AI use case

AI trade promotion optimization for consumer goods manufacturers

Machine learning and optimisation that plans and prices trade promotions, discounts, displays and rebates a consumer goods manufacturer runs with retailers, forecasting the volume each promotion mechanic would generate and searching the combinations of timing, depth and mechanic that best meet a revenue, profit or volume target, for a key account manager to review and agree with the retailer.

By Len Debets · Last verified 29 September 2026 · 2 public deployments

USD 7.5 million to USD 90 million
Indicative value per year
A consumer goods manufacturer with USD 1 billion in annual trade promotion spend. Worked example, see how it is calculated.

What problem does it solve?

Henkel's chief digital and information officer, Michael Nilles, describes trade promotion management as "still a complicated animal." Working with SAP, Henkel built a generative AI natural language layer over trade promotion data, a tool that gives key account managers plain language hints rather than one that forecasts lift and searches combinations itself, and Nilles says it made the process "much more intuitive for the key account managers." Independent trade press reporting on that deployment describes account managers going from a plan that took months to one ready to share by the next day.

PepsiCo's own account of building PromoAI describes a move away from spreadsheet based planning toward optimisation supported workflows, at a scale that puts the underlying problem in perspective: PepsiCo describes itself as a global leader in the consumer goods sector with annual revenues exceeding USD 90 billion, and a single promotional calendar for that portfolio can be chosen from millions of product, promotion and timing combinations, while still respecting retailer specific rules on margin, volume and promotional frequency. Manual planning cannot search that space; it can only check a handful of scenarios a planner already suspects will work.

How does it work?

  1. Forecast the response. A machine learning model forecasts the incremental volume, the lift above baseline sales, that each promotion mechanic would generate for a product, account and time period, learning from past promotions, price and competitor activity.
  2. Search the plan, not one scenario. An optimisation model, not the forecast alone, searches the combinations of timing, depth and mechanic across the whole account and calendar for the plan that best meets a chosen goal: manufacturer revenue, retailer revenue, margin or volume, within business rules such as minimum margin or a promotional frequency cap.
  3. Let the planner set the goal. Key account managers or revenue management teams choose what to prioritise for a given account or period, and can adjust the weighting between objectives such as revenue and margin rather than accepting a single fixed answer.
  4. Review before it goes to the retailer. Planners review the recommended calendar, can override individual promotions, and only then take it into the negotiation with the retailer.
  5. Learn from what happened. Actual sell through after each promotion feeds back into the forecasting model, so the next planning cycle starts from a better estimate of lift.
Audience
Employee facing
Autonomy
Copilot
Adoption
Early adopters
Channels
Internal tools

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

No public deployment has disclosed a measurable outcome yet.

Value drivers: Revenue growth, Lower cost to serve, Employee productivity.

Indicative value

A consumer goods manufacturer with USD 1 billion in annual trade promotion spend

USD 7.5 million to USD 90 million

Trade spend efficiency recovered per year per year

How this is calculated

Formula: tradeSpend * ineffectiveShare * recoveredShare. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
Annual trade promotion spend tradeSpend, USD per year500,000,0001,500,000,000Editorial assumption for a large consumer goods manufacturer's annual trade promotion spend, replace with your own trade spend figure.
Share of that spend estimated to generate no incremental volume ineffectiveShare, fraction of trade spend0.10.2Editorial assumption, replace with your own promotion effectiveness analysis; this is not sourced from either evidence record on this page.
Share of that ineffective spend an optimised plan recovers recoveredShare, fraction of the ineffective spend0.150.3Editorial assumption, kept conservative because neither evidence record on this page discloses a comparable recovery figure.

What it leaves out: Entirely editorial: PepsiCo discloses that about 85% of PromoAI's optimised recommendations are accepted and executed, and that planning cycles for specific pricing and promotional tasks have compressed from weeks to minutes, but it explicitly withholds the underlying financial figures as commercially sensitive, and Henkel's own reporting gives no quantified figure at all. Every input here should be replaced with the manufacturer's own promotion effectiveness data before this number is used for anything but illustration. It also leaves out the cost of the platform and data integration and any change in retailer relationships from a more disciplined promotion plan.

Who already uses it?

2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

PepsiCo

United States · Manufacturing · 2026

ScaledGrade B

PepsiCo describes, in its own published account, two large scale optimisation systems built and deployed to support its revenue growth management: PromoAI, which couples machine learning promotional forecasts with a mixed integer linear programming model to search millions of product, promotion and timing combinations for the calendar that maximises PepsiCo and retailer revenue within business constraints, and PricingAI, which optimises base prices across the portfolio using Bayesian hierarchical models of price elasticity. The paper reports that planners and revenue management teams reviewed model outputs during validation before the systems were put into production use.

No outcome disclosed.

Henkel

Germany · Manufacturing · 2026

ProductionGrade C

Henkel, the German consumer goods manufacturer, worked with SAP's co innovation team to build a generative AI natural language layer for trade promotion management on top of the Just Ask feature of SAP Analytics Cloud. It gives key account managers plain language hints over trade promotion data rather than forecasting lift and searching combinations itself, so it is not an optimisation engine of the kind PromoAI is. Henkel's chief digital and information officer, Michael Nilles, says the tool made trade promotion planning more intuitive for the company's key account managers, and independent trade press reporting describes account managers producing a finished plan by the next day instead of the months it used to take.

No outcome disclosed.

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • Promotion history per product, account and mechanic, with actual lift where it is known
  • Trade term and margin rules per retailer
  • Base price, cost and competitor activity data

Systems to integrate

  • Trade promotion management or revenue growth management system of record
  • Sales and account planning tools key account managers already use
  • Finance systems for trade spend accrual and settlement
  • Point of sale or shipment data for measuring actual promotion lift

Complexity: High

The forecasting model is the easier half. The harder part is the optimisation layer that respects each retailer's own margin, volume and frequency rules at the same time as the manufacturer's own revenue and profit targets, and integrating with the trade spend, sales and finance systems that already hold the numbers a key account manager is measured on.

  1. 1

    Start with the categories where promotion spend is largest

    Optimise the accounts and categories that carry the most trade spend first, where a percentage improvement in efficiency is worth the most in absolute terms.

  2. 2

    Build the forecast before the optimisation

    Get promotional lift forecasting accurate and trusted on its own before adding the optimisation layer on top of it; an optimiser built on a poor forecast just searches confidently for the wrong answer.

  3. 3

    Encode real retailer rules, not generic ones

    Load each retailer's actual margin floors, volume commitments and promotional frequency limits, agreed with the account team, so the recommended plan is one a planner can actually take into the room.

  4. 4

    Let planners adjust the objective, not just the output

    Give key account managers a way to reweight revenue, margin and volume for a specific account or period, so the system supports a negotiation instead of replacing the planner's judgement about that account.

  5. 5

    Close the loop with actual results

    Measure the lift each promotion actually delivered against the forecast, and feed that back into the model, or the forecast quietly drifts from reality.

Guardrails

  • Retailer specific margin, volume and compliance rules enforced as hard constraints, not suggestions
  • A planner reviews and can override any recommended promotion before it reaches a retailer
  • Version history on every promotion plan, so a planner can see what changed and why
  • Segregation between the system that recommends a plan and the system that pays out trade spend

KPIs to instrument

  • Forecast accuracy of promotional lift, before and after each planning cycle
  • Share of recommended plans planners accept without changes, and what they change when they do not
  • Trade spend efficiency, incremental volume per unit of trade spend
  • Time from the start of a planning cycle to an agreed calendar

Human in the loop

Key account managers and revenue management teams choose the objective, review every recommended plan, and take the final call on what goes to a retailer, especially in accounts where the relationship, not the model, decides the right approach. Finance reviews trade spend accruals against the plan actually executed, not the plan the model proposed.

Common failure modes

Optimising for revenue while margin erodes
A plan can hit a revenue target by promoting deeply and often, at the cost of margin. Set margin and frequency as hard constraints, not just one line on a scorecard.
A forecast that has not seen this year's conditions
A model trained on stable years overstates lift when input costs, competitor activity or the account relationship has shifted. Monitor forecast error by account, not only in aggregate.
Treating the recommended plan as final
A plan that goes to a retailer unreviewed misses account context the model cannot see. Keep a real review step, not a rubber stamp, especially in strategic accounts.
No segregation between planning and payout
If the same system both recommends promotions and approves the resulting trade spend payout, errors and manipulation are harder to catch. Keep planning and financial settlement in separate systems with their own controls.

What are the risks and rules?

EU AI Act

Minimal risk

The system optimises promotion terms between a manufacturer and a retailer; it does not decide a natural person's access to credit, employment, essential services or a regulated product, so it falls outside Annex III. It would need reassessment if a manufacturer used the same technique to set individually targeted prices or offers for identified consumers.

Rules that apply

Controls to put in place

  • Hard constraints for retailer margin, volume and compliance rules, checked before a plan is finalised
  • Human review of every plan before it is presented to a retailer
  • Separation of duties between the planning system and trade spend payout approval

Frequently asked questions

Does AI decide which promotions run?
In the deployments on this page, no. Henkel's chief digital and information officer describes the tool as making trade promotion "more intuitive" for key account managers, and PepsiCo's own paper describes PromoAI generating multiple optimised calendars under different objective settings and presenting them to the retailer "as a menu of analytically grounded options," with planners able to apply manual modifications before a calendar is finalised. For PepsiCo, the system narrows millions of combinations to a short list of calendars; Henkel's tool works differently, surfacing hints from the data rather than searching combinations itself. Either way, a person still decides.
What results have manufacturers reported?
PepsiCo's own paper reports that in most markets about 85% of PromoAI's optimised promotional recommendations were accepted and executed by the business, and that planning and execution cycles for specific pricing and promotional tasks compressed from weeks to minutes; it explicitly withholds the underlying financial figures as commercially sensitive. Henkel itself has not published a quantified result. Independent trade press reporting, not a Henkel quote, describes account managers producing a finished plan by the next day instead of the months it used to take.
Is this the same as dynamic pricing to individual shoppers?
No. Trade promotion optimisation sets the terms of a promotion a manufacturer runs with a retailer, at the product and account level, not a price shown to an individual shopper. A system that personalised prices to identified consumers would raise different questions and likely a different EU AI Act analysis.

How to cite this page

Blits.ai AI Use Case Library, "AI trade promotion optimization for consumer goods manufacturers", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/trade-promotion-optimization-for-cpg. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 29 September 2026: First published

Related use cases

Retail and ecommerce

AI demand forecasting and automated replenishment for retail

Machine learning that forecasts demand for every item in every store or fulfillment center, day by day, from sales history, promotions, prices, weather and local events, and turns the forecast into automatic store and warehouse orders within limits set by planners, who handle the exceptions.

Deployments
4 public, best grade B
Autonomy
Supervised agent
Retail and ecommerce

AI dynamic pricing and markdown optimization for retail

AI that recommends or automatically sets the regular price, promotion or markdown for every item in every store, from demand, cost, competitor prices, inventory and, for perishables, how close the item is to its best before date, so pricing analysts manage exceptions and strategy instead of spreadsheets.

Deployments
3 public, best grade B
Autonomy
Supervised agent
Cross industryManufacturing

AI quote and estimate generation from customer requirements

AI that turns what a customer sends, such as a product list, a drawing, a roof photo or a request for quotation, into a draft quote: it reads the input, matches items to the catalog, calculates quantities and applies the organization's price rules, and hands the draft to a seller or estimator who checks and sends it.

Deployments
3 public, best grade C
Reported handling time reduction
87.5%
Enpal, vendor claim
Telecommunications

AI assistant for B2B telecom quoting, sales and service

An AI assistant that serves business customers of a telecom operator and the sellers who look after them: it answers product, pricing and contract questions, prepares configurations and quotes for connectivity, mobile fleets and devices, drafts responses to tenders, and handles routine service requests and fault tickets, with a sales or service specialist approving anything binding.

Deployments
5 public, best grade B
Reported containment rate
70%
SoftBank Corp., organization claim