AI and Enterprise Planning: Cutting Through the Noise

If you run planning, operations, or technology for a large organization, the ‘AI’ noise right now is hard to tune out. AI-native platforms, agentic workflows, composable architectures. The message is urgent and consistent: what you have is not keeping pace, and the cost of waiting is rising.

Some of that is true. Some of it reflects the enthusiasm of a market with a lot of capital looking for returns. The hard part is separating the two while you are also trying to run a planning organization, hit operational targets, and manage a technology portfolio that was never simple to begin with.

The same dynamics show up across planning, operations, and technology functions regardless of industry. Most planning leaders are working through three questions at once: what do I do with what I have, where does new technology earn its place today, and when does a hard problem actually require a different kind of solution. This post takes them in order: what is actually broken in most planning environments, what the technology can and cannot yet do, what the vendors are really selling, and a simple framework for figuring out your next move.

What’s actually broken

Enterprise planning systems were built to record and coordinate, not to decide. ERP is a transaction ledger. It is excellent at capturing what happened and structurally limited at helping you figure out what to do next. The planning layers built on top were designed around exception-based human workflows. At large manufacturers, entire teams would start each day working the exception queue, reviewing alerts, chasing issues, making calls. Advanced analytics and algorithms came later, to surface better exceptions or tighten the plan.

Another limitation is assumption decay. Planning logic gets calibrated against conditions that then change. Demand patterns shift, lead times evolve, operational parameters drift. A better AI model applied on top of stale assumptions produces output that reflects the quality of the logic underneath, not the sophistication of the model. Fixing the assumptions can matter as much as improving the technology, and it is the step many organizations skip when they buy something new.

Underneath both limitations sits data readiness, and the picture there is uneven. Some organizations have done the data work and are well-positioned to layer AI onto a sound foundation. Gartner’s 2025 Hype Cycle for Artificial Intelligence found that 57% of organizations estimate their data is not AI-ready. Where that fragmentation exists, it caps what any AI investment can deliver, no matter how good the underlying models are. Where the work has been done, the path to value is considerably shorter.

With that foundation in place, it is worth understanding the technology claim every vendor now leads with, before it shapes how you evaluate what you are being sold.

Agentic AI: a real direction, an early market

Gartner predicts 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in early 2025. That is what you see on all their LinkedIn posts, blog articles, etc. In most cases, there is a gap between the message and what is actually running in production. At the core of this is Gartner’s term, “agentwashing”: rebranding existing automation and assistants as agents without new capability underneath. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, not because the models fail but because of escalating costs, unclear business value, and inadequate risk controls.

What is working in production today is often narrow and bounded: exception routing, automated supplier communications, replenishment triggers within set parameters, order expediting, and load acceptance, where an agent weighs an inbound shipment offer against lane history and capacity targets and decides whether to take it. They share a profile: a clear scope, clean inputs, and explicit criteria for when to escalate to a human. That profile is also the boundary. It defines how far current agentic approaches can reliably go.

What stays hard is the step from bounded transaction automation to complex planning decisions. Scheduling across multiple resources with competing constraints, plans that have to satisfy conflicting requirements at once: these are not problems generative AI was built to solve. AI is not a single technology. Machine learning models, optimization solvers, constraint-based algorithms, and simulation engines each handle different problem types with different strengths. Generative AI reasons well through language and excels at synthesis and summarization. Problems that demand good decisions across large, constrained solution spaces draw on a different toolkit, one with decades of production history in planning environments.

That said, this field is advancing quickly. Many brilliant people are working to make AI agents operate at scale on truly difficult problems. But that distinction matters the moment you weigh a vendor’s agentic claim against the planning problem you actually need solved.

With an honest read on what is broken and what the technology can currently deliver, the three positions being pitched in the market come into much clearer focus.

What the market is selling

Enterprise software has always cycled between fragmentation and consolidation. Specialized point solutions in the 1990s created the integration nightmares ERP was meant to solve. Then cloud SaaS lowered the barrier to adding new tools, and the sprawl came back in a new form. That history is why three distinct positions are being pitched at the same time, and why every one of them claims that AI changes everything.

The large incumbent platforms, ERP vendors, established planning tools, are largely in ‘stay with us’ mode. Your data is already here, switching costs are real, and AI capabilities are improving, so disrupting what works is a risk you do not need to take. There is truth in this, particularly around data integration and governance. Consolidation around a well-governed core has sound logic, and the integration overhead of adding new systems is a genuine cost that vendors in the other camps tend to minimize. What these platforms are less forthcoming about: architecture decisions made fifteen years ago do age, AI features marketed as agents are often considerably simpler than that label implies, and roadmap promises tend to track vendor development priorities rather than your hardest problems.

The AI-native vendors argue that old architecture has a structural ceiling. Layering new capability onto systems designed for a different era hits diminishing returns, and the only real path forward is a modern foundation built for AI from the start. This has merit, I think specifically around information velocity. Traditional planning systems run on batch cycles, making decisions on conditions that are already hours old, while modern architectures work from real-time or near-real-time data streams. The pitch overstates how fast most organizations can pull off a clean-slate transformation, but the underlying architectural critique is not wrong.

The third position is the hybrid approach: keep your existing investment and extend it beyond what it does natively. The reflex is to read this as a compromise, the option you settle for when you cannot afford to rebuild. It preserves proven infrastructure and adds targeted capability exactly where the platform was never designed to go. The logic is straightforward: lower integration risk than a rebuild, faster time to value than a full replacement, and domain-specific depth that no general platform will build out for every vertical. The cost: you are adding a seam that has to be owned and kept healthy. You are trading dependence on a single large incumbent for a combo including a specialist. Done in a focused manner, around a few problems that warrant it, that trade is sound. Done reflexively across every gap, it recreates the sprawl that started this cycle in the first place.

Testing these pitches against what is actually broken in your environment, and what the technology can deliver today, is the starting point for any decision worth making. The framework below is built for exactly that.

Three paths: where you are, where you should be, and how to tell

Most organizations do not have one planning problem. They have several, at different stages of maturity, across different parts of the operation. So the three paths are not a one time choice, and not a single verdict on the whole stack. You apply them problem by problem, and you will usually be activating one capability, augmenting another, and building a third at the same time.

The trap is treating those as fully independent calls. Solve every problem on its own terms and you recreate the exact sprawl this cycle keeps producing: a pile of disconnected fixes that no longer add up to anything. The harder discipline is making the separate paths roll up to one architecture, a shared foundation the individual solutions sit on, so the portfolio stays coherent as it grows. Which path a given problem is on matters less than two questions asked together: is it moving forward, and do the pieces still fit?

The first path is activation. Your platform can handle the problem, but the data foundations are weak, the planning assumptions have decayed against current conditions, or the AI capabilities you already own have never been properly deployed. This is more common than most organizations want to admit. The symptom is persistent underperformance that gets blamed on platform limitations when the real issue is upstream. Buying something new before fixing the foundation just reproduces the same result on newer software. The right investment here is data quality, process control, and deploying what you are already paying for. Generative AI can be helpful at this stage too, summarizing disruption impacts and surfacing context for planners, all without requiring any platform change.

The second path is augmentation, and the test is simple: does a product already exist for this problem? When one genuinely does, this is the best deal on the board, faster and cheaper than anything you could build, with most of the risk already wrung out by every customer who came before you. The catch is that genuine fit is rarer than it looks. Most planning products solve the problem they were built for and approximate everything else, and the gap between a product that sounds applicable and one that actually fits your operation is where these projects quietly stall. So the question is never whether the demo is impressive. It is whether this vendor has solved your specific problem before, in an environment like yours, with outcomes you can verify, and what the product does when two of your constraints conflict and there is no clean answer. Push on what sits underneath the interface, too: real optimization, or rules and heuristics dressed up to look like it. A product that genuinely fits earns its place fast. One that nearly fits becomes a customization project with no end.

The third path is depth, and you reach it when the honest answer to that test is no. The problem is specific enough, or consequential enough, that no packaged product fits, which does not mean starting from a blank sheet. The answer is a solution built around your actual constraints, increasingly on a platform designed for building this kind of decision and optimization application, combining solvers, ML, your data, and domain expertise into something that fits the problem instead of bending the problem to fit a product. It is an engineering commitment rather than a purchase, but the instinct that it must therefore be the slow, expensive path is wrong. Scoped to one problem and built on the right platform, depth can move as fast as a product rollout, sometimes faster, because you are not waiting on a vendor’s roadmap or reshaping your operation around their assumptions. The discipline is the same one that works everywhere else: start contained, prove value on one problem, then extend.

Consider a manufacturer running a major planning platform for network-level supply and demand planning. The platform does that job well. But production scheduling (machine sequencing constraints, changeover logic, material dependencies) is something it was never built to handle at operational depth. Rather than replace the platform or attempt a full rebuild, they add a specialist scheduling engine that reads from the same data, solves the specific problem, and returns a production plan that feeds back into the platform their planners already use. The existing investment is preserved, the capability gap is closed, and the hybrid architecture proves itself not in theory but in throughput recovered and planning time redirected. That same organization might be on path one for a data quality issue in its demand signal and path two for a specific inventory problem at the same time: different paths, different problems, one coherent operation.

The way forward

The signal that you are on the wrong path is rarely a failed implementation. It is quieter than that: years of incremental investment that never compounds, the same problems persisting while the upgrades pile up, because the spend keeps answering a question no one actually asked.

Cutting through the noise comes down to an honest read on what you already have, the same three questions this post opened with. What is the platform genuinely good at and still serving well? That you protect rather than replace, and that is activation. Where was it never designed to go, and do your hardest problems live in exactly that gap? That is where you augment or build. And how much runway does it really have? Funding the next decade of a sound system is a very different call from paying down the last years of one that is aging out. Read honestly, those answers do not leave you choosing between three paths. They tell you which one each problem is already on.

That read is the work the noise is built to skip. The vendors are selling you their position; your job is to know your own. Get it right and you can move deliberately while the market reacts. With the technology shifting this fast, a clear-eyed view of what you already have is the closest thing to a durable advantage there is.