McKinsey Wants Retailers to Deploy Agentic AI. Their Own CFO Survey Explains the Problem.
Two McKinsey pieces dropped in 48 hours this week. One told retailers to push decision-making authority down to AI agents - moving from dashboards that show what happened to systems that act on what is happening. The other documented CFOs across industries quietly building cash reserves and liquidity buffers against geopolitical disruption. Neither piece referenced the other. They should have.
The gap between those two stories is not a contradiction. It is a map of where agentic AI deployment fails - and occasionally works.
What "Dashboards to Decisions" Actually Requires
The McKinsey retail piece is framed as an opportunity brief: agentic AI can automate routine merchandising work and improve decisions at scale. The specific capability being described is a shift from passive monitoring - dashboards that surface what happened - to active execution, where AI agents take steps autonomously based on what they observe.
This is a real capability shift, not marketing language. An agent that monitors inventory levels, detects a shortfall against projected demand, queries supplier availability, compares landed cost across sources, and initiates a purchase order is doing something qualitatively different from a dashboard that shows low stock. It is collapsing what was a multi-step human workflow into a single automated loop.
The case for this is legitimate in a constrained context. When the optimization objective is clear (minimize stockouts while controlling working capital), when the decision variables are bounded (which SKUs, which suppliers, what quantities, what timing), and when the environment is relatively stable, agentic AI performs well. The system does not get distracted, does not lose track of context across a 40-minute meeting, and does not defer obvious decisions because it is waiting for a quarterly business review.
The Assumptions Underneath the Promise
What makes that context work is also what limits it. Agentic AI for merchandising assumes a world where the supply chain behaves roughly as modeled, where supplier relationships hold, where logistics costs fall within historical ranges, and where demand signals reflect actual consumer behavior rather than pre-panic buying ahead of a tariff announcement.
Strip any of those assumptions, and the optimization problem changes character. You are no longer minimizing stockouts - you are making a judgment call about whether to lock in inventory at current prices against the risk of a 25-percent tariff landing in 90 days. That is not a merchandising decision. It is a capital allocation decision dressed as a merchandising decision. And agentic AI, absent very careful design, will solve the problem it was trained to solve rather than the problem you actually have.
What the CFO Survey Reveals About the Macro Assumption
The McKinsey CFO survey is a more interesting piece than its headline suggests. The finding is not that CFOs are newly worried about geopolitical risk - they have been worried for months. The response has been to double down on performance measurement and build cash and liquidity buffers.
That behavioral pattern is worth reading carefully. Building cash buffers is not a precision hedge against a specific risk. It is a bet that the distribution of outcomes has widened - that the variance in what might happen is large enough that holding optionality is more valuable than deploying capital into optimized operations. When the CFO of a mid-market retailer is building liquidity reserves, it is not because they have a view on which tariff regime will land. It is because they no longer trust their forecast models to bound the problem.
This is the environment in which retailers are being told to delegate more decisions to AI agents. The agents will optimize within the parameters they were given. The parameters may no longer reflect the world.
This is the kind of systemic signal that STI's research tracks across market environments - the divergence between operational AI confidence and macro-level financial risk posture is a leading indicator worth watching.
Why Both Can Be Right Simultaneously
The important point is not that agentic AI is wrong for retail. It is that both McKinsey pieces are compatible if you accept a distinction that neither paper makes explicit: there is a category of retail decisions for which agentic delegation is correct right now, and a category for which it is not.
Routine restocking of high-velocity, low-risk SKUs from established suppliers: good candidate for agentic automation. The optimization function is stable, the variables are bounded, and the cost of an error is recoverable.
Committing to 90-day inventory positions on imported goods during a period of active tariff renegotiation: not a good candidate. The optimization function has changed, the variables are unbounded, and the cost of an error is not recoverable. Edward Jones figured this out early - their agentic AI deployment deliberately stops short of autonomous execution on anything that touches external relationships or irreversible commitments.
What neither McKinsey piece provides is the framework for sorting your decision portfolio into those two buckets. That is the actual work.
The Quality Premium That Emerges From This Gap
Nick Maggiulli at Of Dollars and Data published something this week that reads as personal finance advice but applies more directly here. His argument, citing James Clear: there is always room for quality. The market is not saturated with good things. It is saturated with average things.
The research he cites is a 2014 meta-analysis across 40 years of performance studies: intrinsic motivation predicted more unique variance in quality than financial incentives. Caring about the work matters more than paying for it. Incentives drive quantity. Motivation drives quality.
In the context of agentic AI deployment, this surfaces an underappreciated dynamic. As AI handles more routine decisions - and it will - the relative scarcity of sound judgment on non-routine decisions increases. The organizations that create differentiated outcomes from AI deployment are not those that automate the most. They are those that correctly identify which decisions deserve automation and preserve human deliberation for the rest.
The Editing Principle Applied to Decision Architecture
Maggiulli's other insight is that quality improvement often looks like subtraction - removing the things that dilute focus rather than adding new approaches. The organizations doing this well with AI are doing something similar: they are not trying to wire every decision into an agentic loop. They are pruning the decision portfolio to identify where AI genuinely belongs, and protecting the remaining decisions from premature delegation.
The McKinsey agentic organization report from earlier this week documents the failure mode of the opposite approach - companies that ran AI everywhere without redesigning the underlying work, and ended up with expensive tooling on top of unchanged workflows. The lesson is structural, not tactical.
If you're evaluating which decisions belong in an agentic loop and which require human oversight, our analysis tools can help surface the variables that distinguish automation-ready decisions from those that require human judgment.
Building the Decision Taxonomy
The practical question is how to build the sorting framework that both McKinsey pieces implicitly require but do not provide. Four variables tend to predict whether a decision is an agentic AI candidate:
Objective clarity. Can you specify, precisely, what a good outcome looks like? Minimize stockouts, maximize margin per square foot, reduce supplier lead time variance. If the objective function requires judgment to define, it probably requires judgment to optimize.
Variable boundedness. Are the inputs to this decision within the range your models have seen? A tariff shock from a geopolitical event is out-of-distribution data. An AI agent trained on normal operating conditions will not flag it as anomalous - it will try to solve the problem with the tools it has. That is a feature in stable environments and a liability in unstable ones.
Environmental stability. How quickly is the ground truth changing? The faster the environment shifts, the faster your agent's training data becomes stale. CFOs are building cash buffers precisely because environmental stability has declined. That should be an input to every agentic AI deployment decision being made right now.
Error reversibility. What is the cost of a wrong call? A merchandise overstock is recoverable over several weeks. A locked-in 90-day import contract during a tariff negotiation is not. Irreversible decisions with high error costs should stay human, regardless of how confident your model appears.
What This Means for AI Strategy in 2026
The interesting thing about this week's McKinsey double-bill is that it reveals a real tension inside how organizations are thinking about AI. The operations function is being pushed toward more agentic autonomy. The finance function is being pushed toward more human discretion and liquidity preservation. These two impulses are going to collide in capital allocation decisions about AI infrastructure investment.
Starling Bank's current campaign - led by CMO Michele Rousseau, built around money tips from 190 real people across the UK - is a different kind of data point. A fintech brand growing through trust-building rather than automation, deliberately deploying human credibility signals in a market where synthetic content is everywhere. That is not a technical statement about AI. It is a strategic statement about where differentiation lives when AI handles competence as a baseline.
The companies that deploy agentic AI well in 2026 will not be the ones that automate the most. They will be the ones that built the taxonomy first - mapped their decision portfolio against those four variables, delegated the right things, protected the rest, and held enough liquidity to absorb the decisions their models did not anticipate.
That is not a technology problem. It is a decision architecture problem. The tools are almost incidental.
If your organization is working through which decisions belong in an agentic loop and which require human oversight, STI's decision intelligence research tracks the patterns across industries. Or if you're earlier in the process, schedule a conversation to work through your specific portfolio.