How can AI help in demand planning?
Connected to your planning system, an AI assistant answers questions about the plan in plain language: what needs attention today, what changed and why, how a number was built, why an item needs buying, and where a gap to budget comes from. It shows its work so planners can check it, and it only proposes changes; people approve them.
Seventy percent of large-scale organizations will adopt AI-based forecasting to predict future demand by 2030.
Most talk about AI in planning is about forecasting algorithms. But the biggest day-to-day gain for planners is simpler: asking the plan a question and getting a clear, checkable answer. Today, answering "why did this number move?" means drilling through a hierarchy, opening edit logs, and asking around. With an AI assistant connected to the planning system, it's one question.
This article, the first in our series on using AI in the planning process, walks through six practical use cases, each with an example of what the conversation looks like. The examples come from the planning assistant we've been building at K3; the names and numbers in them are made up.
Chat is a second way into the same plan
An AI assistant doesn't replace the planning grid. A grid works when you already know where to look. A conversation works when you know the question but not where the answer is. Both read the same numbers, history, and forecasts:
| Today | With an AI assistant |
|---|---|
| Drill down the hierarchy to find the problem | Ask "where am I most exposed this month?" and go straight to the cells that matter |
| A number is just a value | A number comes with how it was built: the model, the edits, the promotions, and the reasons |
| Edits are cell values | Changes start as intents ("move the spring promotion to week 32"), recorded beside the cells they change |
| Context lives in people's heads | Reasons, events, and assumptions become records the next person can see |
1. The morning briefing: "What needs me today?"
Instead of opening five reports, a planner asks one question. The assistant checks what moved since the last published plan, the biggest forecast misses, edits that came in without a reason, items that will run short after the order cutoff, and open supplier issues, then lists what deserves a look first.
What needs me today?
Five things, most urgent first:
- Harbor lace oxford, black: short 1,200 pairs in March, past the order cutoff supplyOptions: expedite, or move stock from the West DC
- Factory response on PO 5120 waiting 3 days: cut 600 pairs, moved 1,800 to May purchase order
- Bellmont spring forecast up 15% since the last publish plan moverReason given: "Buyer confirmed extra doors for spring"
- Carroway's March forecast cut 8% with no reason given no reason
- Trail boot, tan: September actuals 24% under forecast, the biggest miss in the category accuracy
Illustrative. Names and numbers are made up.
2. "What changed, and why?"
Plans change constantly: edits, new published versions, stock loads, purchase order updates. An AI assistant can read all of it as one feed, with who made each change, how, and why. The "why" is there because planners are asked for a reason when they change a number, so the explanation is stored next to it. Changes with no reason stand out, so they can be followed up while people still remember.
What changed in the plan this week, and why?
- Oct 9 · Maria L. raised Bellmont spring 14,000 → 16,100 edit"Buyer confirmed extra doors for spring"
- Oct 13 · Factory response on PO 5120: cut 600, moved 1,800 to May purchase orderFactory note: "Leather allocation reduced for April"
- Oct 12 · Plan version 14 published by the planning manager publish
- Oct 11 · J. Patel cut Carroway's March 3,900 → 3,590 editNo reason given
- Oct 10 · Weekly stock load for 3 distribution centers stock load
Illustrative. Names and numbers are made up.
3. Explain any number, showing the work
Ask "why is this number what it is?" and the assistant breaks it down: which measures it's made of, how it compares with actuals and the statistical baseline, the edits behind it with who and why, the promotions in it, and how accurate the plan has been. Every answer separates the evidence, the assumptions it relied on, and what it didn't check, so a planner can verify it in seconds instead of trusting it blindly. We show a full example of this answer layout in our article on corporate memory.
The assistant can also start from what's on screen. A planner selects a cell in the grid and chooses "Ask about this," so "why is this up?" means that cell, without retyping the item, account, and month.
4. "Why do we need to buy this?" in plain arithmetic
Open-to-buy and replenishment numbers are often the hardest to explain, because they come out of a calculation nobody sees. An AI assistant can lay the calculation out period by period, the way a buyer would check it by hand: opening stock plus what's arriving gives supply; demand plus safety stock gives need; the difference is what to buy, adjusted for minimum order quantities and periods that are already too late to order for.
Why does this need a buy of 2,400 pairs?
April needs 2,000 pairs that nothing covers, and the factory sells in multiples of 1,200, so the order rounds up to 2,400. March is short too, but it's already past the order cutoff.
| Month | Supply opening + arriving | Need demand + safety stock | To buy | Ending stock supply − demand |
|---|---|---|---|---|
| Feb | 1,200 + 800 = 2,000 | 1,500 + 400 = 1,900 | 0 | 500 incl. 400 safety stock |
| Mar | 500 + 0 = 500 | 1,700 + 450 = 2,150 | 1,650 too late to order | 0 short 1,200 |
| Apr | 0 + 0 = 0 | 1,600 + 400 = 2,000 | 2,000 → 2,400 | 800 after the buy arrives |
- The 800 pairs on PO 4471 arrive in February as confirmed purchase order
- Minimum order: multiples of 1,200 pairs factory terms
- March's shortfall is lost sales, not carried into April as backorders planning settingIf it were backordered, April would need 1,200 more pairs
Illustrative. Names and numbers are made up.
5. Purchase orders and factory performance
When a factory responds to a purchase order by cutting quantities, splitting shipments, or moving dates, the planner needs to know two things: where the negotiation stands, and what accepting the latest response would do to the plan. The assistant can show each round with its note and the effect month by month, along with how the factory has performed over time: cut rate, split rate, how often and how far dates move, response time, and on-time arrival.
What happens if we accept the factory's response on PO 5120?
April receipts drop by 2,400 pairs and May gains 1,800, so the order ends 600 pairs smaller. Two styles would run short in April. This factory has cut or split 3 of its last 10 orders.
| Round | Change | April | May |
|---|---|---|---|
| Ordered, Sep 22 | 6,000 pairs for April | 6,000 | 0 |
| Factory, Oct 13 | Cut 600; move 1,800 to May. "Leather allocation reduced for April" | 3,600 | 1,800 |
Illustrative. Names and numbers are made up.
6. Explaining the gap to budget
Finance asks a different question: why is the plan above or below budget? The assistant can split the gap into volume (selling more or fewer units), mix (selling a different blend of products), and price, with the three adding up exactly to the total, then point to the items driving each.
Why is the spring plan $420K under budget?
Lower volume explains more than the whole gap. Mix and price recover part of it, because the units we're keeping are the higher-priced boots.
| Driver | Effect | Main items |
|---|---|---|
| Volume | −$610K | Sandals, down 18% after a key account cut its order |
| Mix | +$95K | More boots in the blend |
| Price | +$95K | Spring price increase on 40 styles |
| Total | −$420K |
Illustrative. Names and numbers are made up.
What makes these use cases work
None of these answers are useful unless planners trust them. Three things make the difference:
- It shows its work. Every answer separates evidence, assumptions, and what wasn't checked, so it can be verified quickly.
- It proposes; people decide. The assistant can suggest a change, with its reason and evidence, but nothing changes until a person approves it.
- It sees only what you can see. The assistant works with the same permissions as the person asking. A sales rep limited to their own accounts gets answers about those accounts only.
There's a longer-term benefit too. When reasons are captured alongside every change, the plan's history becomes something the company can learn from. That's the subject of the next article in this series: using AI to build corporate memory in planning.
Frequently asked questions
How is AI used in demand planning?
Beyond forecasting models, AI assistants connected to the planning system answer planners' questions in plain language: what needs attention, what changed and why, how a number was built, why an item needs buying, and where gaps to budget come from. They show their evidence and propose changes for people to approve.
Does an AI assistant replace the planning system?
No. It's a second way into the same data. The planning grid is still the fastest way to make repeated edits; the assistant is faster for questions, explanations, and changes that span many cells.
Can AI change the forecast on its own?
It shouldn't. A well-designed assistant proposes changes with a reason and evidence, and a person approves them. The approved change is that person's edit.
What is an AI morning briefing for planners?
It's a single question, such as "what needs me today?", that returns a prioritized list: plan changes since the last publish, the biggest forecast misses, edits without a reason, items short after the order cutoff, and open supplier issues.
Can AI explain open-to-buy and replenishment numbers?
Yes. It can lay out the calculation period by period: opening stock plus arrivals for supply, demand plus safety stock for need, and the difference as the quantity to buy, adjusted for minimum order quantities and order cutoffs.