Your demand planner forecasts what the market wants. Your ERP system calculates lead times. Your production scheduler optimizes the line. But none of them talk to each other about what your plant can actually deliver. Equipment fails mid-shift. A key supplier misses their window by a week. You're short-staffed on the night shift and no one flagged it until the backlog hit the floor. The forecast was accurate. The constraint killed the plan.
This is where most manufacturing supply chains break. Demand sensing and constraint planning have to be one signal, not two silos. The RealAI manufacturing stack treats it that way: predictive maintenance, live OEE and constraint-aware planning feeding the same picture of what the line can do this shift, so the plan reflects real capacity instead of optimistic spreadsheets.
The Forecast-vs.-Reality Gap
A typical demand plan runs like this: you feed a forecasting model months of sales history, POS data, promotional calendars and maybe weather. It outputs a demand curve. Your planners feed that into the ERP. The ERP calculates raw-material lead times, suggests reorder points, and publishes the MRP schedule. Production takes the schedule, assigns it to lines and shifts, and starts the week. By Wednesday, a critical compressor goes down. The line that was supposed to make its quota falls short. By Friday, the demand forecast is still accurate — but the line is now behind, and you're expediting to cover a customer commitment you can fulfill, just not on the original plan.
The accuracy of the demand forecast was never the problem. The forecast said you needed the units. You have the orders. But the plan assumed full uptime and full labor. In the real world, equipment runs below nameplate, suppliers slip, and staffing is a negotiation with whoever is available on Tuesday morning. (Plant-level uptime and supplier on-time rates vary widely by sector and asset — treat any single figure as an industry benchmark, not a universal.)
Constraint planning folds those real numbers back into the demand signal. Not as a safety stock or a static buffer (which masks the real problem and ties up working capital). But as live feedback: when equipment is down, the demand that can be pulled forward shrinks; when a supplier delays, the raw-material window for downstream make-to-order work shifts; when staffing drops, the throughput ceiling for that shift moves. The plan has to adapt in real time to the actual constraints, not pretend they don't exist.
Constraint Mapping: What's Actually Stopping You
The first step is unblinking clarity on what stops the line — and for how long.
In a mature manufacturing operation, you already have this data in fragments:
- SCADA and historian logs capture equipment runtime and stop reasons (programmed stop, fault, changeover, throttle). Your real availability is sitting in that data already — you know which equipment, which shifts, which days of the week drive the loss.
- Supplier scorecards show on-time delivery rates, lead-time variability and defect patterns. If one supplier ships on-window most but not all of the time, your raw-material replenishment window for products that depend on them needs to open earlier than the nominal lead time implies.
- Labor tracking systems (or shift rosters) flag staffing headcount per shift per line. If you're regularly short on nights, you don't have a labor shortage; you have a line-throughput constraint designed into every night shift.
- Quality and rework records show what percentage of production hits the line and what percentage comes back for rework. That's not just a quality metric; it's a yield loss that shrinks available capacity.
The assessment maps all four against your top product families by volume and margin. You're looking for the constraint that is active — the one actually limiting throughput right now, not the theoretical worst-case. Often there is only one, sometimes two. A bottleneck press. A supplier with a long lead time and a real slip rate. A night shift that runs well below day-shift capacity.
From Demand Forecast to Constraint-Aware Production Plan
Once you know the active constraints, the demand forecast changes.
Instead of publishing "you need this many units of SKU-X in week 12," the model publishes what you can make in week 12 — because the press has long changeovers and runs below nameplate, and a key supplier delivers only part of its committed volume on-window. That sounds pessimistic. It isn't — it's honest. And it is actionable.
Now the production scheduler can make real choices:
- Pull demand forward — if the press is the constraint and it is sitting idle earlier in the window, can you move SKU-X orders into that idle time to use the spare capacity?
- Smooth demand across multiple lines or shifts — if you are short-staffed on nights, can you break a large order across day and night and accept a slightly later delivery?
- Negotiate with suppliers — if a supplier delivers on-window only part of the time, can you go dual-source for this material, or can you open your raw-material window earlier to absorb their slip?
- Invest in constraint relief — if a press is genuinely the bottleneck and it is booked solid for months, adding a shift, a new press, or a third-party contract becomes a defensible capital decision because you have the numbers.
The demand forecast does not change. The plan does.
- 45%
- Fewer unplanned stops · predictive maintenance
- Live
- OEE per asset
- 100%
- In-line inspection
- 4-6 weeks
- Assessment and constraint mapping
Live OEE: The Signal That Closes the Loop
OEE — Overall Equipment Effectiveness — is usually a monthly KPI. You run the numbers at month-end and see how a line averaged. Useful for tracking. Not useful for planning tomorrow's demand pull.
Live OEE inverts that. Every asset in the plant streams availability, performance and quality data into one real-time signal. When availability dips — a bearing is warming up, a servo is drifting out of spec, a die is beginning to wear — the planning system sees it today, not in the month-end report. The demand that that line can absorb next shift adjusts immediately.
This is where the constraint model stops being a static number ("line 3 runs at its average OEE") and becomes a dynamic one ("line 3 is running below its average today because of a thermal throttle; its capacity is reduced; the orders that were pencilled in for it on Thursday should move to Friday or to line 2 if line 2 has room").
That real-time feedback folds into the broader supply chain: if line 3 drops, and SKU-X is made only on line 3, and the demand for SKU-X is firm for Thursday, then either you expedite maintenance, or you talk to the customer, or you pull that SKU's demand into line 2 if a changeover is acceptable. But you make the choice before Thursday, not after the miss.
The systems that ship this — that run live OEE into the production planning engine — shorten replanning cycles from once a week (Monday morning, all damage done already) to once a shift or once a day. Unplanned stockouts drop because you are not gambling on uptime that did not happen.
A demand forecast is only as good as the constraints it runs against — the equipment that breaks down, the suppliers that slip, the shifts you're actually staffed for.
Supplier Lead Time as a Demand Constraint
Raw-material availability is often the invisible constraint. Your demand forecast says you need a batch of connectors in week 14. Your supplier's nominal lead time is several weeks. So you should have ordered well in advance. But supplier slip is real: they miss a portion of their commitments by a week. And a week of slip means the connectors arrive late, your assembly line is dark when it should be running, and the orders that depend on those connectors slip.
Constraint planning addresses this by folding supplier performance data into raw-material replenishment windows. If one supplier delivers on-time more reliably than another, their effective lead times are different: the less reliable one needs a time buffer built in. If a critical supplier slips, that is a capacity constraint for as long as it takes material to arrive. The demand forecast stays the same; the production plan absorbs the gap by either delaying orders or pulling demand into weeks where materials are already in stock.
Dual-sourcing — a second supplier for critical materials — becomes measurable: if two suppliers slip on uncorrelated timing, their combined on-time rate can be materially higher than either alone. The incremental cost of the second supplier becomes a capital decision because you can quantify how much more throughput and how much less safety stock you can carry. (On-time and lead-time slip rates differ by supplier, geography and category; treat any rule-of-thumb number as an industry benchmark and measure your own.)
Where to Start
The 4–6 week assessment asks three questions:
-
What is the active constraint right now? Run SCADA data (you probably have years), pull supplier on-time and slip metrics from your purchase-order history, and list staffing levels by shift and line. One thing will stand out: the bottleneck press, the critical supplier, the night shift that always runs short. That is your starting constraint.
-
How much capacity is that constraint leaving on the table? If the bottleneck asset runs below what it could theoretically produce, you are leaving capacity on the table every day. How much revenue is that? What is the inventory carrying cost of the safety stock you hold to cover that variability? That is your payback floor.
-
Do you have the data to feed a constraint model? SCADA logs, historian, ERP data, supplier scorecards, labor roster. You probably have all of it. The question is whether you can connect them. That is the transform-phase roadmap — asset-specific models wired into the SCADA and maintenance workflow where technicians already work.
The output is one ranked constraint to attack first, a clear picture of the payback (lower inventory, fewer stockouts, fewer unplanned order slips, faster replanning cycles), and a roadmap to build the constraint model, integrate the data feeds, and run it as a pilot on the highest-value product family. Brownfield-first: the models bolt onto the SCADA, historians and PLCs you already run, with no rip-and-replace of plant infrastructure. And because production conditions move — new tooling, recipe changes, seasonal throughput — the models are monitored for drift and retrained against current operating data, not the day-one baseline.
You don't start with a perfect model. You start with the constraint you are certain about — the one causing real cash loss right now — and you build the planning logic that accounts for it. Everything else follows.
“A demand forecast is only as good as the constraints it runs against — the equipment that breaks down, the suppliers that slip, the shifts you're actually staffed for.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production — sovereign models, governed data, and agents you can audit. Start with a value-first assessment.