Agentic AI in the Warehouse: Why Labor Contracts Are the Real Constraint

Two things are happening in the same warehouse right now, and almost nobody is looking at both.

The first is the one that gets written about. Warehouse management software is shipping agents — software that doesn’t just recommend an action but takes it. Manhattan Associates told investors on its most recent quarterly call that more than 10% of its Active install base has now touched its agent product through a pilot or subscription, with every pilot so far converting to a subscription. Blue Yonder has released warehouse, logistics, and inventory operations agents, and a system for testing agent behavior before it runs live. The capability is real, and it is arriving faster than most distribution operators expected.

The second thing is a document being signed in the same buildings. Across food distribution and grocery warehousing in 2026, a wave of collective bargaining agreements has been ratified — and a growing number of them contain explicit protections against automation. Not aspirational language. Contract language.

The agentic AI warehouse conversation almost never includes that second document. It should, because in the physical economy that document is usually the binding constraint — not model capability, not integration difficulty, not budget.

The gap between “it works” and “it changes anything”

I’ve sat on the buying side of warehouse and labor management system selections, with the capital authority to approve them and the operating P&L that absorbed the result. The lesson that transfers directly to agentic AI is this: a system working is not a system deployed.

Between those two states sits a stretch that vendors rarely model and buyers consistently underestimate. Engineered labor standards have to be re-timed. Supervisors have to be retrained to trust an assignment they didn’t make. Productivity typically dips before it climbs, because the workforce is learning a new pattern while the standard is still being validated. That dip is normal, survivable, and — critically — it happens after you’ve already cleared the harder gate.

The harder gate is permission.

What automation language actually does

Automation protections in a distribution contract generally do some combination of the following: require notice before introducing technology that changes bargaining-unit work, require the employer to bargain over the effects of that change, protect incumbent headcount or classifications, and commit to retraining or placement rather than displacement.

Read that against what an execution agent does. An agent that assigns labor, re-sequences pick paths, or replaces an engineered standard is changing bargaining-unit work by definition. That is not a technology decision an operations team can make unilaterally in a represented facility. It is a bargaining event.

This produces a distinction that matters far more than the pilot-conversion numbers:

  • Decision-support agents — ones that monitor, alert, summarize, and recommend to a supervisor who still decides — deploy under most contracts without triggering anything. They are also the ones in most of the pilots being counted today.
  • Execution agents — ones that direct work — are a different legal object. They can be fully functional, fully integrated, fully paid for, and still sit switched off pending a negotiation that runs on the contract calendar, not the software roadmap.

Most vendor roadmaps march from the first category to the second. Very few price the transition.

Why the market can’t see it

The blind spot is structural, not stupid.

Enterprise AI is sold to IT and supply chain executives. Labor relations sits with HR and legal. In most organizations those groups meet at implementation, not at selection — which means the constraint surfaces after the contract is signed and the ROI case is already circulating.

The same split shows up on the investment side. An analyst modeling adoption for a warehouse software vendor will reach for pilot counts and conversion rates, because those are the numbers disclosed. But a pilot conversion measures purchase intent. It does not measure deployability, and in a represented facility those can be years apart. Nobody on that diligence call is reading a collective bargaining agreement.

So you get adoption curves built on a metric that is real, publicly reported, genuinely encouraging — and silent on the thing most likely to determine the timeline.

What this means if you’re buying

Three things change if you take the constraint seriously.

Sequence by contract status, not by site readiness. The obvious rollout order is largest site first, or best-instrumented site first. The order that actually works is: non-represented sites first, then represented sites with contracts recently ratified and automation language you’ve already read, then sites with expirations approaching. The technically ideal pilot site is sometimes the worst first deployment.

Bring labor relations into selection, not implementation. The question to ask before signing is narrow and answerable: which functions in this product direct bargaining-unit work, and which advise a supervisor who directs it? A vendor who can’t answer that cleanly is telling you something.

Put the bargaining timeline in the ROI case. If the benefit case depends on execution-level autonomy, the payback period includes a negotiation. Building the case without it doesn’t remove the negotiation; it just guarantees the number is wrong.

What to watch

A few signals will tell you which way this goes over the next two years.

Whether automation language becomes standard rather than notable in distribution contracts. Right now it’s a bargaining win worth a press release. If it becomes boilerplate, the constraint stops being site-specific and becomes sector-wide.

Whether contract expiration calendars start correlating with vendor deployment announcements. They should, and when they do it will be visible.

And whether any vendor starts selling a deployment tier designed around this — an explicit advisory-mode product with a documented boundary, sold on the basis that it doesn’t trigger bargaining. That would be the clearest sign the industry has noticed. As of now, I’m not aware of one.

The general point

The interesting failures in applied AI are rarely about intelligence. A model that can direct a warehouse is a solved-enough problem. Whether it is allowed to is a question answered by people in a room with lawyers, on a schedule nobody in the technology stack controls.

That’s not a warehouse quirk. It’s the shape of the thing everywhere AI meets the physical economy and the physical economy has representation, licensure, or liability — freight, healthcare, construction, utilities, ports. Capability arrives first. Consent arrives on its own calendar, negotiated and legally binding.

Anyone forecasting the second by measuring the first is going to be early, and being early on an adoption curve is expensive.

Deploying agents into a physical operation?

I advise operators and investors on where applied AI actually stalls — from the seat that signed the capex and ran the P&L afterwards.

See AI advisory services →

The adjacent question — why so many of these deployments stall before they reach production at all, and why the widely quoted failure rate is measuring the wrong denominator — is taken apart in why AI pilots fail in warehouses.

Sources

  • Manhattan Associates Q2 FY2026 earnings call — Active Agents install-base penetration and pilot-to-subscription conversion.
  • Blue Yonder — agentic warehouse, logistics, and inventory operations agents; AI agent testing system (DC Velocity, Logistics Viewpoints).
  • International Brotherhood of Teamsters — 2026 food distribution and grocery warehousing contract ratifications and strike authorizations (multiple releases, teamster.org).

Two further constraints sit alongside the contractual one. What an agent costs once it is actually running, as opposed to what the rate card implies, is worked out in what an AI agent actually costs to run. And for anyone moving food through a network, the regulatory constraint is now a patchwork rather than a standard — fifteen states restrict PFAS in food packaging and they do not agree on what PFAS is.

Colin Reed is an operations executive and independent consultant. He writes on applied AI in the physical economy — distribution, warehousing, fleet, and food service — from the seat that approves the capital and runs the P&L afterwards. AI advisory services →

Written by Colin Reed

Colin Reed is an operations executive and independent consultant whose career spans foodservice distribution, airline catering, last-mile delivery, and hospitality, from restaurants and bars to luxury resort food and beverage programs. He has held distribution P&L with capital authority and run enterprise system selections across warehouse management, labor management, and ERP platforms, and advises funds, corporates, and management consultancies through expert networks and research firms. On Modern Wealth Model he writes on applied AI in the physical economy, Bitcoin, and long-horizon retirement planning — and builds the calculators.

Full bio · Contact · Consulting