The Silos Never Left. Now They’re Getting Agents.

You've seen the demo by now. A demand agent spots a spike. A supply agent checks capacity. A logistics agent rebooks the freight. An orchestration agent supervises, and the agents — this is the part that gets the room — talk to each other. It's impressive engineering, and the first time you see it, it feels like the future.

Look again, and count two things. Count the agents: one per planning function — demand, supply, logistics — each optimizing its own objective. Then listen to what they're passing each other: numbers. The forecast is 12,400 units. The lead time is 45 days. The yield is 96%. Confident, singular numbers, about a world that has never once honored them.

Two inheritances, in one demo. The org chart decided who gets to optimize. The point estimate decided what they optimize against. Enterprise planning has spent thirty years learning — expensively, in escalation meetings and write-offs — what both assumptions do to enterprise value. Agentic AI is now automating both, faithfully: software that gets faster exactly where it needed to get wiser.

The first inheritance: who gets to optimize

Here's the uncomfortable truth about the last three decades of planning technology: silo optimization engines, plural, arrived years ago. The enterprise optimization engine, singular, never did.

Every function got its own engine and its own number — demand planning got forecast accuracy, supply planning got utilization, logistics got cost per unit, inventory got days on hand. And each function, year after year, got measurably better at its number, while the enterprise didn't get proportionally better. Every operator knows why. The demand plan that maximized accuracy ignored margin. The supply plan that maximized utilization built the wrong inventory. The freight plan that minimized cost per unit missed the customer that mattered. Functional excellence rose; enterprise value leaked through the handoffs, invisible to every function because no function was measuring it.

The only cross-functional layer we ever built was made of meetings. S&OP — and its more ambitious cousin, IBP — is a monthly conference room invented to reconcile plans that should never have been separate. Its existence is a confession: coordination is not integration. Coordination means each party solves its own problem and the parties then reconcile — around a table or over an agent-to-agent protocol; the mechanism doesn't matter. Integration means there is one problem, solved whole, where every trade-off is visible in one frame. The trade-offs that matter most are cross-functional by nature — take the freight hit to protect the strategic customer; build the "wrong" inventory before the capacity window closes — and each one is invisible to every individual function, and therefore to every individual agent. No volume of conversation between agents surfaces a number none of them is computing.

Integration was always the promise. The silo engine was always the product. And now the demos are showing us one agent per silo, negotiating — the conference room rebuilt in software, with the meeting deleted. The meeting, whatever its faults, was a governor: the one place a human occasionally caught cross-functional damage before it shipped.

The second inheritance: what they optimize against

Open any planning system and look at what it actually contains: thousands of confident numbers. The lead time is 45 days. The yield is 96%. The commodity cost is $1.84. The pipeline deal closes on the 15th. Every one is a distribution flattened for the system's convenience — the software was built to compute with points, so the organization learned to feed it points.

The demand forecast is merely the most famous of these fictions. It has a thousand quieter siblings, and they do most of the damage. Run the diagnostic on your own operation: pull last quarter's escalations and sort them by cause. How many were true demand surprises — and how many were a lead time that slipped, a yield that dipped, a cost that moved, a deal that slid a quarter? The forecast takes the blame in the steering meeting. The supporting cast starts the fires.

The bitter part: the evidence of every one of those uncertainties already exists, in your own systems. The supplier's actual lead-time record is in the ERP, spread and all. The way deals churn their close dates is in the CRM. The yield's real behavior is in the MES. Your organization has the distributions. The planning layer threw them away — then hired planners to firefight the gap between the point and the truth, and called the firefighting "experience."

The twist: fixing the first mistake amplifies the second

Now suppose someone finally builds what the industry promised — one model, one connected plan, every function in a single solve. It sounds like the happy ending.

Watch what actually happens. In the siloed world, a supplier slip blew up one planner's week; the silo walls, for all their sins, were accidental firebreaks. Wire everything together on point estimates, and every local tremor propagates through the whole connected model — one violated assumption re-solves the factory, which moves the freight, which re-promises the customers. The integrated-but-certain plan doesn't fail less often than the fragmented one. It fails more often, and everywhere at once, because it holds thousands of point assumptions and honors each one absolutely.

Operators know the symptom by name — plan churn: Monday's confident plan contradicting Friday's. And they know the immune response, because it's endemic in every "integrated" deployment: planners stop trusting the lurching system, pad their lead times, hide safety stock in unofficial corners, quietly stop entering what they know. Integration without resilience doesn't just fail — it drives the organization to rebuild fragmentation by hand, inside the integrated system, as self-defense. For thirty years we treated integration as the destination. It's necessary. It is also, on deterministic mathematics, an amplifier.

The same holds one level up, where discrete risks live — the port strike, the tariff, the wobbling single-source supplier. A one-question audit: if that strike's likelihood moved from 25% to 50% tomorrow, would your plan change? In most organizations the honest answer is no — the risk lives in a register reviewed quarterly, while the plan lives in a system that has never heard of it. A risk whose probability cannot move the plan isn't being managed; it's being memorialized. And that's for the risks someone wrote down — which leaves a question worth asking any vendor and any team: who is watching for the ones nobody wrote down?

Put agents on top of all this and the inheritance completes. Human latency was the throttle: someone had to notice, convene, decide. Remove the humans and keep the point estimates, and the whiplash runs at machine speed — every wobble re-planning the enterprise before anyone convenes. The market is already booking the consequences: Gartner publicly predicts over 40% of agentic AI projects canceled by end of 2027 on costs, unclear value, and inadequate risk controls — and this year added that organizations treating agent trust as binary, locked down or fully autonomous, will end up demoting and decommissioning their agents at scale. MIT researchers found some 95% of enterprise GenAI pilots deliver no measurable P&L return — a failure they attribute to tools that don't embed in real workflows, not to weak models. The receipts pile up the same way (McKinsey's "gen AI paradox," IBM's CEOs calling their own stacks piecemeal, S&P's abandonment rates — next Monday's article walks the full ledger). The pattern underneath is architectural, not algorithmic.

What follows, in order

Accept the diagnosis and the requirements stop being a wish list; they become a sequence, each step forced by the last.

If the decomposition is wrong — if no functional agent can see the enterprise trade-off — then the agents need one number: a single economic objective that rolls to the P&L, which no silo can improve at the enterprise's expense without the cost showing.

If the mathematics is wrong — if points can't carry truth — then the plan must be chosen across the ranges: lead times, yields, costs, capacities, close dates learned from your own history rather than typed into parameter screens, with named risks priced in at their probabilities. This is also where stability comes from: a decision chosen because it holds across the range doesn't lurch every time one number wobbles. Robustness, not freezing, is the cure for churn. And it's a rebuild, not a feature — a planning core must reason over distributions to do it, which is precisely why it never shipped as an upgrade.

And only when both are fixed does the interesting question even become askable: has the machine earned the authority to act? Not the binary Gartner warns about — locked down or fully trusted — but authority the way every functioning organization grants it: conditionally, on evidence, scoped, and revocable, decision by decision. Integrated, resilient, autonomous — in that order, because each makes the next one safe.

The evaluation changes with the diagnosis

Once you accept that the problem is architectural, vendor evaluation changes completely. You stop scoring demos and start auditing assumptions — with questions any serious platform, including ours, should have to answer in the room. What single number do your agents jointly maximize, and can you trace it to my P&L? When two agents disagree, what resolves it — a negotiation protocol, or one solve that sees the whole trade-off? Which of my planning parameters do you treat as uncertain — demand only, or lead times, yields, costs, and capacity too — and do the ranges come from my history or your assumptions? If a named risk moves from 25% to 50% likely, show me, live, what changes in the plan. And which decisions may the machine take alone today, on what evidence was that right granted, who revokes it — and what did last quarter's autonomous decisions actually earn?

If the answers are protocols and hand-waving, you're looking at the org chart, automated — betting its numbers, faster.

The next generation of planning platforms won't win by having the most capable agents. They'll win because they stopped handing their agents assumptions that were wrong before AI ever arrived. Thirty years of silo engines taught us that optimizing one function at a time destroys value across the whole; thirty years of parameter screens taught us that certainty typed into a field was never certainty at all. The irony of this moment is that the technology being rushed into that old architecture is the first technology capable of running the right one — one objective, many futures, authority earned. The silos never left. The question is whether we finally retire them, or hand them agents.

On August 25 we'll show what the alternative looks like, live, in working software: "Autonomous, Where It Counts — The Five Levels of Earned Decision Rights."

Registration link: vyan.ai/resources/webinars/autonomous-where-it-counts