This started as a quick catch-up between two people who keep crossing paths in the same corner of the world: pharma and manufacturing supply chains. One of us maps and rebuilds planning processes for manufacturers; the other builds and deploys the AI agents that run inside them. Between us, we've walked into well over 400 factories.

The call was supposed to last fifteen minutes. It turned into an hour of comparing scars — and almost everything we landed on traced back to a single idea. So that's the story we want to tell, start to finish: the version from the battlefield, not the webinar.

It begins with a line one of us saw on LinkedIn a few days earlier and couldn't shake: companies are racing to be AI-first, but most of them aren't even process-first.

That gap is the whole story.

But there's a second gap sitting right behind it, one that's harder to name: most companies aren't even decision-first. They have the data. They have the meetings. They run the cycles. What they often don't have is a clear design for what a decision looks like — who makes it, with what information, and what it costs to get it wrong.

Walk into the average mid-market manufacturer and you'll find the process isn't really written down anywhere — it lives inside one or two experienced people. They know when to follow the standard procedure and, more importantly, when to break it. They're the reason things work. And the day they leave, everyone stares at the official process with big eyes, because the real logic just walked out the door.

You cannot put an agent on top of that. There's nothing underneath to put it on top of.

Which is why most of the market is far more conservative than the headlines suggest. The customers who let you push hard, redesign workflows, and deploy agents end-to-end are a thrilling minority — and it's easy to mistake them for the norm. One of us has a rule of thumb for staying honest: whatever you see at the leading edge, divide it by five. Talk to a random person doing the same job at a typical company and the reaction is closer to "I'd never let a machine do that" — and that person is the majority.

So the real constraint isn't appetite for technology. It's whether there's a process solid enough to automate in the first place.

And for planning processes, there's a further constraint: whether the process was ever designed to produce decisions, or just reviews. Most IBP cycles we've seen produce the latter. Information gets consolidated. Scenarios get presented. And somewhere between the numbers and the table, the actual decision disappears.

When the process does hold, the value shows up somewhere deeply unglamorous. Not in the futuristic "ask the data anything" demos, but in the backbone: order-to-cash and procure-to-pay. Order confirmations. Inbound deliveries. Cross-checking documents, batch numbers, shelf life. Getting clean data into the ERP, every time, without a human re-keying it.

It sounds trivial. It isn't — and this is the part that surprises people. Take a plain order confirmation. Skip it, or enter it late, and MRP is suddenly working from a stale picture. Your production plan is wrong. Your promises to customers are wrong. It all propagates downstream. And the window is bigger than anyone realizes: if your average lead time for materials is fifty days, you have fifty days to quietly resolve every discrepancy — why the quantity shrank, why the date moved, why the price changed — before the invoice ever arrives. Do that work continuously and invoice matching at the end becomes effortless. Skip it, and you're digging through old emails trying to reconstruct what everyone meant three weeks ago.

The boring process is load-bearing.

What strikes us, looking across both levels — operational and planning — is that the pattern is identical. Every data quality problem is a decision problem in disguise. A gap in the ERP represents a moment where someone either didn't have clear ownership or didn't have what they needed to act in time. A gap in the planning cycle represents the same thing, one floor up: a trade-off that was never made explicit, a scenario that was never stress-tested, a number on a slide that nobody was willing to be accountable for. Fix the decision design and the data improves. Try to fix the data without touching the decision logic and the same gaps come back in a different shape.

This matters especially for supply chain planning, because planning is where the two levels meet. The transactions feed the plan. The plan drives the decisions. And if the plan isn't designed to produce clear decisions — with explicit trade-offs, financial consequences, and named owners — then clean transaction data just means you're looking at accurate numbers in a process that still won't give you an answer.

A well-run IBP cycle doesn't start with the forecast. It starts with the trade-off. What are we protecting — margin, service level, or working capital? If demand shifts fifteen percent, which do we sacrifice first, and who has the authority to make that call? Those are not questions a planning system answers automatically. They are questions a planning system should be designed to surface. Most aren't.

The scenarios that matter in IBP are not optimistic, base, and pessimistic. They're: what happens to our cash position if this assumption is wrong. What margin do we put at risk if we commit to this volume. What is the cost of waiting one more month before deciding. Those are financial questions sitting inside a supply chain process — and the connection between the two is exactly where most planning cycles break down.

How you make it stick matters just as much as what you automate. The agents that survive contact with reality live where people already work — inside Outlook and Teams — not in yet another platform nobody wants to log into. And you don't flip a switch to full autonomy. You start at one hundred percent human control: nothing posts to the ERP automatically, even with strong guardrails making sure nothing malformed ever could. Then, after a few weeks, the users themselves say it: "Why am I confirming things that are obviously fine? Just send them through." Trust isn't sold. It's accumulated.

The same dynamic plays out in planning. You don't start by automating the executive decision. You start by making the decision moment visible. What's the actual choice on the table? What are the consequences of each path? Once planners and executives experience that clarity consistently — once the meeting stops being a review and starts being a decision — the appetite for more follows naturally. The resistance was never to the technology. It was to the accountability that came with it.

And when something does go wrong, it's almost never the software. We watched a company spend nine months on a major AP implementation that never went live. A focused agent deployment had it posting to the ERP in three weeks — and then immediately surfaced the real problem: receiving was lagging, there were no goods receipts in the system, so accounts couldn't close. The bottleneck was never AP. It was three steps upstream, in a process nobody had fixed. That's the pattern, every time: the technology is rarely the hard part. The process underneath usually is.

We've seen the same pattern in planning transformations. The IBP redesign stalls not because the tool is wrong or the data is missing, but because nobody agreed on what a decision actually looks like in that organization. The meeting runs. The deck gets updated. And three months later, the same conversation happens again with slightly different numbers. The bottleneck is never the system. It's the three steps upstream where someone chose not to be explicit about the trade-off.

None of this happens as fast as the hype promises — and faster than the skeptics think. As Gates put it, we overestimate what technology does in two years and underestimate what it does in twenty. Manufacturing, remember, is fundamentally a cost-reduction business, so urgency always loses to economics. Today's early adopters are spending real money and may not see strong short-term returns. But they're learning what works in their own environment — and that is the one thing you cannot buy off a shelf. When the second wave arrives, it'll be far bigger than the first, because by then the pressure is competitive survival rather than curiosity. The companies who got their hands dirty early will already know what works.

So if we had to compress everything into advice, it's this.

Get your hands dirty — you'll never discover what works in your industry from a slide; run a real use case on a real process and find out where it breaks for you.

Define the process before you automate it, and get leadership to commit, so you end up with something repeatable, auditable, and finally living outside one person's head.

And then take the uncomfortable step: redesign the process for agents, not just for humans. The real transformation isn't automating a human-centered workflow — it's rethinking the decision logic for one where agents do the operational work and people supervise the exceptions. That means handing over some decision authority, and most of the resistance you'll hit lives right there, not in the technology.

And if you're in planning: ask yourself honestly what your IBP cycle is actually producing. Information, or decisions? If the answer is information, that's where to start — before the AI conversation, not after it.

The honest summary is that this work is hard, the market is more conservative than the headlines, and nearly every failure traces back to a broken process rather than broken software. The same is true one level up: nearly every planning failure traces back to a process that was never designed to produce clarity — not to a shortage of data or technology. But that's also exactly where the opportunity is. The companies building the process foundation now — quietly, unglamorously — are the ones who'll be ready when being merely "AI-first" stops being a slogan and starts being the price of entry.

We'd rather talk about where things actually break than add to the pile of everything-is-fine stories. If this matched what you're seeing in your own corner of the trenches, we'd genuinely like to hear it.

‍