Seven S&OP challenges that outlive your ERP — and how AI-assisted decisions close the gap

Every company running a serious sales and operations planning process has an ERP system. Most of them still have the same S&OP problems they had before the ERP went in.

That’s not an indictment of ERP. It’s a category error about what ERP is for. An ERP system is a system of record: it captures transactions, keeps the books consistent, and tells you — accurately — what already happened. S&OP is a system of decision: it exists to reconcile what sales wants to sell, what operations can make, and what finance can afford, and to commit the company to one plan. The gap between recording what happened and deciding what to do next is where S&OP actually lives, and no amount of ERP configuration closes it.

Our thesis is simple: the persistent S&OP challenges are decision problems, not data-storage problems, and they yield to AI-assisted decision-making — models and agents that read the data continuously, surface what matters, and put a defensible recommendation in front of a planner — layered on top of the systems you already run. Here are the seven that come up most often, and what closing each one actually takes.

1. Forecasts that repeat history

Ask a demand planner how the forecast is built and you’ll usually hear some version of: last year’s actuals, adjusted by judgment. The ERP aggregates historical sales and applies a smoothing method; the planner overlays what they know about promotions and customers; sales adds its number, which is really a target wearing a forecast’s clothes.

The result is a forecast that faithfully reproduces the past — including its biases. Optimistic sales estimates, sandbagged quotas, and risk-averse operations padding all flow straight into the demand signal, and nothing in the ERP is designed to catch them.

Machine-learning forecasting changes two things. First, it widens the inputs: order patterns, pricing moves, seasonality interactions, and external signals that a monthly spreadsheet exercise can’t absorb. Second — and less discussed — it makes bias visible. When you systematically compare forecast to actual by team, region, and product family, the persistent over- and under-calls stop being anecdotes and become correctable patterns. The forecast stops being a negotiation and starts being a measurement.

2. Everyone plans from their own spreadsheet

S&OP is supposed to align sales, operations, and finance. In practice, each function exports the same ERP data into its own spreadsheet, applies its own assumptions, and arrives at the monthly meeting with its own version of the truth. The meeting then spends its first hour arguing about whose numbers are right instead of deciding anything.

Centralizing the data didn’t fix this, because the problem was never storage — it was that each function needs to see the plan through its own lens, and the tools forced them to rebuild that lens privately. What changes the meeting is a shared working layer: one demand picture, one supply picture, and scenario analysis that every function can interrogate from its own angle without forking the data. When sales, operations, and finance are reacting to the same what-if — “here’s what the plan looks like if this product line runs 15 points hot” — the argument shifts from whose data is right to what the company should do. That’s the argument S&OP was invented to have.

3. Inventory set by rules nobody revisits

Most ERP inventory logic comes down to static reorder points and safety-stock formulas that someone parameterized years ago. Demand variability has changed since then. Lead times have changed. Supplier reliability has changed. The parameters haven’t, because revisiting thousands of SKU-level settings by hand is nobody’s idea of a good quarter.

So the organization oscillates: a stockout scare pushes safety stocks up everywhere, capital quietly accumulates on shelves, a working-capital review pushes them back down, and the next disruption starts the cycle again.

This is exactly the kind of problem continuous optimization is built for. Models that watch actual demand variability, actual lead times, and actual supplier performance can keep inventory targets current at SKU-location level — recomputing them as conditions move rather than waiting for the annual parameter review that never comes. The planner’s job shifts from maintaining settings to reviewing exceptions, which is where their judgment was always most valuable.

4. A planning cycle slower than the market

The monthly S&OP cadence made sense when data arrived monthly. It doesn’t anymore. A demand surge, a supplier failure, or a freight disruption in week one waits until the next cycle for an official response — and by the time the revised plan is approved, conditions have moved again. Planners know this, which is why so much real decision-making happens off-process, in hallway escalations and emergency calls the S&OP plan only hears about later.

The fix is not more meetings. It’s separating the cadence of decisions from the cadence of the calendar. AI-assisted planning keeps the underlying picture current continuously — demand sensed daily, supply positions refreshed as data lands — and runs scenarios on demand rather than on schedule. Routine adjustments within agreed guardrails can execute without waiting for the meeting; the monthly forum is reserved for the trade-offs that genuinely need executives in a room. The cycle stops being the bottleneck and becomes what it should have been: governance.

5. Data you can’t plan on

“Garbage in, garbage out” is the oldest line in planning, and it survives because it keeps being true. Duplicate customer records, product codes that don’t match across sites, unit-of-measure inconsistencies between modules — every planner has a private cleanup ritual before the numbers are usable, and every cleanup ritual is another place versions of the truth diverge.

ERPs store data; they don’t referee it across modules, sites, and the CRM next door. That refereeing is the job of data orchestration: automated pipelines that collect, reconcile, and validate data across systems on a schedule, applying consistent definitions so that anomalies get flagged at ingestion rather than discovered mid-meeting. It’s the least glamorous item on this list and the most important, because everything else here — forecasting models, scenario engines, inventory optimization — inherits the quality of the data underneath it. Orchestration is where any serious AI effort in S&OP has to start; models pointed at unreconciled data just automate the confusion.

An ERP sees your company. It doesn’t see your suppliers’ schedules slipping, the vessel that missed its rotation, or the tier-two shortage that will surface in your production plan six weeks from now. So planners find out about upstream problems the way everyone dreads: when the shipment doesn’t arrive, and the response is expediting, air freight, and allocation calls — firefighting priced at a premium.

Extending visibility beyond the four walls is partly a data problem — bringing supplier commitments, logistics milestones, and external risk signals into the same orchestrated layer as internal data — and partly a decision problem: someone, or something, has to watch that wider picture continuously and judge which of a thousand small deviations actually threatens the plan. That watching is work humans do badly at scale and models do well. An agent that monitors supply signals and raises its hand with “this delay breaks these three orders unless you act by Thursday” converts blind-spot surprises into decisions made while options are still cheap.

In many companies the operational plan and the financial plan are reconciled exactly once a year, at budget time, and drift apart every month after. Operations plans in units; finance plans in dollars; the S&OP process nominally connects them, but when demand shifts mid-year, nobody can quickly say what the revised plan does to margin, cash, or the commitments made to the board. So finance hedges, operations pads, and the plan of record is the one nobody quite believes.

Integrated financial modeling closes this by evaluating operational choices in financial terms as they’re made, not at year-end. Every scenario carries its margin and cash-flow consequences with it: holding extra safety stock on a volatile line, chasing an upside demand case, protecting a launch. When the S&OP meeting can see that trade-off in dollars in real time, finance stops being the function that says no afterward and becomes a participant in the decision — which is what integrated business planning was always supposed to mean.

Start with the decisions, not the technology

A pattern runs through all seven challenges: in each case the data mostly exists, and the decision process around it is what’s broken. That ordering should discipline how you adopt AI in S&OP.

Don’t start with a platform replacement — big-bang planning transformations have a long history of arriving late and underdelivering, a pattern McKinsey and the University of Oxford documented in their joint research on large-scale IT projects, which found such programs routinely run far over budget and deliver less value than promised. Start by picking one or two decisions that are made badly today, on stale or contested data, and fix the path from data to decision for those. Usually that means orchestration first (get one reconciled, trusted picture of demand and supply), then automation of the repetitive analysis around it, then models and agents that recommend — with planners reviewing and overriding, and every override feeding back into the system. Confidence compounds. So does scope.

This is the philosophy we build on at A2go. The A2go Decision Intelligence Platform (ADIP) is built to work alongside your ERP rather than replacing it: orchestrated data underneath, AI agents on top to monitor conditions and support specific planning decisions inside your existing S&OP process. We’d rather change how one decision gets made this quarter than promise to change everything eventually.

Your ERP is doing its job. The seven challenges above were never its job. They belong to the decision layer — and that layer is now buildable.