A distribution center on 42 hours: picking, inventory and zero shrinkage with AI agents

In April 2026 the maximum ordinary workweek in Chile drops to 42 hours (see our full analysis of Law 21,561) and in 2028 it reaches 40. For a retail or e-commerce distribution center —where every minute of human time counts double during the peak-hour pick— this reduction is not a payroll adjustment: it is a problem of operational architecture. The operations managers we work with raised the same dilemma: "I don’t want to hire 20 more people, but I also don’t want to lose SLA or have shrinkage go up". This article proposes a concrete way out.

1. What exactly you lose when 2 hours are cut

A full-time worker on a 5x2 regime goes from 44 to 42 hours. Two hours per week. It seems like little until you multiply by the DC’s 180 people: 360 fewer person-hours per week, 1,440 per month, 17,280 per year. In a DC whose productivity per hour is between 120 and 180 lines picked, that is between 2 and 3 million fewer lines per year with everything else equal.

The instinctive response is to hire more people. Mathematically it works (you need 4.5% more workforce) but it creates four new problems:

• Learning curve: a new picker takes 6-8 weeks to reach average productivity.
• Turnover: when you add low-skilled staff, picking errors rise 15-25% in the first 60 days.
• Fixed cost: the extra workforce stays all year, but the peak demand is in 3-4 specific weeks.
• Compliance: more heads means more shifts to rotate, more risk of exceeding the weekly cap through an administrative error.

The alternative is not "to work faster" (that is no longer possible), but to eliminate unnecessary work and better distribute what remains. That is what an architecture of agents does.

2. The 6 critical tasks of a DC and the cost of each one

In a typical DC, time is consumed roughly like this:

Cycle counting is the first to slip when there is time pressure and, ironically, it is the one that sustains the health of the inventory. Ignoring it produces the typical effect of a stressed DC: everything seems to be going fine until the annual inventory shows adjustments of 1.8% of the total value and no one can explain where they came from.

3. Why shrinkage rises when the workweek is squeezed

Shrinkage in a DC does not come only from theft (although it exists). In order of contribution to the typical total:

• Picking errors: the wrong SKU is picked, less quantity is shipped, the customer requests a reversal. ~35% of DC shrinkage.
• Internal breakage and damage: poorly built pallets, fallen boxes, products burst in internal transit. ~22%.
• Expirations: FIFO not respected in hot zones, product that goes from "sell at a discount" to "throw away". ~18%.
• Administrative differences: inventory adjustments with no clear explanation, receipts not reconciled. ~15%.
• Shrinkage (theft/loss): ~10%.

When the workweek drops and the pressure rises, supervisors prioritize picking and dispatch because those are the ones the customer sees. Counts are postponed, damage inspection is skipped, receipts are accepted without reconciliation. The first four types of shrinkage grow in silence. In a medium-sized DC, 0.3 percentage points of additional shrinkage on CLP 18,000 M of inventory are CLP 54 M per year. That pays for several agents and then some.

4. Proposed architecture: 5 agents in parallel

The proposal is not a magic agent: it is five specialized agents that coordinate via the same data layer (WMS + attendance system + vision). Each one attacks a specific pain and together they recover the lost hours without adding headcount.

1WorkforceAgent — a living shift schedule

It calculates optimal shifts by day/hour considering historical order demand, skill profiles (pickers, replenishers, forklift drivers, load-certified staff), the new 42-hour cap, parental care and the collective agreement. It redistributes the workforce toward the real peak hours, not the "assumed" ones. It typically frees up 3-5% of productive hours without touching headcount. Full detail in the 40-hour Law article.

2Internal RouteAgent — optimal picking routes

It is not the truck routing out on the street: it is the routing inside the DC. Each picking order becomes an optimal sequence of aisles and positions. It uses real history of time per position, SKU density and cross-docking constraints. Real implementations reduce 8-14% of the time per line in large DCs. The roughly 2 hours lost per week are recovered with this agent alone.

3CountingAgent — vision-based cycle counting

Fixed cameras in the highest-turnover zones (ABC) observe every movement and compare against the WMS. They flag discrepancies in minutes, without waiting for the annual count. A worker confirms or dismisses the alert from their device. It reduces the time spent on physical counts by 60-80% and detects shrinkage while the cause can still be investigated. In high-value retail (electronics, perfumery) the payback is 2-4 months.

4QualityAgent — vision at receiving and dispatch

At the docks, cameras inspect each pallet on the way in (damage, quantity, match with the delivery note) and on the way out (breakage, labeling, correct order). It detects 95% of the cases that the operator misses today due to time pressure. It avoids accepting damaged merchandise (supplier discount) and avoids dispatching pallets with problems (customer complaints). It reduces breakage and errors by 20-35%.

5DemandAgent — prediction-based pre-replenishment

It predicts 48-72h in advance which SKUs will be picked the most. It replenishes the hot positions during off-peak hours (before 8:00 a.m. or after 10:00 p.m.), so that during the peak the pickers don’t stop for lack of stock. It reduces pickers’ idle time and eliminates rushed replenishment in the middle of the peak (the #1 cause of pallet breakage).

Optionally, ProcessAgent is added to automate invoices, delivery notes and electronic documents (SII), freeing administrative staff who return to operational tasks.

5. Case: a 15,000 m² DC with 180 workers

A B2C logistics operator in the Santiago Metropolitan Region. Consumer goods retail (non-perishable food, cleaning, personal care), ~45,000 active SKUs, 180 direct workers across two shifts, 22,000 weekly orders on average with a peak of 38,000 during campaigns (Cyber, Christmas).

Aggregate net annual savings: ~CLP 340 M (shrinkage + overtime + adjustments + added capacity without hiring). Total stack investment: ~CLP 160 M (year 1) including vision hardware. Effective payback: 5.6 months.

6. A 4-month implementation plan

The most common mistake is wanting all the agents operational at the same time. It fails: the DC doesn’t have the bandwidth for five parallel implementations. We recommend a sequence:

1. Month 1 — Diagnosis + WorkforceAgent. We model demand, the agreement and costs. The new 42h shift schedule is delivered. The DC recovers the lost hours almost immediately.
2. Month 2 — Internal RouteAgent. Integration with the WMS, we pilot in 2 critical aisles, validate the gain, full rollout. Picking gains 10-14% in speed.
3. Month 3 — CountingAgent + QualityAgent. Cameras are installed in the ABC zones and at the receiving/dispatch docks. The agents are trained with 2-3 weeks of video; in parallel the counting team is redirected to confirming alerts.
4. Month 4 — DemandAgent + ProcessAgent. The stack is completed. The administrative team’s operational load drops ~30%. The DC closes the month with high SLA, minimal overtime and 100% compliance.

The human always stays in the loop: each agent proposes, the supervisor approves when appropriate. No one is delegating critical decisions to an AI in blind mode.

7. Frequently asked questions