How Much Autonomy Should an AI Agent Have in Your Warehouse?
Quick answer. An AI agent should have as much autonomy as the cost of a wrong decision allows. Sort every warehouse decision into three tiers before go-live. Tier 1, auto-execute: low-exposure decisions the agent makes and logs on its own. Tier 2, recommend and approve: the agent proposes and a human confirms. Tier 3, always escalate: compliance, customer commitments, large financial exposure, or no precedent, where the agent flags and stops. Unless the operation sets these boundaries, the vendor's default configuration sets them.
At a Glance
- 88% of organizations regularly use AI in at least one business function, and 23% are scaling an agentic AI system somewhere in the enterprise. (McKinsey, The State of AI in 2025)
- 51% of respondents whose organizations use AI report at least one negative consequence, with inaccuracy the most commonly cited. (McKinsey, The State of AI in 2025)
- 40% of respondents from large organizations (more than $1 billion in annual revenue) report scaling AI agents, up from 27% a year earlier. (McKinsey, The State of AI in 2026)
Agent use is growing, and inaccuracy is already the most common problem organizations report. Defining where an agent's authority ends, before it goes live and not after, is a decision leaders can get ahead of. Treating it as an operating model choice, not a technology setting, is what makes bounded autonomy work.
The Autonomy Question Nobody Resolves Before Go-Live
Warehouse teams are running AI agents today across receiving, wave planning, carrier selection, slotting, and exception handling. Unless the operation sets the boundary between what the agent decides on its own and what it hands to a human, the vendor's default configuration sets it. Vendor defaults are generic by design and are not calibrated to your operation's risk. A setting that looks conservative during a pilot can authorize consequential calls in production at the speed of software, not the speed of a morning walkthrough.
Sophisticated models do not answer the autonomy question. The operation does, and it is far easier to answer before the agent goes live. The research points the same way. McKinsey's 2025 survey found that AI high performers, about 6% of respondents, are more likely than others to have defined processes for when model outputs need human validation, one of the factors that most distinguished them. McKinsey's 2026 survey adds that high performers are much more likely than others to work to mitigate unauthorized or unintended AI actions.
A useful reference point is how the same principle was applied before agents: a custom WMS deployed across 60 3PL facilities, with exception handling built around the operation's own workflows and business rules, not adapted from a generic platform. For the broader sequencing of automation investments, see In What Order Should You Automate Your Supply Chain? A Practical Strategy Framework. For the data and operating-model foundations beneath agentic deployments, see our earlier piece on agentic AI readiness.
Defining the Terms
Bounded autonomy is a formally defined set of constraints that governs which decisions an AI agent may execute independently, which require human confirmation before execution, and which must always be escalated regardless of the agent's confidence level.
A decision-rights framework is the operational document that assigns each category of warehouse decision to one of three tiers, with specified financial thresholds, volume parameters, and escalation conditions.
Decision rights in a warehouse have always existed. Supervisors carry them. The question of who can authorize a short-ship, override a receiving variance, or release a held load is not new. What changes with agentic AI is that the agent needs those same rights codified explicitly, because it cannot read organizational context or ask a floor manager before acting.
The Three-Tier Decision-Rights Framework
Tier assignment is based on financial exposure, reversibility, downstream impact, and compliance relevance, not on the agent's confidence score.
- Cycle count adjustment within pre-defined tolerance bands
- Wave release when labor, equipment, and dock availability meet pre-set thresholds
- Carrier selection within pre-approved rate cards
- Replenishment trigger for high-velocity SKUs within approved reorder parameters
- Receiving variance decisions that exceed the defined tolerance band
- Priority order insertion that disrupts an active wave already in progress
- Carrier re-tender within a pre-approved spot rate ceiling (spot rates vary by market, unlike contracted rate cards)
- Any decision that modifies a confirmed customer-facing commitment, including short-ship authorization
- Vendor claim processing above a defined dollar amount
- Regulated product decisions, compliance holds, or quarantine releases
- Any exception that has no precedent in the historical decision log
Calibration signal: override rate. Everest uses a working threshold, not a published benchmark. If managers override Tier 2 recommendations more than 30% of the time, the boundary is likely mis-calibrated, and the tier assignments need revision, not just a log review. A very low override rate is not proof that the boundary is right either, since it can also mean recommendations are being approved without real review.
Where This Gets Missed
Five mistakes undermine bounded autonomy:
Keeping the vendor's default settings
A default reflects no knowledge of your financial thresholds, your customer commitments, or your compliance obligations. The boundary has to be set by the operation, then configured into the platform.
Assigning tiers by the agent's confidence score
Confidence describes how certain the model is, not what an error would cost. A decision with high exposure or limited reversibility belongs in a higher tier however confident the agent is.
Not tracking override rates
Without the override rate, nobody can tell whether the Tier 2 boundary is set correctly. Gaps then surface through production incidents instead of through calibration.
Widening the boundary too early
A good period is not a mandate to expand. Tier 1 should grow through the documented process described below, not on the strength of recent results.
Letting the agent act on stale data
A boundary assumes the agent can see the current state of the operation. When inventory positions, dock status, or labor capacity are out of date, a well-drawn tier becomes an unsafe one.
Case Study: The Same Principle, Applied Before Agents
Custom WMS Across a 60-Facility 3PL Network | Seasonal Retail, North America
This engagement predates agentic AI, and the case study does not describe a tiered model. It shows the discipline underneath one: exception scenarios defined and tested before go-live.
A seasonal retailer ran its supply chain through 60 distributed 3PL warehouse partners across North America, with fragmented data exchanges, limited real-time visibility into carton and pallet movements, and manual coordination across systems. After evaluating traditional enterprise WMS platforms and custom-built options from other providers, the retailer concluded that its seasonal, high-velocity 3PL model required a purpose-built approach.
Everest designed and implemented a custom SaaS WMS built around the client's workflows, seasonal rhythms, and business rules, with exception-driven processing for operational mistakes and data inconsistencies. Delivery included 500+ test cases and regression cycles, with phased releases timed ahead of peak demand. Exception scenarios such as mis-sorts, blind receipts, and unlabeled cartons were handled inside the system, with reduced manual reconciliation. Deciding in advance how a system treats each exception type is the same exercise the tiers above formalize for an AI agent.
Results: 4.6M+ inbound and 4.6M+ outbound cartons processed during peak seasonal cycles, with weekly peak volumes exceeding 850,000 cartons across 60 3PL distribution centers.
Source: Everest Technologies, everesttech.com/case-studies/seasonal-retail-warehouse-execution
Widening the Boundary Over Time
The initial framework is a starting point, not a permanent configuration. As override patterns stabilize and the exception log matures, Tier 1 can expand. The process matters as much as the boundary itself. Everest recommends these starting parameters, and each operation should set its own:
- Set a 90-day observation window before adjusting anything. Widen only after the exception log has stabilized and the override rate has held consistent for at least 30 consecutive operating days, across a full operating cycle that includes non-peak periods.
- Move one process at a time from Tier 2 to Tier 1. This isolates the effect and makes calibration straightforward.
- Require a formal sign-off for each expansion. Operations, finance, and compliance should approve it and the process should be documented.
- Do not change tiers immediately before or during peak. Stable results at peak show that the current boundary holds under stress. They are not, by themselves, a reason to widen it.
Where Everest Can Help
Building a decision-rights framework before an agentic deployment is straightforward when the right people are in the room. The harder part is knowing which questions to ask, which exception types carry the most risk, and where the integration layer will surface gaps that change the boundary. That is where Everest Technologies fits in. As a Manhattan Associates Platinum Partner with 300+ warehouse management system go-lives, Everest works across the platform, the integration layer, and the operation itself.
Before the platform conversation starts
Everest's assessment work starts with how a distribution center runs today. For an agentic deployment, that scope can include mapping current decision flows, identifying who owns each exception type, and setting the financial and volume thresholds that define tier assignments. The result is a decision-rights document that goes into implementation rather than being built retroactively from production incidents.
At the integration layer
An agent's boundary is only as reliable as the data it reads. Everest's integration work, built on platforms such as MuleSoft, Boomi, and Kafka, ensures the signals the agent acts on, including inventory positions, dock status, and labor capacity, are accurate, current, and consistently sourced. An agent acting on stale data is operating outside its actual boundary regardless of what the framework says.
After go-live
The override rate does not manage itself. Everest's post-go-live work includes SLA-backed L2 and L3 support, production monitoring, incident and release management, and continuous optimization for Manhattan Active® environments. Reviewing override rates and tier boundaries belongs in that same cadence, and it is something Everest can help put in place.
If an agentic deployment is in planning or a current one is generating more exceptions than it resolves, the decision-rights framework is the right place to start. Everest Technologies is glad to walk through what that review looks like.
Let's talk about your warehouse autonomy strategyFrequently Asked Questions
Bounded autonomy means giving an AI agent a written limit on its authority. Some decisions it executes on its own, some it recommends for a human to approve, and some it always escalates. The limits are set before go-live and reviewed against data afterward.
Decisions in Tier 3 stay with a named human role, and the agent flags and stops. These include any decision that modifies a confirmed customer-facing commitment, including short-ship authorization; vendor claim processing above a defined dollar amount; regulated product decisions, compliance holds, and quarantine releases; and any exception with no precedent in the historical decision log. Dollar thresholds are set by the operation.
Tier 1 decisions are high-frequency, low-exposure, and correctable within the same shift at limited cost. The agent acts and logs. Tier 2 decisions carry moderate financial exposure, downstream dependencies, or exposure above the Tier 1 threshold. The agent recommends, and a human confirms before any action is taken. Assignment is based on exposure and reversibility.
Everest uses a working threshold rather than a published industry benchmark. If managers override Tier 2 recommendations more than 30% of the time, the boundary is likely mis-calibrated and the tier assignments need revision, not just a log review. A very low override rate is not proof that the boundary is right either, since it can mean recommendations are being approved without real review.
Everest recommends a minimum 90-day observation window, with the override rate holding consistent for at least 30 consecutive operating days, across a full operating cycle that includes non-peak periods, before any boundary change. Tiers should not change immediately before or during peak. Each operation should set its own parameters.
Ownership typically sits with the VP of Operations or the equivalent supply chain leader, with sign-off required from Finance and Compliance for any change to tier boundaries. The framework document is not an IT artifact. It is an operating model document. IT implements what the framework specifies; it does not define the thresholds.
Treat the vendor's settings as a starting point to review, not as the boundary. The decision-rights framework review maps your actual decision flows against whatever the vendor has configured and adjusts thresholds to reflect your operation, not the vendor's reference model.
Sources
- McKinsey & Company, "The state of AI in 2025: Agents, innovation, and transformation," McKinsey Global Survey, November 5, 2025 (1,993 participants in 105 nations, in the field June 25 to July 29, 2025). https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai-2025
- McKinsey & Company, "The state of AI in 2026: On the road to ROI," McKinsey Global Survey, August 25, 2026 (1,719 participants in 97 nations, in the field May 4 to June 8, 2026). https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- Everest Technologies, "Seasonal Retailer Deploys a Custom-Built WMS Across 60 3PL Facilities." everesttech.com/case-studies/seasonal-retail-warehouse-execution
- Everest Technologies, "Your Manhattan Active® Implementation Partner" (Platinum Partner status, 300+ WMS go-lives, SLA-backed L2 and L3 support). everesttech.com/manhattan
About the Author
