AI Demand Planning Software Picks Depend on ERP and Workload

Kinaxis for manufacturers off SAP, SAP IBP on S/4HANA, RELEX for retail and grocery, o9 for one plan across functions. Blue Yonder loses a planning-only purchase. Run the forecast bake-off on your own history before you sign.

By Rajesh Beri·November 20, 2026·14 min read
Share:
A warehouse planning office at dusk: a desk with two monitors showing a weekly demand forecast line chart against actual sales, a printed S&OP binder and a stack of pallet labels beside the keyboard, with racked pallets

Illustration generated using AI

The best AI demand planning software for a multi-site manufacturer that does not run S/4HANA is Kinaxis Maestro. On an S/4HANA estate it is SAP IBP, for retail and grocery it is RELEX, and o9 fits when demand, commercial and financial planning must sit on one model across a complicated network. Blue Yonder loses a planning-only purchase and wins only when you also buy its execution software. In every case the forecasting model decides less than you think. Implementation length, data quality and the discipline of your consensus process decide more, and no vendor competes on those in a demo.

Prices were checked on each vendor's own pages on November 20, 2026.

Best fit What the AI layer does today Published price (checked 2026-11-20) Do not pick it if
Kinaxis Maestro Discrete manufacturers who replan constantly (high tech, industrial, life sciences) Concurrent what-if recalculation, Maestro Agents, Agent Studio None. Contact sales You need a retail store and SKU replenishment engine
o9 Digital Brain Multi-tier networks wanting demand, supply, commercial and financial plans on one graph Agents and the March 2026 APEX capability on a knowledge graph None. Contact sales You can't staff a long build with your own planners
SAP IBP Companies already on S/4HANA or ECC A planning assistant and supply chain agents; a partner-built demand planning agent checks data readiness Metered in blocks of annual cost of goods; price on request Your ERP is not SAP
Blue Yonder Retailers and distributors buying planning plus warehouse, transport and workforce Explainable AI for planning decisions, new agents from ICON 2026 None. Contact sales You only want planning
RELEX Grocery, retail and consumer goods with store-level replenishment and fresh product Machine learning forecasts tied to ordering, production and markdowns None. Contact sales You are a discrete manufacturer with deep bills of material

The workload behind these judgements is a company with about 20,000 SKUs across 300 ship-to locations, weekly buckets, an 18-month horizon and a monthly S&OP cycle. Change that shape and the ranking changes with it; the decision section at the end says how.


Which Demand Planning Software Fits Which Company?

Two facts you already know decide most of the choice: which ERP holds your orders, and whether your hard problem is on the supply side (constrained materials, long lead times) or the store side (millions of SKU-location pairs, short shelf life). Gartner's 2026 Magic Quadrants for supply chain planning come in separate Discrete Industries and Process Industries editions, and buyers should make the same split.

Kinaxis Maestro: The Pick for Constant Replanning

Kinaxis is the pick when your planners spend their week answering "what happens if this supplier slips three weeks." Its architecture recalculates the whole plan concurrently, so a change in demand ripples through supply, capacity and inventory in one pass without waiting for an overnight batch. Kinaxis says it was named a Leader in both 2026 reports and placed highest on Ability to Execute and furthest on Completeness of Vision in the Discrete Industries edition. That is the vendor's own summary of a paid report, so read the report yourself before quoting it upward.

The agent features are newer than the marketing suggests. Maestro Agents launched in October 2025, and Maestro Agent Studio arrived in February 2026, with orchestrator agents and connections to outside agents planned for later in 2026. If an agent capability is on your requirements list, ask to see it running at a production customer.

Who should not buy it: a retailer. Kinaxis is built around manufacturing supply problems, and store replenishment for fresh food needs a different engine. Budget time for the rollout too. A Nucleus Research case study published by Kinaxis describes a pharmaceutical company whose implementation began in 2018 and took more than three years, because each country or manufacturing site needed its own set-up. That is a document the vendor distributes to show itself at its best, and it still says three years.

o9 Solutions: The Pick for One Planning Model Across the Business

o9 is the pick when demand, supply, commercial and finance each keep their own plan in their own tool and the monthly reconciliation eats a week. Its Digital Brain platform puts those plans on one knowledge graph. In March 2026 o9 launched APEX, which it calls an autonomous planning and execution capability, and said it completed more than 130 go-lives in 2025 and moved several AI pilots into production. Both are company figures.

Reviewers rate it well. Gartner Peer Insights shows the o9 Digital Brain Platform at 4.8 from about 190 reviews, and o9 was a Customers' Choice in the 2025 Voice of the Customer report. Gartner labels Peer Insights content as individual users' opinions, so read the low-star reviews for deployment detail.

Who should not buy it: a company that wants a partner to switch on a planning tool while the planners keep doing their day jobs. A single model across functions means every function has to agree on its definitions, and no platform does that organisational work for you. If finance will not join, you are paying for breadth you will not use.

SAP IBP: The Pick if S/4HANA Already Runs Your Orders

SAP IBP is the default for an SAP shop because of integration. Master data, orders and stock already live in S/4HANA, and the integration path is SAP's own. Peer Insights shows IBP for Supply Chain at 4.7 from roughly 185 ratings.

The generative layer is real and narrow. SAP says it is delivering more than 60 purpose-built agents across supply chain processes. The most telling example is a partner-built IBP Demand Planning Agent configuration guide: the agent's job is to check whether sales history is ready before a demand planning run. That is useful, and it is a data quality check, which tells you where the hard part still is.

Who should not buy it: anyone whose ERP is not SAP. You give up the main advantage and keep the SAP licensing model. SAPinsider's review of IBP failures names the most common trap as switching on every available feature instead of what the business can adopt.

Blue Yonder: The Loser for a Planning-Only Purchase

Blue Yonder is a Leader in the 2026 Discrete Industries Magic Quadrant and a Visionary in the Process Industries edition, by its own account, and at ICON 2026 it announced explainable AI features for planning decisions, new agents and a Model Training Factory built on NVIDIA Nemotron. Panasonic bought the company for $7.1 billion, so its owner has deep pockets.

It loses this comparison because its advantage is breadth across planning and execution: warehouse, transport, store workforce and commerce. Buy planning alone and you pay for that breadth without using it. The November 2024 ransomware attack shows what concentration means in practice. The Termite gang hit Blue Yonder's hosted environment, and the disruption reached Starbucks' barista scheduling and pay, and warehouse systems at Morrisons, with Sainsbury's also affected; the gang claimed 680 GB of stolen data. Any SaaS planner carries that risk, and the more of your operation sits with one vendor, the larger the blast radius.

Who should buy it anyway: a retailer or distributor already running Blue Yonder warehouse or transport software, where one data model from forecast to shipment is the reason to buy.

RELEX: The Pick for Retail and Grocery

RELEX is the pick when the problem is millions of SKU-store combinations, promotions and product that spoils. It became a Leader in Gartner's supply chain planning Magic Quadrant for the first time in April 2025, and Peer Insights shows 4.7 across 104 reviews. Its grocery page claims up to an 80% reduction in out-of-stocks; treat that as the vendor's best case.

Who should not buy it: a discrete manufacturer whose constraint is a multi-level bill of materials and a contract manufacturer's capacity. RELEX can plan supply, but its centre of gravity is the shelf.


What Do the Generative and Agentic Layers Actually Do?

Today the generative layer explains and checks the plan, and the forecast comes from older machinery. Underneath every one of these products, the number in the demand plan still comes from statistical and machine learning models: exponential smoothing, regression on demand drivers, gradient-boosted trees and ensembles of them. The chat interface sits on top. It tells a planner why the forecast moved, summarises exceptions, or, in SAP's case, confirms that the history is clean enough to run.

That ordering matches the research. The M5 competition, run on Walmart unit sales with 5,507 teams in the accuracy track, found that the strongest entries learned across many related series at once and combined several models. None of that needs a language model. So when a vendor demo shows an agent answering "why did the forecast for SKU 4471 jump," ask a second question: what model produced the number, and can you see its accuracy history at the level you plan?

Agents that act on their own, moving orders or rebalancing stock without a planner's sign-off, are mostly still announcements. Kinaxis says its orchestrator agents come later in 2026. SAP's agents are rolling out across a portfolio. o9 says pilots reached production but has not published which decisions they make unattended. Treat any autonomous action as a pilot with a rollback plan.


Why Vendor Forecast Accuracy Numbers Are Not Comparable

Vendor accuracy claims cannot be compared because each one is measured on a different customer, at a different level of aggregation, at a different lag, with a different metric. "Accuracy improved 20%" can describe monthly product-family forecasts at a one-month lag, which are easy, or weekly SKU-location forecasts at a twelve-week lag, which are hard. The same history scored both ways gives very different numbers.

The only number worth buying on is one you computed on your own data. Use these metrics:

  • WMAPE (weighted mean absolute percentage error) is the sum of absolute forecast errors divided by the sum of actual demand. Weighting by volume stops slow movers with a handful of units from swamping the score, which plain MAPE allows.
  • Bias is the sum of forecast minus actual, divided by actual. A model with good WMAPE and a persistent 8% over-forecast will still bloat your inventory.
  • Forecast value added (FVA) is the change in accuracy each step of the process contributes compared with a naive forecast. SAS's Michael Gilliland, who popularised the method, has written that in a study of supply chain forecasts by Steve Morlidge, about half were less accurate than a naive forecast.

How to Run the Bake-Off

  1. Hand every vendor the same extract: three years of weekly shipments or point-of-sale data at SKU-location level, plus promotions, price changes and known stockouts. Hold back the most recent 26 weeks.
  2. Score at the level and lag you plan at. If production locks eight weeks out, score the eight-week lag. Use rolling origins, so each vendor forecasts from several cut-off dates instead of one.
  3. Include two baselines you build yourself: a seasonal naive forecast and a free time-series foundation model run on your warehouse (see below). A suite that cannot beat both on your data is selling you workflow, and you should price it that way.
  4. Freeze the configuration before you reveal the holdout. A vendor who tunes after seeing the answers is showing you a number you will never get in production.
  5. Report WMAPE, bias and FVA per segment (fast movers, slow movers, new products), because the averages hide where each tool fails.

How Long Does Implementation Take, and What Does It Cost?

Plan for 12 to 18 months to a first stable cycle in a single region, longer for a global rollout, with a systems integrator on the critical path. The 12 to 18 month figure is our judgement; none of the five vendors publishes a standard timeline, and the one concrete public figure found in this research, the three-plus-year Kinaxis pharmaceutical rollout above, comes from the vendor's own case study. Use the range to stress-test the integrator's plan, and ask every reference customer how long it took to reach a cycle the planners trusted.

The cautionary tale every supply chain VP should know is Nike's. Problems in its new i2 demand planning system in 2000, which became public in February 2001, ordered thousands more of some shoes than the market wanted and thousands fewer Air Jordans than it needed, and the glitch cost Nike more than $100 million in lost sales.

What Is Public on Pricing

As of November 20, 2026, none of the five publishes a list price for a full deployment.

SAP IBP is the only one that publishes its metric. SAP's pricing page bills the base plan in blocks of annual cost of goods (350 million per block), with prices on request, and meters add-ons for demand, supply and inventory optimisation in blocks of annual revenue. Kinaxis, o9, Blue Yonder and RELEX are contact-sales only.

The best public signal of deal size is Kinaxis' own disclosure: in 2025 it signed more than 20 deals above $1 million in average annual software contract value, and ended the year with annual recurring revenue of $433 million.

Third-party sites quote ranges for o9 and Blue Yonder. Most of them sell contract benchmarking and none shows its data, so do not budget off them. Budget for services on top of the subscription, and ask each integrator to price the data cleansing work as its own line item.


When a Time-Series Foundation Model Beats Buying a Suite

A time-series foundation model on your own warehouse beats a suite when the forecast itself is the problem and your planning workflow already works. If your planners run a functioning S&OP process in the ERP and spreadsheets, and the complaint is that the statistical forecast is bad, a better forecast model may be all you need.

The options are cheap to test. Google's BigQuery AI.FORECAST function runs TimesFM 2.5, with TimesFM 3.0 in preview in SQL and bills at the BigQuery ML prediction rate. Amazon's Chronos-2 is Apache 2.0 licensed, has 120 million parameters and handles covariates such as promotions and price. Our TimesFM vs Chronos vs Moirai comparison covers the licences and running costs. On Databricks Mosaic AI or Snowflake Cortex AI, a data engineer can score a model like these against your history in about a week.

Buy the suite when the pain is constrained supply planning, multi-echelon inventory, or getting sales, operations and finance to sign one plan. A foundation model produces a forecast. It does no capacity-constrained supply planning, and it will not run your consensus meeting.


The Decision: What Predicts Regret

The criteria that predict regret are the ones a demo does not show. In order:

  1. ERP fit. On S/4HANA, the bar for not choosing IBP is high. Off SAP, IBP's main advantage disappears.
  2. Where your constraint lives. Factory and supplier constraints point to Kinaxis or o9. Shelf and store constraints point to RELEX, or to Blue Yonder if you also want its execution software.
  3. Your own bake-off result, by segment, against a seasonal naive forecast and a foundation model baseline.
  4. Integrator depth for that product in your industry and region. Ask for named consultants, and ignore the firm's logo slide.
  5. Who owns the data clean-up. If the answer is "the vendor," the timeline is fiction.

Two things change the answer. In our judgement, a company under about $500 million in revenue will find all five more platform than it needs, and a foundation model plus a lighter planning tool gets most of the forecast gain. If you are midway through an S/4HANA migration, let it land before you choose a planner, because your master data will change under you.

This Week: Pull three years of weekly demand at the level you plan, and compute WMAPE, bias and FVA for your current forecast against a seasonal naive. That is your baseline, and most vendors will ask for it.

This Month: Run a TimesFM or Chronos-2 forecast on the same extract in your warehouse. If it beats your current number by a wide margin, decide whether you are buying a forecast or a planning process.

Before Your Next Budget Cycle: Send the extract and the holdout rules to two or three vendors from the table above, chosen by ERP and constraint, and require frozen configurations and per-segment results before any commercial conversation.

Continue Reading

Share:

Frequently Asked Questions

What is the best AI demand planning software in 2026?

It depends on two facts you already know. Kinaxis Maestro fits multi-site manufacturers off SAP who replan constantly, SAP IBP fits companies on S/4HANA, RELEX fits grocery and retail, and o9 fits companies that want demand, supply, commercial and financial plans on one model. Blue Yonder is strongest when you also buy its warehouse or transport software.

How much does SAP IBP cost?

SAP does not publish a dollar price. As of November 20, 2026, SAP's IBP pricing page meters the base plan in blocks of annual cost of goods (350 million per block), with prices on request. Kinaxis, o9, Blue Yonder and RELEX are contact-sales only.

How do I compare forecast accuracy between demand planning vendors?

Give every vendor the same three-year extract, hold back the latest 26 weeks, and score each at the level and lag you plan at using WMAPE, bias and forecast value added against a seasonal naive baseline. Freeze configurations before revealing the holdout, because vendor headline accuracy figures use different data, levels and lags and cannot be compared.

What do generative AI agents do in demand planning tools?

Mostly they explain and check the plan. A partner-built IBP demand planning agent for SAP Joule checks whether sales history is ready before a run. The forecast itself still comes from statistical and machine learning models underneath, and fully autonomous agents that move orders are mostly still announcements.

When is a time-series foundation model better than a demand planning suite?

When the forecast itself is the problem and your S&OP process already works. TimesFM runs in BigQuery through AI.FORECAST and Chronos-2 is Apache 2.0 licensed, so you can test either on your history quickly. Buy a suite when the pain is constrained supply planning, multi-echelon inventory or cross-functional consensus.

Newsletter

More on Models & LLMs, every Sunday

The week's enterprise AI news, Models & LLMs included, in one Sunday email.

Weekly, free. Unsubscribe any time. Privacy