Last week I had the chance to join some really smart people on stage at CSCMP EDGE in Nashville to talk about where S&OP is headed in the AI era. Ajay Natarajan of Manhattan Associates moderated. I shared the panel with Bart De Muynck, who brings 34 years in supply chain including eight at Gartner; Joyce Qu, a data scientist at Niagara Bottling; and Kedar Kulkarni, who ran S&OP at Amazon and Microsoft before founding Strom AI.
We covered a lot of ground in 45 minutes: data foundations, decision velocity, even agents negotiating with other companies' agents. But the point I kept coming back to, and haven't stopped thinking about since, is this:
Agentic S&OP isn't going to stall because the technology isn't ready. It's going to stall because people don't trust it.
The technology is ready enough. We can now sense changes continuously, simulate dozens of scenarios, and orchestrate workflows that used to take weeks of email chasing. That part is real.
But I've been through enough large-scale change to know how these programs actually end. They live or die less on the tech and more on whether people adopt it. And people don't adopt what they don't trust. Skip that part and you've spent a lot of time, money, blood, sweat and tears on what becomes one more failed project.
So let's talk about trust: what earns it, what kills it, and where I think the biggest untapped opportunity sits.
Continuous S&OP was never a desire problem
Practitioners have talked about moving past the monthly S&OP cycle for twenty-plus years. Most companies never got there, and it wasn't for lack of wanting to. It was capacity.
Joyce nailed it on the panel: we don't hold monthly S&OP meetings because demand only changes once a month. We hold them monthly because it takes that long to pull the data, reconcile the spreadsheets, build the storyline and make the deck.
I hear this from customers frequently; in fact, one large multinational manufacturer I work with told me flatly, "We just don't have the capacity to run different scenarios. We go in with one and hope the execs don’t ask a lot of questions." Their planners are underwater.
This is where the agentic era changes the math. When agents handle the reconciliation, spot the exceptions and draft first-pass scenarios continuously in the background, the monthly review can become weekly, or event-triggered. Continuous doesn't mean more meetings. It means fewer wasted ones, and the ones you keep are about decisions, not data.
But here's the shift people underestimate. When software moves from advising to acting, drafting the scenario, routing the workflow, executing within guardrails, the trust bar goes way up. You can ignore a dashboard. You can't ignore an agent that's doing work on your behalf. And if your team doesn't believe in what it's doing, they will quietly rebuild the spreadsheet next to it. We've all seen that movie.
Trust is earned three ways
I don't think trust is a soft, fuzzy thing you hope shows up after go-live. It has to be designed in, both in the product and in how you implement it. In my experience it comes down to three things.
1. Value at every level, not just one
S&OP is the decision connective tissue of a company. It runs across functions, and it also runs up and down the organization. So AI has to pay off for everyone whose time it asks for.
Planners: automate the toil. Getting the data, cleaning the data, chasing people down. Instead of reviewing hundreds of exceptions, they see the handful where service, inventory or margin is actually at risk, and get to use their well-earned judgment.
Managers: pre-game and post-game analysis (more on this below, because I think it's the sleeper).
C-suite: translate the operating plan into their language, which is cash, margin and risk. Not units.
If AI delivers for one level and ignores another, you don't get trust. You get frustration. And that undermines the whole point of S&OP, which is alignment.
2. Transparency: show the work
A black box is simply unacceptable anywhere a decision touches customers, inventory investment, service commitments or financial trade-offs. It's not enough for AI to say what should happen. It has to show why.
Say demand shifts toward a higher-margin segment while capacity is constrained. A useful recommendation tells you which signal changed, which constraint became binding, which inventory position is exposed and what it means for service and margin. Now commercial, supply and finance are arguing from the same facts.
Transparency also starts upstream, with an honest forecast. A single-number forecast is a guess wearing a costume. Nobody has ever hit a monthly forecast exactly, yet we keep planning, and measuring people, as if we could. A probabilistic forecast shows the range of outcomes you actually need to plan for. Gartner reports that 76% of companies are forced to act outside their S&OP plans when disruption hits, and only 30% have a planning process that handles uncertainty well. That's a trust problem hiding in plain sight.
3. Change management that starts with the decision
Please don't sprinkle AI on top of the thousands of alerts your team already ignores. That's AI-washing, and it doubles the work instead of reducing it.
Start with the decision, not the technology. The question isn't "Where can we add AI?" It's "Which decision do we need to make faster, better or with less manual effort?" Then make governance concrete: who can accept a recommendation without review, when a human has to sign off, and how overrides get logged.
And I'd flip the old mantra. Forget fail fast. Aim to succeed fast. Prove it on one decision, earn a track record, and let success build on success.
The real opportunity: pre-game and post-game
If you take one idea from this post, make it this one. I think it's the most underrated opportunity in agentic S&OP.
The pre-game is knowing the risks before you commit. Picture a manager whose product launch is tracking below plan. Today, someone builds that analysis from scratch, under pressure, usually late. Instead, they could pull up three ready-made scenarios: hold the plan, expedite supply, or reset the forecast. Each one shows the service and margin trade-off, upstream and downstream.
I like to borrow an idea from cybersecurity here. Security teams don't wait for an incident to decide what to do. They run scenarios and write runbooks ahead of time. Your demand range & inventory response should work the same way. Plan to the most likely case, but decide in advance what you'll do if reality drifts toward the low end or the high end. Then week three of a launch isn't a fire drill.
The post-game is keeping score. And I want to be really clear about what I mean, because this is where it often goes wrong. Keeping score is not about blame. S&OP is a team sport: you win or lose together. The point is a shared fact base:
Which assumptions held, and which didn't?
Which signals did we miss or underweight?
Which trade-offs paid off?
Where does the decision process itself need to change?
That's how a miss becomes a better decision next cycle. It's structural learning. The organization gets smarter, not just the model.
The same goes for overrides. A planner overriding the system isn't a failure. They may know about a promotion, a customer or a supplier issue the model can't see yet. What matters is that the override is logged, its effect is measured and the team learns from it.
So why doesn't everyone already do this? Because nobody has time. The after-action review dies the moment the next cycle starts. This is exactly where agents shine. They can monitor, remember and explain why a KPI was or wasn't hit, at scale, across any time horizon. That's what turns S&OP from a heavy, monolithic process into a living way of working.
It also changes how you grade a forecast. Stop asking only "how close was the number?" Start asking: Did the actual land inside the range we gave? How fast did we narrow it as real data came in? Did we make the right call at each decision point?
And there's a side effect I love. Sales and operations don't always get along. As Kedar put it on the panel, they don't have to like each other; S&OP is there to seek out the truth. A shared, honest scorecard means teams spend less time debating whose numbers are right and more time deciding what to do next.
Where I'd start
Pick a persistent, chronic pain point. Get the stakeholders in a room and ask how you'd do it if you were starting from scratch today. Don't just bolt AI onto the process you have.
One use case I think is ready now: demand scenario creation on top of a probabilistic forecast. An agent interrogates the forecast across its range, then builds scenarios at different probabilities so planners can see what each one means for supply, revenue and margin. The guardrail is simple. People still select, review and submit the scenario. The agent does the legwork; the team owns the call.
Then measure it with the KPIs you already have. Revenue, margin, inventory turns, working capital, service. Those don't change. What changes is that you can watch them improve, and explain why.
And, as Bart said, stop calling it a pilot. Call it Phase One.
Trust is the product
The technology is going to keep getting better, fast. Trust won't improve on its own. It has to be built into the product, into the implementation and into how teams work together.
The companies that pull ahead won't be the ones with the most AI. They'll be the ones whose people believe the plan enough to act on it.
So I'm curious: where is trust breaking down in your S&OP today? Is it the data, the black box or the people? Tell me in the comments.
#SOP #SupplyChain #AgenticAI #SupplyChainPlanning #CSCMPEDGE
