, pallets in place in under 3.5 hours.
- Money. Every plan states how much it saves per day and how many days it takes to pay back.
- Accidents. A faulty forklift shouldn't work a full shift without anyone finding out. Now there's an alarm that says so.
- Efficiency. Accurate inventory and pallets where the ABCD class says they should be.
We're not done. Several of our numbers are still red, and that's precisely the reason for all of this: now we see them. A smart warehouse isn't one that's never wrong. It's one that knows how to measure, explains why it recommends something and leaves the final word in the hands of the people who know the floor.
Every week I'll share here a piece of what we're building with AI inside Silodisa and why it matters. If you're interested in how to apply AI agents to a real logistics operation, write to me. I'm glad to share what we've learned.
Comments
comments · I answer every one. If you disagree, even better: say it with numbers.