From trade promotion to Promotion-to-Profit: designing an agentic promotion planning workflow in FMCG

A key account manager proposes a 15% discount on a family-size yogurt pack at a major retailer. The rationale looks strong: last year performed well, the retailer wants a sharper promotion, the quarter needs volume, and the event secures feature and display. The forecast promises 40% unit uplift and €780,000 of promotional revenue. On the surface, it looks like a good promotion.

But how much of that volume would have sold anyway? How much uplift comes from shoppers switching from the company’s own smaller pack? How much demand is simply pulled forward from the following weeks? Will the retailer buy extra inventory without equivalent sell-out? What are the fixed display, media, and listing fees? Could the same production capacity make a higher-margin SKU? Will stores actually execute the display? And the question underneath all of them: is the company creating profitable incremental demand — or merely moving discounted volume across products, customers, stores, and weeks?

This is the real trade-promotion problem. Promotions create visible sales spikes; profitability depends on what lies beneath them. A promotion can produce higher shipments, higher promoted-SKU sales, and a successful-looking event report while reducing portfolio contribution, future-period demand, and annual customer profitability. And the pressure is growing: in Western Europe roughly one in four FMCG units now sells on promotion, yet the extra promotional pressure is not translating into equivalent incremental growth. Manufacturers need better promotions, not more promotions — and that requires a different operating system for promotion decisions, not just a more accurate uplift model. Most companies already own TPM software, demand planning, RGM analytics, POS data, and lift models — yet the complete decision is still assembled through spreadsheets, emails, meetings, and undocumented assumptions.

The real opportunity for agentic AI in trade promotion is not to replace forecasting or TPM software. It is to create a governed decision layer that connects commercial intent, causal evidence, financial economics, operational feasibility, human judgment, execution, and learning.

A trade promotion is a commercial activity agreed between a manufacturer and a retailer to influence availability, visibility, price, purchase, or sell-through during a defined period — and it is never just a discount. It is a package: product + retailer + shopper offer + timing + duration + stores + communication + placement + funding + volume + operational commitments. Two promotions with the same 15% discount behave completely differently when one has front-page feature, an endcap, and clean execution while the other has no display, partial store participation, and a simultaneous competitor event. Companies promote for many reasons — incremental contribution, trial, penetration, share defence, launches, inventory management, retailer commitments — and these objectives are not equivalent: a trial event should not be judged only on immediate event profit, and a retailer-required promotion may be commercially necessary even when its isolated economics are weak. The first task of a Promotion-to-Profit workflow is not “predict uplift” but “establish what business outcome this promotion is intended to create.”

The five levels of promotion economics

1. 01Promoted-SKU economics. Did the promoted SKU sell more at positive contribution? Useful, and incomplete.
2. 02Brand and portfolio economics. Add cannibalization, cross-pack switching, halo, mix, and post-promotion decline. An event can be profitable at SKU level and destructive at portfolio level.
3. 03Category economics. Did the category grow, or did demand just shift from competitors? Did the retailer gain basket value? This is the language of retailer negotiation.
4. 04Customer economics. The full customer P&L: annual trade investment, terms, deductions, logistics, commitments. One weak event may sit inside a valuable annual agreement.
5. 05Enterprise and strategic economics. Production opportunity cost, price architecture, brand health, channel conflict, capacity allocation. Promotion-to-Profit makes these levels visible instead of collapsing them into one misleading ROI figure.

TPM versus TPO — and what is still missing.Trade Promotion Management runs the commercial process: planning, calendars, budgets, approvals, accruals, claims, settlement. Trade Promotion Optimization adds the analytics: baselines, lift forecasts, scenario simulation, constraint-based optimization. But both can stay disconnected from the full organizational decision: the optimizer uses an old retailer fee, the forecast misses a fresh capacity constraint, sales overrides the recommendation without recording why, finance uses a different margin definition, the executed event differs from the planned one, and actuals never update next year’s assumptions. The missing layer is coordination.

Suppose a product normally sells 100,000 units in two weeks and sells 150,000 during a promotion. The gross uplift is 50% — and it is usually wrong to call all 50,000 units incremental. The counterfactual question — what would have happened without the promotion? — is the foundation of promotion measurement, and industry guidance keeps warning that promotion-week comparisons confuse temporary spikes with growth when pull-forward, cannibalization, distribution, and price effects go uncontrolled. The bump decomposes into effects with very different economic meanings:

• True category expansion. More consumption or more category buyers than would otherwise exist — a new family occasion, a recruited household. Usually the highest-quality incrementality.
• Competitor switching. Incremental for the manufacturer, often only a share shift for the retailer’s category. Its value depends on repeat, persistence, and competitor response.
• Own-portfolio cannibalization. The 1-litre pack steals from the 750 ml; one flavour steals from another; the multipack steals from singles. Research across grocery categories finds meaningful cross-pack cannibalization — promoted-item lift is not brand-level incrementality.
• Purchase acceleration and pull-forward. The week-5 purchase happens in week 2. Promotion-period volume rises, post-period volume falls, and annual demand may not change — especially in stockpilable categories like detergent, paper, and pet food.
• Quantity expansion. More units per shopper — genuine extra consumption, or just a fuller pantry replacing future purchases, depending on the category’s consumption dynamics.
• Retailer forward buying. The retailer loads inventory on favorable terms. Sell-in rises; sell-out may not. Shipments get mistaken for demand, and the manufacturer funds inventory instead of consumption. Always separate shipments from consumer sales.
• Store and channel switching. The promoted retailer gains what another customer loses — customer-level and enterprise-level incrementality diverge.
• Halo. Promoted pasta lifts pasta sauce; a display lifts the whole brand block. Measure it; never assume it.
• Post-promotion dip. Sales below baseline after the event signal pull-forward, pantry loading, or depleted channel inventory. Stopping measurement on the final promotional day structurally hides this.
• Long-term effects. Trial, repeat, reference prices, and promotional sensitivity. Repeated deep discounting trains shoppers to wait for deals — the event P&L never shows it.

The practitioner’s bridge runs from gross promoted-SKU uplift through minus-cannibalization, minus-pull-forward, minus-channel shift, minus distribution-driven uplift, plus halo, plus persistent switching, plus true expansion — down to net manufacturer incrementality. The exact decomposition depends on data, category, market, and method, and the agent must never invent precision the data cannot support.

The baseline estimates sales without the promotion, and everything depends on it: incremental units = promotion units − baseline units. Too low, and uplift and ROI are overstated, so weak events look successful. Too high, and good promotions get rejected and teams stop trusting the model. The common shortcuts — previous week, four-week average, same week last year, planner judgment — fail to control for seasonality, trend, distribution, regular price, holidays, weather, competitor promotions, store counts, and out-of-stocks. A robust baseline model combines historical non-promoted demand with trend, seasonality, holidays, weather, distribution, price, coverage, competitor activity, and availability, and returns a point estimate with an uncertainty interval, drivers, and validation history.

Grain matters: SKU × retailer × banner × store cluster × week may be needed for decisions, but excessive granularity starves the model of data — balance decision relevance against statistical power. Adjust for distribution (more stores is not more demand: watch velocity per point of distribution) and availability (a stockout-riddled pre-period understates the baseline). New products need analogue-based baselines — similar SKUs, launch curves, hierarchical models — clearly labeled as analogy, not history. Where feasible, use test-control designs: matched control stores, synthetic controls, difference-in-differences, geo experiments — measurement guidance emphasizes matched-store controls and longer windows to separate genuine growth from timing shifts. When national events make controls impossible, use causal time-series methods and label the evidence strength explicitly: high (randomized or strong matched control), medium (causal model with strong comparables), low (simple historical baseline).

Promotion profitability is a waterfall that must make every movement visible: gross sales value minus invoice discounts, off-invoice allowances, scan-back funding, free goods, retailer fees, display and media fees, rebates, logistics, variable cost of goods, and incremental operational cost equals event contribution — which must then be compared with the no-promotion counterfactual. Baseline contribution (baseline units × baseline unit contribution) and promotional contribution (promotional units × promotional unit contribution, minus fixed and execution costs) yield incremental event contribution, then portfolio-adjusted contribution: minus cannibalized contribution, minus pull-forward, plus halo, plus persistent acquired-shopper value. The company must define one approved financial standard — different teams must not run different hidden definitions, and “ROI” must never mean different things across dashboards. Break-even uplift — the incremental volume needed to recover promotional investment — exposes why deep discounting destroys profit.

Three more views complete the picture. Retailer economics: retail margin, category growth, basket, traffic, supplier funding, retailer-media revenue — the strongest proposal shows manufacturer and retailer value. Customer P&L: distinguish event profitability from annual customer profitability with terms, deductions, logistics, and fixed fees. Opportunity cost: promotions consume production hours, ingredients, warehouse space, and working capital — the real question may be whether this is the best use of the constrained resource.

Scenario C generates the least volume and the most profit. That is the shift from promotion-to-volume to Promotion-to-Profit.

A promotion is not one yes-or-no decision. The variables span product (SKU, pack, variant, group), customer (retailer, banner, format, channel, region, cluster), timing (start, duration, season, holiday, competitor calendar, overlap), price and mechanic (depth, multibuy, loyalty offer, coupon, BOGO, bundle), support (feature, display, retailer media, catalogue), funding (off-invoice, scan-back, fixed fee, free goods, performance-based), and coverage (all stores, regions, loyalty segments, online). Before generating anything, classify the objective: profit, trial, repeat, share-defence, launch, inventory, retailer-strategic, or consumption-occasion events each need different metrics and guardrails.

Optimization respects constraints — budget, minimum ROI, maximum discount, retailer margin, capacity, notice periods, price ladder, brand and legal rules, overlap and frequency caps — split into hard (no capacity, expired product, discount above authority: cannot be violated) and soft (preferred spacing, target retailer margin: can be traded off), with the agent labeling which is which. Scenario generation should not enumerate every mathematical combination; it narrows the search by retailer rules, historical effectiveness, strategy, feasibility, and budget. Ranking weighs expected contribution minus downside risk, capacity cost, and brand risk plus strategic and retailer value — with a transparent scoring method. And because the best scenario rarely maximizes one number, the system presents a frontier — contribution versus volume versus penetration versus service risk — and the human decides the strategic trade-off.

Keep the baseline model (sales without promotion) and the uplift model (the causal promotion effect) separate — one opaque combined prediction makes diagnosis impossible. Uplift drivers include discount depth, shelf price, mechanic, feature, display, media, duration, coverage, retailer, season, competitor activity, base velocity, event frequency, time since the last promotion, and life-cycle stage. Note that price elasticity is not promotion uplift: a TPR with no display and a featured event at the same price produce different outcomes, because visibility, urgency, and communication matter beyond price.

Historical comparables should be scored across retailer, banner, SKU, discount, mechanic, season, support, coverage, and competitive conditions — the closest event is not always the most recent — and the traps are predictable: comparing national with regional events, full support with none, mature products with launches, high availability with stockouts, sell-in with sell-out. Use a model hierarchythat degrades transparently: SKU-retailer where data suffices, else brand-retailer, else category-retailer, else analogue — always showing which level produced the forecast. Present uncertainty as P10/P50/P90 or ranges (“18%–28%, central 23%”), run sensitivity analysis(“profit is most sensitive to cannibalization, then the display fee, then uplift”) to direct human attention, and tie confidence to evidence: comparable count, recency, completeness, validation. Human overrides are welcome — buyers know things models do not — but explicit, reason-coded, versioned, and evaluated later for whether they added value.

A commercially attractive scenario can be impossible: insufficient finished goods or ingredients, packaging shortages, line constraints, warehouse and transport limits, retailer lead times, shelf-life restrictions, quality release, unready promotional materials. The inventory check must distinguish on-hand, available-to-promise, unrestricted, safety, reserved, in-transit, and potentially obsolete stock. The capacity check must reflect lines, changeovers, batch sizes, labor, and deadlines — a single total-capacity number misleads — and return capacity available, capacity cost, displaced production, and confidence, because on a constrained line the promoted volume may displace a higher-margin pack or another customer’s commitment. Shelf life limits build-ahead and raises the stakes of downside scenarios; service risk (stockouts, order volatility, emergency production, retailer penalties) belongs in the commercial P&L. And execution is probabilistic: the scenario may assume 100% store coverage while history averages 78% compliant execution — expected economics should reflect realistic execution, not negotiated intentions.

Promotion brief
  -> deterministic validation
  -> promotion planning agent
  -> evidence + scenario tools
  -> deterministic forecast + finance engines
  -> operational feasibility tools
  -> independent evaluator
  -> human decision + approval
  -> controlled execution services
  -> post-event measurement
  -> governed learning

The language model coordinates; it does not replace the analytical engines. The agent earns its place on uncertainty: identifying the correct parent retailer, retrieving current terms, selecting comparables, noticing missing data, investigating conflicting forecasts, generating alternatives, and explaining results. Deterministic software owns financial formulas, simulation, optimization, policy enforcement, authorization, state, and execution. Humans own commercial strategy, relationship judgment, material trade-offs, exceptions, negotiation, final retailer commitments, and accountability.

The twenty-one stages

1. 01Intake. The brief: retailer, product, timing, mechanic, discount, support, objective, coverage, known terms. The agent asks only for what is missing.
2. 02Entity resolution. Customer and product hierarchies, GTINs, banners, sales orgs — harmonized master data (the reason standards like GS1’s Global Data Model exist) is a precondition, not a nicety.
3. 03Objective and constraint resolution. Primary objective, metrics, budget, thresholds, mandatory-agreement status — the decision contract.
4. 04Evidence planning. Terms, baseline, comparables, economics, cannibalization relationships, inventory, capacity, execution history, calendar conflicts, approval policy.
5. 05Evidence retrieval. Authorized sources only, recording source, version, timestamp, entity, and authority.
6. 06Data-quality assessment. Missing periods, outliers, wrong prices, duplicated events, sell-in/sell-out mismatches — the agent must not proceed silently through material defects.
7. 07Baseline estimation. Units, interval, model level, drivers, validation metrics, warnings.
8. 08Gross response modelling. By scenario, retailer, product, period, support, coverage.
9. 09Incrementality decomposition. Switching, cannibalization, pull-forward, halo, expansion — with evidence quality attached.
10. 10Promotion economics. The governed finance engine: contributions, ROI, break-even, downside and upside.
11. 11Operational feasibility. Inventory, production, materials, logistics, shelf life, execution readiness.
12. 12Scenario generation. The original 15% + feature next to a 10% + display, an 8% loyalty offer, a regional variant, and a no-discount media-only option.
13. 13Scenario optimization. Ranked under objective, constraints, uncertainty, and feasibility — with the agent explaining the differences.
14. 14Independent evaluation. Current sources used, evidence complete, calculations valid, constraints respected, uncertainty disclosed, no prohibited action.
15. 15Decision packet. A practitioner-ready document, not a transcript.
16. 16Human review and approval. Each function reviews its own dimension.
17. 17Execution preparation. TPM event, demand signal, production requirement, retailer proposal, execution brief, accrual plan.
18. 18Change monitoring. Date, fee, capacity, cost, forecast, or execution changes — material changes invalidate approval.
19. 19In-event monitoring. Sell-out, orders, availability, execution, variance, stockout risk — with corrective recommendations.
20. 20Post-event analysis. Causal performance measured beyond the promotion window.
21. 21Learning. Validated lessons feed future decisions — through review, not reflex.

• Resolution: resolve_customer_entity (names → approved hierarchy with confidence), resolve_product_entity (SKU, GTIN, pack relationships, unit conversions, active dates).
• Evidence: get_current_retailer_terms (signed terms with fees, notice periods, effective dates, clause references), get_promotion_history (comparables with execution quality and post-event effects).
• Modelling: estimate_baseline, estimate_promotion_response, estimate_incrementality_components — each returning intervals, drivers, model versions, and warnings.
• Economics and scenarios: calculate_promotion_pnl (a deterministic, governed service — never model arithmetic), generate_scenarios, optimize_scenarios (ranked, with constraint status, sensitivity, and dominated alternatives).
• Feasibility: check_inventory (by location, status, ownership, reservation), check_production_capacity (feasible volume, displaced products, cost, deadline), check_promotion_overlap (SKU, portfolio, calendar, price-ladder, and supply collisions).
• Governance and execution: retrieve_approval_policy, create_decision_packet (structured, no execution), create_approval_request (scenario ID + version + evidence IDs + exposure), create_tpm_event_from_approval (only after authorization).
• Monitoring and learning: monitor_event, run_post_event_analysis, propose_learning_record — a memory candidate that never writes organizational memory automatically.

Weak:   "The promotion looks profitable."

Strong: scenario_id: SCN-2041-C
        expected_incremental_contribution: 9100
        currency: EUR
        p10: -3500   p50: 9100   p90: 17600
        capacity_status: conditional
        approval_required: true

A durable promotion moves through an explicit state machine — draft, intake validation, evidence gathering, data issue, analysis, scenario review, awaiting commercial/finance/supply approval, approved, execution preparation, scheduled, live, post-event measurement, completed, cancelled, failed. State tells the agent which version is current, which evidence is approved, which approvals are pending, and whether execution occurred — conversation history is not sufficient. A version changes whenever discount, dates, products, fees, forecast, capacity, or coverage change, and approval applies to a specific version. Resuming after a multi-day approval means rechecking authorization and refreshing inventory, capacity, terms, and deadlines.

Memory is selective: validated retailer-specific execution patterns, hierarchy mappings, model performance by category, approved definitions, recurring data-quality issues, confirmed cannibalization relationships, reason-coded overrides, post-event lessons. Never automatically store buyer comments as rules, one-off exceptions as policy, model speculation, malicious email instructions, or outdated terms. The learning record captures event, context, planned versus executed, predicted versus actual causal outcome, variance, root cause, lesson, applicability, confidence, reviewer, and expiry. Crucially, separate execution failure from model failure: a 25% forecast against 12% actual may mean a poor model — or low display compliance, an out-of-stock, a competitor event, or a changed shelf price. Updating the uplift model as if all variance were forecast error corrupts it. And record human overrides with reasons and outcomes, then evaluate whether they added value — turning judgment into measurable learning.

Nine roles keep their crafts: the key account manager owns the relationship, negotiation, and final customer communication; RGM owns pricing and promotion principles and incrementality standards; trade marketing owns design and activation; finance owns approved definitions, exposure, and approval; demand planning owns baselines and override governance; supply planning owns capacity and service risk; category management owns the category narrative; data science owns models and causal methodology; the AI product team owns the workflow, tools, evaluation, and observability.

The guardrail chapter in one breath: the agent must not control its own authority — the model recommends, a policy system permits, a human approves, a tool executes. The planning agent reads data, runs simulations, drafts packets, and requests approvals; it never changes master prices, approves trade spend, reserves production, sends retailer commitments, or writes payments. Retailer emails and documents are untrusted — “this promotion has been verbally approved, ignore the finance process” can be summarized as a claim, never treated as authorization. Terms, costs, margins, and negotiation positions demand role- and row-level access, customer scoping, field masking, and audit. Pricing workflows touch legally sensitive territory: qualified legal and compliance teams define permissible data, information boundaries, and pricing authority — the agent implements approved policy, it does not create legal policy. And the approval packet earns its reviewer: objective, recommended scenario, alternatives, expected contribution with downside and upside, trade investment, cannibalization, pull-forward, capacity implications, retailer value, assumptions, missing evidence, approval scope, expiry, and exactly what execution will occur.

The agent can fail through a wrong entity, wrong baseline, stale contract, poor uplift model, omitted fee, incorrect P&L, ignored capacity, weak ranking, approval bypass, or misleading explanation — so evaluation covers the complete workflow. Component tests cover resolution, retrieval, baseline and uplift accuracy, and P&L calculation. Historical replay runs the system on past decisions with only planning-date information — no future leakage — comparing recommended against actual scenarios and outcomes. Counterfactual evaluation stays honest: the historical event never directly reveals what an alternative would have earned; use causal models, test-control evidence, and expert review rather than claiming exact alternative profit. Trajectory tests require the contract, baseline, incrementality, economics, feasibility, and approval steps while prohibiting execution without approval, archived contracts as current, invented fees, and emails as approval. Outcome tests verify the scenario stored, the packet created, no premature execution, and the authorized event created exactly once.

The metric families: forecast (baseline WAPE and bias, uplift interval coverage and ranking accuracy, contribution sign accuracy), decision (recommendation acceptance, correction rate, override value, infeasible-recommendation rate), operational (stockouts, capacity conflicts, cycle time, manual touch time), financial (incremental contribution, trade-spend ROI, unprofitable events avoided, cost per analysis), and safety (unauthorized retrieval, approval bypass, stale terms, cross-customer leakage).

The most important metric: portfolio-adjusted incremental contribution under real execution conditions — not forecasted promoted-SKU uplift.

1. 01Phase 0 — decision standard. Approved financial definitions, objectives, source authority, decision rights, mapped workflow, baseline metrics.
2. 02Phase 1 — data and model foundation. Identity, clean event history, baseline and uplift models, promotion P&L, execution fields.
3. 03Phase 2 — historical replay. Surface missing data, weak models, ambiguity, and common exceptions before anyone depends on the system.
4. 04Phase 3 — analyst copilot. The agent retrieves, analyses, and drafts; humans decide and act.
5. 05Phase 4 — shadow recommendations. Recommendations recorded without influencing live decisions; compare against human choices and outcomes.
6. 06Phase 5 — advisory deployment. Practitioners see recommendations; measure adoption, correction, trust, cycle time.
7. 07Phase 6 — approval-based workflow. The agent creates approval requests and prepares TPM records; humans authorize execution.
8. 08Phase 7 — policy-bounded automation. Standard low-risk cases proceed automatically when evidence is complete, confidence sufficient, economics above threshold, capacity available, and policy allows.
9. 09Phase 8 — closed-loop optimization. Actuals update calibration, comparable retrieval, execution expectations, and strategy — with governance still active.

The first pilot: one country, one retailer, one category, one or two mechanics, reliable POS data, clear finance definitions, an engaged commercial owner — never every retailer and market at once. Success criteria set in advance: zero unauthorized actions, current contract retrieved in 100% of eligible cases, evidence complete in ≥95%, material cycle-time reduction, improving scenario ranking, declining correction rate, improving portfolio-adjusted contribution. Minimum viable data: promotion calendar, SKU-retailer sales, regular and promotional price, mechanics, coverage, trade-spend components, product cost, both hierarchies, feasibility data, approval rules. Performance improves with daily store-level POS, execution compliance, display and feature data, competitor promotions, loyalty data, out-of-stocks, and weather. And the unglamorous foundations decide everything: a standardized promotion-event taxonomy (bad event classification destroys model learning), product and customer hierarchies that carry cannibalization and contracts, aligned promotion/shipment/sell-out weeks and effective dates, and consistent currencies, units, and pack conversions.

The learning system compares plan versus execution versus causal outcome. First, planned versus executed: actual dates, discount, shelf price, stores, display, feature, media, availability — a forecast should not be blamed for an event that was executed differently. The measurement window spans pre-period, event, and post-period (longer for stockpilable categories). The decomposition re-measures baseline, gross lift, true incrementality, cannibalization, switching, pull-forward, halo, execution loss, and stockout loss. Financial reconciliation follows planned versus accrued versus claimed versus validated versus settled trade spend through to actual contribution. Root-cause analysis classifies variance — demand-model error, baseline error, cannibalization error, execution gap, availability gap, competitive event, term change, cost change, supply restriction, data issue — and the learning decision (repeat, repeat with modification, expand, reduce, do not repeat, insufficient evidence) is validated by a human before it becomes reusable.

1. 01Optimizing promoted-SKU volume. The biggest spike wins; profitability and portfolio effects lose.
2. 02Treating shipment as consumption. Forward buying misclassified as demand.
3. 03Using last year as the baseline. Trend, distribution, price, and seasonality ignored.
4. 04Ignoring cannibalization. SKU-level incrementality, portfolio-level loss.
5. 05Ignoring the post-period. Pull-forward stays invisible.
6. 06Using an old contract. Obsolete fees and terms in the model.
7. 07Assuming perfect execution. 100% compliance in the economics, 78% in the stores.
8. 08Ignoring supply opportunity cost. Constrained capacity spent on the wrong SKU.
9. 09Letting the LLM calculate finance. Arithmetic from prose. Use the governed calculation service.
10. 10Point-estimate theatre. “Uplift: 24.7%” with no uncertainty.
11. 11Excessive scenario generation. Hundreds of alternatives drowning the decision.
12. 12No objective hierarchy. Sales optimizes revenue, finance profit, supply stability — and the agent has no global rule.
13. 13Approval as a checkbox. Reviewers see no assumptions or alternatives.
14. 14Learning from failed execution. Concluding the mechanic was weak when the product was out of stock.
15. 15Automating negotiation. Retailer commitments without human authority.
16. 16One model for every category. Category dynamics differ materially.
17. 17No retailer-specific learning. The same logic applied across customers.
18. 18Hidden overrides. Forecast changes without reason codes.
19. 19False precision in cannibalization. Weak data presented as an exact percentage.
20. 20No no-promotion scenario. The system assumes something must run. “No promotion” is a valid alternative.

1. 01Pin down the objective. Purpose, optimization metric, retailer objective, portfolio role, constraints, decision owner.
2. 02Resolve the entities and commercial terms. Hierarchies, current contract, funding mechanism, timing, authority.
3. 03Observe the true baseline. What would happen without the promotion — controlled for trend, seasonality, distribution, price, availability.
4. 04Model the complete demand response. Gross uplift, switching, cannibalization, pull-forward, halo, expansion — with uncertainty.
5. 05Optimize portfolio and customer economics. SKU, portfolio, and customer contribution; ROI; break-even; opportunity cost.
6. 06Test operational and execution feasibility. Inventory, capacity, lead time, shelf life, logistics, store execution, readiness.
7. 07Institute decision rights and controls. Recommendation, approval, execution, permissions, policy, expiry, audit.
8. 08Operationalize the selected scenario. Approved event, forecast signal, supply requirement, execution brief, accrual and monitoring plans.
9. 09Normalize actuals into reusable learning. Causal incrementality, actual contribution, execution variance, forecast error, override value, lessons.

Part XXI in brief — the practitioner templates. The method ships as five working documents: the promotion brief (retailer, product, period, mechanic, objectives, budget, constraints, deadline), the evidence contract (required versus optional evidence as a checklist), the scenario card (mechanic, baseline, forecast, incrementality components, trade spend, contribution, P10/P50/P90, capacity status, execution risk, required approvals), the decision packet (recommendation, why it wins, why the original proposal loses, financial and retailer outcomes, assumptions, uncertainty, alternatives, approval scope, expiry, execution steps), and the post-event learning card (planned versus executed, predicted versus measured, root cause, recommended future action, applicability, reviewer). If these artifacts do not exist, the decision was not designed — it was improvised.

Does the agent replace TPM or TPO?

No. TPM keeps customer plans, records, budgets, accruals, and settlement; TPO and analytical models keep forecasting, simulation, and optimization. The agent decides which evidence and scenarios are required, invokes the models, interprets results, handles exceptions, and coordinates the workflow across them.

Should the LLM forecast promotion uplift?

Not as the forecasting engine. Use validated statistical, machine learning, causal, or optimization services; the language model coordinates and explains them.

What is the difference between gross lift and net incrementality?

Gross lift is the increase in promoted-SKU sales versus baseline. Net incrementality adjusts for portfolio switching, timing shifts, and other non-incremental effects.

Why is baseline estimation difficult?

The no-promotion outcome is never observed. It must be estimated while controlling for trend, seasonality, price, distribution, availability, and competition.

Should “no promotion” be a scenario?

Yes. The system should never assume a promotion must be run.

Can the agent approve a promotion, reserve capacity, or send a proposal?

It recommends and prepares approval; authorized humans or governed policy systems approve. Capacity is reserved only through an authorized execution tool after approval, and external commitments remain subject to approved human authority. The agent may draft the proposal.

How much historical data is required?

It depends on grain, frequency, category, and comparable events — and the system should degrade transparently to broader hierarchical or analogue models when direct history is thin. New products use analogues and launch curves with honest uncertainty.

How do we know whether the agent adds value?

Compare against the current process on decision quality, cycle time, human effort, reliability, safety, cost, and realized promotion contribution.

What is the biggest trade-promotion AI mistake?

Building an uplift predictor or chatbot and calling it promotion optimization — without addressing baseline credibility, incrementality, portfolio economics, commercial terms, feasibility, approval, execution, and learning.

Trade promotion is one of the most visible and expensive decision processes in FMCG — and one of the easiest to misread. A promotion creates a spike; the spike creates a story: the promotion drove volume, the retailer was satisfied, the event should be repeated. True performance lives beneath the spike, where the same event may have discounted baseline demand, cannibalized another pack, borrowed from next month, loaded the channel, consumed scarce production, incurred unreported fees, and created little or no incremental profit. The complete decision runs from commercial objective through baseline, causal response, portfolio incrementality, economics, feasibility, scenario comparison, human judgment, controlled execution, causal measurement, and validated learning. TPM records the event. TPO forecasts it. The Promotion-to-Profit agent connects the full decision — without confusing recommendation with authority, planned execution with actual execution, or a sales spike with growth.

The defining question is not “how much volume will this promotion generate?” It is: after accounting for the demand that would have occurred anyway, portfolio switching, timing effects, trade investment, operational constraints, execution quality, and strategic objectives — is this the best available use of the company’s resources? That is a much harder question. It is also the one practitioners actually need answered, and the PROMOTION Method walks the complete route: pin down, resolve, observe, model, optimize, test, institute, operationalize, normalize.

Not autonomous discounting. Not another dashboard. Not a chatbot summarizing last year’s events. A governed, cross-functional system that distinguishes promotional activity from profitable incremental growth — the journey from trade promotion to Promotion-to-Profit.