The 5 Classes of AI Agents in FMCG: a framework by Asseco Platform for AI that supports sales, retail and HoReCa execution teams. Data unification, recognition, scoring, context and next best action — used across 125,000+ users in 74 markets.

KI-Agenten-Framework von Asseco Platform: fünf spezialisierte Klassen von KI-Agenten — Datenvereinheitlichungs-Agent, Erkennungs-Agent, Bewertungs- und Priorisierungs-Agent, Kontext-Agent, Nächstbeste-Aktion-Agent — die zusammenarbeiten, um FMCG-Außendienst, Retail und HoReCa-Execution zu unterstützen.

The 5 Classes of AI Agents in FMCG: a framework for AI that equips humans.

A framework from Asseco Platform for how AI supports people — not replaces them — across sales, retail and HoReCa execution. Built on the foundation of 125,000+ users in 74 markets.

125,000+ Users across 74 markets on Asseco Platform
POI Best-in-Class 2025 including AI/Machine Learning & Advanced Imaging
ISO 27001 Certified for AI AI-based solutions explicitly in certification scope
Ein Agent für jeden Moment

Jeder Moment der FMCG-Execution hat einen Agenten dahinter.

Jeder Moment in der FMCG-Execution hat dieselbe Form: Daten sammeln, beurteilen, was wichtig ist, handeln. KI-Agenten in Asseco Platform stehen an jedem dieser Momente — dreizehn spezialisierte Agenten, bereit, wann immer jemand eingreift.

Wenn Distributordaten in fünf verschiedenen Formaten eintreffen

Distributor-Datenvereinheitlichungs-Agent

Wenn ein Außendienstmitarbeiter ein Regal fotografiert

Regalerkennungs-Agent

Wenn eine handschriftliche Bestellung von einem Handelspartner eintrifft

Bestellerkennungs-Agent

Wenn die Routen für morgen für das gesamte Team geplant werden müssen

Routenoptimierungs-Agent

Wenn ein Vertrag mit einer HoReCa-Kette zur Überprüfung ansteht

Vertragskonformitäts-Bewertungs-Agent

Wenn ein Mitarbeiter den nächsten Kunden aufsucht

Vorbesuchs-Briefing-Agent

Wenn ein Vertriebsleiter einen Verhandlungsraum betritt

Vorverhandlungs-Briefing-Agent

Wenn die nächste Bestellung die richtige Menge zum richtigen Zeitpunkt benötigt

Bestellempfehlungs-Agent

Wer auch immer in diesem Moment agiert – ein Außendienstmitarbeiter, ein Trade Marketing Manager, ein Vertriebsleiter, ein CTO, ein Category Manager, ein Feldmitarbeiter, ein Supply Chain Planer – ein Agent ist für diesen Moment da. Dreizehn spezialisierte Agenten, eine Kette von Rohdaten bis zur konkreten Aktion. Das Framework begegnet Ihnen dort, wo die Arbeit stattfindet.

Wie das Framework funktioniert

KI auf drei verschiedenen Ebenen der Autonomie.

Nicht jeder KI-Agent funktioniert auf dieselbe Weise. Jede der fünf Klassen hat ein unterschiedliches Autonomieniveau — das widerspiegelt, wo KI am besten allein arbeitet, wo sie mit Ihnen arbeitet und wo sie nur auf Anfrage arbeitet. Human-in-the-loop, dreifach.

01

Im Hintergrund

Klassen 1 und 2: Datenvereinheitlichung und Erkennung. Daten werden vereinheitlicht, Bilder erkannt, eingehende Bestellungen in strukturierte Daten umgewandelt — ohne dass jemand etwas tun muss. Eine unsichtbare Grundlage. Keine Benutzeroberfläche, keine Benutzereingabe — nur zuverlässige, standardisierte Daten, auf die alles andere aufbaut.

Hohe Autonomie — läuft ohne Benutzerinitiative

02

Nach Zeitplan, zur Analyse

Klasse 3: Bewertung & Priorisierung. Rankings, Scores und Routen werden regelmäßig generiert — jede Nacht, jeden Morgen, jeden Planungszyklus. Das Management überprüft sie, passt sie an und nutzt sie für Entscheidungen, die Hunderte von Geschäften auf einmal betreffen.

Hohe Autonomie — läuft nach Zeitplan oder auf Abruf

03

Im Moment, mit dem Menschen

Klassen 4 und 5: Kontext und Nächstbeste Aktion. Kontext vor dem Besuch. Empfehlung am Regal. Vorschlag während der Bestellung. Diese Agenten arbeiten in Echtzeit mit dem Benutzer — und der Benutzer entscheidet. KI bereitet den Boden. Der Mensch handelt.

Niedrige bis mittlere Autonomie — Mensch entscheidet

Dies ist keine Automatisierung um der Automatisierung willen. Es geht darum, Gewicht vom Menschen auf das System zu verlagern — dort, wo das System besser abschneidet: bei der Aggregation und Korrelation von Daten, deren Umfang und Geschwindigkeit die Wahrnehmung einer einzelnen Person übersteigen. Der Mensch wird nicht ersetzt. Der Mensch wird mit einem Werkzeug ausgestattet, das mehr weiß, weiter sieht und schneller vorschlägt.

The 5 Classes — one question each

Each class answers one question FMCG teams face every day.

Together, they form a chain from scattered data to concrete action — with thirteen specialized agents across five products.

Class 1 — Data Unification Agent: How do you turn scattered data into a single source of truth?

Works autonomously in the background, with no end-user involvement. Its job is to bring data from multiple sources into one consistent, analysis-ready dataset — mapping records, deduplicating entries, enriching profiles with external attributes. Engages a human only in exceptions that require business judgement. Autonomy: High.

In Asseco Platform:

  • Distributor Data Unification Agent — unifies sell-out data from distributors into one data model in Trade Data Hub.
  • Customer Profile Enrichment Agent — enriches customer profiles with external data (industry codes, licensing registries, tax records) in Trade Data Hub.
  • Menu Data Acquisition Agent — sources public menu data for HoReCa outlet prospecting in Menu Recognition (coming H2 2026).
Class 2 — Recognition Agent: How do you turn observation into structured data — without human interpretation?

Processes unstructured inputs — images, text, HTML — and converts them into structured, verifiable KPIs. Does not infer, recommend or judge. Delivers an objective, repeatable picture of reality, independent of subjective human assessment. Works in real time or near real time. Autonomy: High, or Medium when triggered by user action.

In Asseco Platform:

  • Shelf Recognition Agent — identifies SKUs, counts facings, detects gaps, reads prices from shelf photos in Retail Image Recognition.
  • Menu Recognition Agent — identifies brands, categories, prices and competitors from menu photos in Menu Recognition.
  • Order Recognition Agent — reads handwritten or image-based orders from trade partners in Trade Data Hub.
Class 3 — Scoring & Prioritization Agent: Which store, outlet or contract needs attention first — and why?

Combines data from multiple sources with business rules or ML models to generate a score, ranking or classification. Answers the “which one” question: which store, which outlet, which contract, which customer. Doesn’t collect data, doesn’t recognize images — it reasons and prioritizes on data already gathered. Autonomy: High.

In Asseco Platform:

  • Store Potential Scoring Agent — scores each store’s sales potential in Sales & Retail Execution.
  • Route Optimization Agent (DRO) — generates optimal route plans for the entire field team in Sales & Retail Execution. In production, service ratio rises from 58% to 71% — 13 percentage points more time with the customer, same working hours.
  • HoReCa Outlet Scoring Agent — scores unvisited HoReCa outlets for prospecting in Menu Recognition (coming H2 2026).
  • Contract Compliance Scoring Agent — aggregates contract-execution scores per outlet, region and brand in Menu Recognition.
Class 4 — Context Agent: What should I know before I walk into this conversation?

Runs just before or at the start of a human interaction. Assembles a situational picture from multiple operational systems the team already uses — visit history, active promotions, open tasks, goals, last audit — so the human walks into the moment already informed. The underlying data is already unified by the platform; this class selects and composes the slice that matters right now. Doesn’t tell what to do — prepares the ground for a better human decision. Autonomy: Low — triggered on demand.

In Asseco Platform:

  • Pre-Visit Briefing Agent — aggregates purchase history, last audit, active promotions, open tasks and goals before a visit in Sales & Retail Execution.
  • Pre-Negotiation Briefing Agent — aggregates field visit data into a contract-execution picture before a negotiation round in Menu Recognition.
Class 5 — Next Best Action Agent: What should I do right now — at this place, at this moment?

Turns the output of other agents — or direct data analysis — into a concrete recommendation: what to do, where, when and why. Delivers the recommendation at the right moment to the right person. Differs from a Context Agent: it doesn’t just inform — it suggests action. Differs from a Scoring Agent: it addresses a specific person with a specific step, not an abstract ranking. Autonomy: Medium — recommends, human decides.

In Asseco Platform:

  • In-store Recommendation Agent — generates immediate action recommendations after a shelf photo in Retail Image Recognition.
  • Guided Selling Agent — guides the rep through the visit: which products to present, which arguments to use, which actions to take based on store profile and current priorities. Deployed in Sales & Retail Execution (RAO feature). Recognized with the POI Best-in-Class distinction 2025.
  • Order Recommendation Agent — recommends optimal order quantities based on purchase history, seasonality and current stock — preventing out-of-stock — in Third Party Field Execution.

Deep-dive: Next Best Action Agent →

Customer Success

AI-driven suggestions, acknowledged by Kraft Heinz.

Industry: Global FMCG — food & grocery Products: Retail Image Recognition, Sales & Retail Execution AI agents deployed: Shelf Recognition + In-store Recommendation

“We find the Asseco Business Solutions AI-driven suggestions useful in terms of our business processes such as e.g. automating order generation in response to inventory shortages.”

— Adam Imielski, Managing Director Poland & Ukraine | Board Member at The Kraft Heinz Company

Global beverage leader — name under NDA

At scale — one of the largest AI deployments in FMCG.

A global beverage leader deploys AI agents across its full field organization — for store scoring, pre-visit briefing, shelf recognition, and in-store recommendations. The client plans to make Retail Image Recognition the central tool for settling global sales targets — a signal that AI-based execution measurement is moving from operational tool to management system.

52

countries

Mobile Touch with AI agents active

~3,700

field users

scoring, briefing and recommendation agents daily

9

IR countries

shelf recognition agents toward global target settlement

Das Framework in Aktion

Jedes Produkt führt Agenten aus einer oder mehreren Klassen aus.

Je umfangreicher die KI-Oberfläche eines Produkts ist, desto mehr Klassen aktiviert es. Sales & Retail Execution und Menu Recognition aktivieren jeweils vier der fünf Klassen — sie sind die KI-dichtesten Produkte der Plattform.

Produkt Klasse 1
Datenvereinheitlichung
Klasse 2
Erkennung
Klasse 3
Bewertung
Klasse 4
Kontext
Klasse 5
Nächstbeste Aktion
Sales & Retail Execution Geschäftspotenzial-Bewertung, Routenoptimierung Vorbesuchs-Briefing Guided Selling
Trade Data Hub Distributor-Datenvereinheitlichung, Kundenprofilanreicherung Bestellerkennung
Retail Image Recognition Regalerkennung Im-Geschäft-Empfehlung
Third Party Field Execution Bestellempfehlung
Menu Recognition (Seite kommt bald) Menüdaten-Erfassung Menüerkennung HoReCa-Outlet-Bewertung, Vertragskonformitäts-Bewertung Vorverhandlungs-Briefing

Einige Produkte (Trade Data Hub, TPFE, Retail Image Recognition) spezialisieren sich auf einen Teil der Kette — und tragen dort bei, wo ihre Domäne den größten Mehrwert bietet. Andere (Sales & Retail Execution, Menu Recognition) orchestrieren Agenten über mehrere Rollen hinweg End-to-End — von der Datenvorbereitung bis zur Empfehlung am Regal.

Industry Recognition

Recognized by analysts, validated by the industry.

The framework isn’t a concept document. The AI agents behind it are benchmarked by third parties — and certified against international standards that explicitly include AI in scope.

Gartner Representative Vendor

Asseco Platform was named Representative Vendor in the Gartner Market Guide for Retail Execution Management in FMCG (G00768127, April 2025) — confirming world-class capacity for multi-channel, multi-region FMCG deployment.

POI Best-in-Class 2025 — 9 distinctions

Nine Best-in-Class distinctions in the POI Consumer Goods Enterprise Planning & Retail Execution Vendor Panorama Report 2025, including AI/Machine Learning, Advanced Imaging Technology, Retail Activity Optimization, Guided Selling, and Data Management.

ISO/IEC 27001:2022 — AI-based in scope

Asseco Business Solutions is certified to ISO/IEC 27001:2022 — the international standard for information security management. AI-based solutions are explicitly included in the certification scope. Valid through February 2029 (Alcumus ISOQAR, 26713-ISMS-001).

Plattform-Ökosystem

Jeder KI-Agent läuft innerhalb einer integrierten Plattform.

Agenten existieren nicht isoliert. Sie laufen innerhalb von Asseco Platform-Produkten — mit demselben Datenrückgrat, derselben mobilen Execution-Ebene und derselben zertifizierten Sicherheitshülle.

Asseco Platform

Fünf Produkte. Dreizehn spezialisierte KI-Agenten. Eine integrierte Suite für FMCG-Execution.

Field Execution
Sales & Retail Execution Geschäftsbewertungs-, Routenoptimierungs-, Vorbesuchs-Briefing- und Guided-Selling-Agenten.
Retail Image Recognition Regalerkennungs- und Im-Geschäft-Empfehlungsagenten.
Menu Recognition Menüerfassungs-, Erkennungs-, Outlet-Bewertungs-, Vertragskonformitäts- und Vorverhandlungs-Briefing-Agenten.
Daten-Backbone
Trade Data Hub Distributor-Datenvereinheitlichungs-, Kundenprofilanreicherungs- und Bestellerkennungsagenten — die Datengrundlage, auf der jeder andere Agent aufbaut.
Ecosystem Extension
Third Party Field Execution Bestellempfehlungsagent für externe Außendienstpartner — Agenturen, Distributoren, Merchandising-Teams.
FAQ

Common questions about the AI Agents Framework

What are AI agents in FMCG?

AI agents are specialized software components that help humans execute sales, retail and HoReCa work — by unifying data, recognizing images, prioritizing decisions, assembling context and recommending actions. Asseco Platform groups them into five classes, each answering one specific question FMCG teams face every day.

Does an AI agent replace field sales reps?

No. Each class of AI agent in Asseco Platform has a clearly defined autonomy level — from fully autonomous (background data work) to collaborative (suggestions at the shelf). None of the classes make the final call with the customer. AI prepares the ground; the human decides and acts.

Can I start with one class of agents — or do I need all five?

You can start with one. The most common entry points are Class 2 (Recognition Agent — e.g. shelf recognition) and Class 5 (Next Best Action — e.g. guided selling). Each class delivers value on its own. Adopting more classes strengthens the chain from data to action, but isn’t a prerequisite.

How accurate are Recognition Agents?

The Shelf Recognition Agent operates at up to 98% accuracy across production deployments. Accuracy depends on training data, store environment and SKU range — but the production benchmark is well above what manual audits deliver.

What data do AI agents need to work?

Class 1 (Data Unification) needs source data: sell-out from distributors, customer records, product catalogues. For image recognition, a SKU reference library is required. For scoring and context agents, operational data (visits, orders, promotions, goals) is typically already in your SRE or ERP.

How long does it take to deploy AI agents in Asseco Platform?

It depends on the agent type and the scope. Shelf Recognition deployments typically go live in weeks. Data Unification projects (Trade Data Hub) are phased per partner — first batch in months, rolling expansion afterwards. Route Optimization (DRO) can deploy as a standalone service, independent of the full Asseco Platform.

Is my data safe when AI agents process it?

Asseco Business Solutions is certified to ISO/IEC 27001:2022 — the international standard for information security management — with AI-based solutions explicitly included in the certification scope. Data handling, retention and access control are governed by the same policies that cover the rest of the platform.

How is this different from a single AI assistant?

A single assistant answers whatever you ask. Our framework is the opposite: five specialized agents, each with one clear role, each working on specific data. This design trades generality for reliability. For a sales rep or a director, it matters that the tool always answers the right question — not that it answers any question imperfectly.

About the Authors

Who built this framework.

Marcin Grzegorczyk

Marcin Grzegorczyk

Director of Market Development & Sales Enablement — Asseco Business Solutions

Leads marketing, sales enablement and visibility strategy for Asseco Platform — the FMCG execution platform deployed across 74 markets with 125,000+ active users. Active on LinkedIn discussing AI, retail execution and field sales technology in FMCG.

LinkedIn →

Tomasz Puławski

Tomasz Puławski

Product Manager for AI & Analytics — Asseco Business Solutions

Owns the product roadmap for AI and analytical products in Asseco Platform, including Retail Image Recognition, Menu Recognition, and the broader AI agent layer across the platform. Author of the original framework.

LinkedIn →

Ready to see the framework running on your data?

We’ll walk you through the AI agents operating in your segment — retail, HoReCa or distributor channel — with real examples from comparable deployments. 30 minutes, tailored to your industry, your team size, and your current AI maturity.

125,000+ users across 74 markets · 9× POI Best-in-Class 2025 · ISO/IEC 27001:2022 certified with AI-based solutions in scope