What agentic AI really changes in CP&R supply chains

Every consumer brand has a plan. The harder question is what happens to it by Thursday.

Picture a planning team on a Thursday morning. A promotion has landed harder than expected; the weather has turned in two regions, and one product has gone quietly viral. The plan that looked sound on Monday is now the best guess.

This is not an edge case. In our latest Capgemini Research Institute study, Resilient by design, 77% of consumer goods organizations named forecasting uncertainty, driven by fast-shifting consumer behavior and external factors, as a top supply chain challenge.

Forecasts will keep missing. What separates one brand from another is what happens next.

The response gap 

Take a heatwave that lifts demand for one product in one region, a week earlier than anyone planned. Or a promotion that does better in one region but inventory has been deployed closer to a different region. How long does it take for that signal to become a decision? Who sees it first? Who must agree before stock moves, production shifts or a shipment is rerouted? Who does a true total cost-to-serve and service analyses to support decision-making.

That interval, between the moment something changes and the moment the organization acts on it, is the response gap. It is where availability, margin and customer trust are won or lost, and it has little to do with how good the forecast was. And increasingly, closing that gap is less a forecasting problem than an orchestration problem: connecting signals, decisions and action fast enough to change the outcome.

This is where I think agents earn their place, and it is also where our latest research gets interesting.

Agents are arriving. Nobody is handing over the keys.

Adoption is moving. 22% of consumer products organizations have now piloted or implemented AI agents in their supply chains, up from 15% a year ago.

But only 24% expect a fully automated supply chain within three to five years. Industrial manufacturing, with its standardized processes and longer lead times, sits at 39%. And across all sectors, the share of organizations expecting agents to make most supply chain decisions in that window has slipped from 66% last year to 56%.

I read that as maturity, not retreat. Teams that have moved from pilots into production have discovered how many decisions sit inside a supply chain, how uneven the data can be, and how often an exception needs human judgment. In consumer goods, where decisions are often made in minutes or days rather than weeks, that lesson arrives faster. The question has shifted from how much to automate to how to design the handoff between agent and person.

Build around the outcome, not the function

Most supply chain technology has been built around functions. Planning has its tools, procurement has its tools, logistics has its own, and the handoffs between them are where time and margin leak away.

In a recent blog, I wrote about PowerThreads thinking: Why AI needs an operating thread across functions rather than another collection of isolated use cases. Our latest supply chain report makes that argument very tangible. When demand changes, the outcome depends on planning, inventory, production, logistics and commercial decisions moving together, not on any one of them becoming individually smarter.

The more promising pattern runs the other way: organizing agents around an outcome the business cares about, such as the product being on the shelf when the customer wants it.

In forecast-to-fulfill, a demand-sensing agent watches signals like point-of-sale data, promotions, weather and social sentiment. A supply resilience agent looks at supplier and plant capacity, while a fulfillment agent determines where inventory should move. Order-to-cash follows the same logic, with agents covering order commitment, disruption management, returns and settlement, all pointed at on-time delivery.

The value comes from the thread connecting these decisions around a shared outcome, rather than optimizing each one in isolation.

The prize is practical: better availability, less excess stock, fewer markdowns. Early signs are already visible across the industry, from multi-agent procurement at Unilever to real-time inventory support for field teams at PepsiCo.

Oversight is a design choice

Here is the number I keep coming back to. 86% of consumer products respondents say that a lack of human oversight would raise operational and reputational risk. That is the highest of any sector surveyed, ahead of the 80% average.

I don’t read that as a resistance to AI. I read it as people who have lived with the consequences of a bad call at scale and want someone accountable for the big ones. 

In practice, agents take monitoring, analysis, coordination, and routine execution. People keep the judgment: setting guardrails, handling exceptions, making trade-offs and approving the high impact moves. This is why agentic supply chains are as much an operating-model question as a technology question. Decision rights need to be designed from the start: what can an agent decide, what requires human judgment, when does an exception trigger escalation, and who ultimately owns the outcome?

It also means expecting to be overruled. Planners will challenge agent recommendations, especially early on, and they should. The mistake is letting those overrides vanish. When a person changes a recommendation, the system should capture why, and what happens next. That is how trust gets built from evidence instead of faith.

“Autonomy is the wrong scoreboard. The brands that win will be the ones that keep people and agents looking at the same picture, at the same moment.”

Mayank Sharma, VP, Global Intelligent Operations Offer Lead, Capgemini

Knowing the risk is not the same as moving on it

Consumer products leaders are not naive about risk. 81% call geopolitical instability a significant challenge for their supply chain today, and 61% worry about keeping operations running smoothly.

But only half say they can adjust quickly and cost-effectively to short-term disruption.

That distance, between seeing the risk and being able to respond to it, is the response gap again, now at network scale. A disruption-management agent that senses a deviation early, tests what it means across sourcing, production and distribution, and lines up a coordinated response gives people hours back when they need them most.

It reaches upstream, too. For brands buying agricultural commodities, monitoring signals like water stress and deforestation can flag sourcing risk before a shortage turns into an empty shelf.

The shelf is only one outcome

On-shelf availability has an online twin: the delivery promise. Whether the customer is standing in an aisle or clicking “buy,” the outcome depends on multiple decisions working together across demand, inventory, fulfillment and logistics.

And the shelf or delivery promise is only one expression of this challenge. The same orchestration logic applies to managing a return, protecting margin or converting an order cleanly into cash. Each outcome crosses multiple functions, systems and decision points. No single agent or function owns the whole result.

This is where PowerThread thinking extends the argument beyond the shelf. The same orchestration principle can connect signals, decisions and human-agent actions around the outcomes that matter across the enterprise. The opportunity for agentic AI is not simply to automate more steps, but to help the enterprise respond as one. 

Where I would start

  1. Find the response gap: Identify a recurring point where a business signal arrives faster than the organization can act.
  2. Anchor it to a business outcome: Choose availability, service level, inventory, margin or another measurable outcome rather than starting with the agent.
  3. Trace where the response slows: Follow the decisions across planning, procurement, production, logistics and commercial teams and identify the handoffs creating delay.
  4. Define human-agent decision rights: Be explicit about what agents can execute, where people intervene and how overrides become learning signals.

Demand will not become more predictable, and supply chains will not become less exposed to disruption. The advantage will come from shortening the distance between signal and action, connecting people, agents, data, and decision rights around the outcomes that matter.

Perhaps that is the real test of an agentic supply chain: move beyond the independent efficiency of its agents to how effectively the enterprise responds as one.