BASF is a German multinational and one of many world’s largest chemical firms, identified for its built-in Verbund manufacturing community, world scale, and broad portfolio spanning from fundamental chemical substances to superior agricultural options. With its sturdy basis in analysis and growth, BASF operates throughout numerous industries whereas repeatedly driving innovation and sustainability.
One among its key operational divisions is BASF Coatings, which makes a speciality of creating, manufacturing, and advertising superior automotive and industrial coatings, together with ornamental paints. As a pioneer in eco-efficient floor applied sciences, BASF Coatings can also be on the forefront of digital transformation, leveraging AI-powered platforms to boost productiveness, innovation, reliability, and design.
In partnership with Databricks, BASF Coatings has applied a production-ready, ruled, and business-impacting multi-agent answer. This method not solely enhances cross-team collaboration but in addition allows smarter, quicker decision-making throughout vital enterprise capabilities — setting a benchmark for the way superior analytics and AI can drive tangible enterprise outcomes.
The Problem: Carry extra Modularity, Specialization and Management to Agent Techniques
As a company with over 11,000 workers throughout greater than 70 websites worldwide, managing the rising complexity and bettering effectivity of cross-department digitalization is a non-trivial job. Extra particularly, turning huge, disparate organizational information into actionable insights, and enabling real-time decision-making and productiveness has grow to be the important thing. Fixing this downside mattered as a result of environment friendly digital collaboration and information utilization immediately have an effect on market responsiveness, innovation velocity, buyer satisfaction, and operational reliability. The stakes had been significantly excessive in industries like coatings, the place agility and precision are essential amid quickly altering buyer calls for and sustainability pressures.
An agentic system – the place autonomous or semi-autonomous AI brokers proactively handle enterprise processes and information flows – was the very best answer as a result of it may automate coordination and evaluation duties that beforehand required intensive guide effort. Agent programs may empower organisations like BASF Coatings to:
- Seamlessly combine AI throughout domains, automating routine operations in gross sales, procurement, and provide chain administration.
- Present sensible, contextual suggestions and automate choice flows, dramatically decreasing bottlenecks and errors.
- Enhance person expertise by enabling “always-on” chat assistants for help, Q&A, or workflow integration throughout departments.
- Drive adoption of on a regular basis AI instruments company-wide, making complicated digital capabilities accessible to enterprise stakeholders and fostering information literacy.
As highly effective as an agent might be, as we develop these programs, they could develop extra complicated over time, making them tougher to handle and scale. For instance, an agent can have too many instruments at its disposal and make poor selections about which software to name subsequent, additionally the context grows too complicated for a single agent to maintain observe of. There’s a want for a number of specialization areas within the system (e.g. supervisor, area orchestration, subject material professional, and many others.)
One other approach to view the problem is thru the variety of knowledge that kinds the agent system’s data base. Many people are already aware of RAG (Retrieval-Augmented Era), a way that mixes giant language fashions (LLMs) with real-time information retrieval to enhance response accuracy and relevance. Nonetheless, RAG programs are primarily designed to deal with unstructured information – Corresponding to paperwork, internet pages, PDFs, or different types of free textual content – somewhat than structured tables with predefined fields and relationships. When working with structured information, Textual content-to-SQL is the most typical method for pure language analytics. Nonetheless, it typically depends on pre-defined instance SQL queries and lacks built-in mechanisms for information governance and permission management.
The Resolution: An Finish-to-Finish Multi-Agent Supervisor for Structured and Unstructured Data
To handle these challenges, we suggest breaking our utility into a number of smaller, unbiased brokers and composing them right into a multi-agent system. This technique will comply with a supervisor sample that coordinates the specialist brokers – particularly, Genie brokers and function-calling brokers – which work together with the Databricks Vector Retailer Retrieval software.
AI/BI Genie, some of the fashionable options inside Databricks, is designed to make structured information akin to Delta tables and views immediately accessible to enterprise customers by leveraging pure language interfaces. It makes use of metadata from Unity Catalog, akin to desk descriptions, PK/FK relationships, and column names/descriptions. This metadata guides Genie in parsing person questions, developing correct SQL, and delivering contextually related solutions – serving to to mitigate errors or hallucinations. As well as, Genie authors can improve the house by domestically enhancing metadata, defining joins, including synonyms, and curating BASF-specific directions. This permits information stewards to actively handle and preserve the standard of their Genie areas thus contributing on to the agent system with their invaluable enterprise area data.
To ease using Genie inside agent orchestration frameworks, there are frameworks supporting devoted Python wrappers for constructing Genie brokers (examine right here for reference). As well as, Databricks product group options instance notebooks that stroll our customers by means of establishing a multi-agent system utilizing Mosaic AI Agent Framework along with Genie. These examples leverage LangGraph (an open-source agent orchestration library) and show how one can compose workflows the place Genie is one specialised agent amongst a number of.
An summary of our structure is as follows. We undertake Databricks’ Mosaic AI framework to simplify the complexities of managing AI agent lifecycles, providing instruments and fast multi-agent coordination prototyping, rigorous analysis, and efficient real-time operational monitoring. Notably, we additionally combine the deployed supervisor endpoint with Microsoft Groups for real-time agent execution, and make AI-powered insights available to all sorts of customers, together with enterprise stakeholders who’re much less aware of information platforms – by embedding conversational deployment endpoints immediately throughout the Groups interface. Clear, reusable accelerators exist for provisioning cloud assets (Azure Bot Service, App Service) and connecting endpoints to Groups.
Actual Enterprise Influence
Whereas BASF Coatings is creating AI brokers that may improve its enterprise processes, the primary touchdown zone challenge, Marketmind, focuses on the Gross sales & Advertising division. The use case allows superior quantitative and qualitative evaluation by consolidating inside Salesforce buyer go to experiences and market consumption insights with exterior market developments together with S&P 500 information. A few of this information is already processed and accessible within the type of Delta tables and views, whereas the remainder exists as free-text information and PDF paperwork, every arriving at completely different speeds and being up to date at various frequencies. Moreover, the info is managed by completely different groups and stewards. For instance, structured tables are primarily supplied by BASF’s central Enterprise Information Lake (EDL) group, with Gross sales & Advertising enterprise consultants enriching them with domain-specific metadata. In distinction, unstructured information is primarily processed by means of code-first ETL pipelines developed and maintained by the Coatings Information & AI workplace group.
Given the complexity of the info panorama, we adopted the multi-agent supervisor structure for the Marketmind challenge and used the template pocket book as our place to begin. We created a Genie house for structured information, enriching it with curated tables, detailed column descriptions, Genie-local be part of relationships, and worth sampling. To enhance accuracy, we added SQL examples and clear directions to information Genie’s responses, and we carried out common Benchmark exams as new information got here in to judge its general efficiency.
For unstructured information akin to Salesforce go to experiences and market information, we constructed vector search indices for every supply utilizing embeddings to allow context-aware similarity search. We then created Unity Catalog capabilities that wrap Mosaic AI Vector Search queries, guaranteeing enterprise-ready governance, discoverability, and automated MLflow tracing. Lastly, we developed a perform tool-calling agent that invokes vector retrieval instruments to deal with task-specific requests handed alongside by the supervisor.
Our Marketmind challenge started its scoping section in April this yr, adopted by a 5–6 week proof of idea (PoC). We then moved into the total implementation section, accompanied by technical upskilling workshops, structure opinions, and product and have discussions with the Databricks’ Mosaic AI product group. We carried out a one‑month pilot with 25 key customers, and at the moment are within the ultimate refinement stage earlier than go‑stay and rolling out to North America by the top of October . As soon as launched, greater than 1,000 gross sales representatives worldwide can be utilizing Marketmind, with inputs up to date continuously.
Marketmind is already altering how BASF Coatings’ gross sales groups put together, have interaction, and comply with up with their prospects. As an alternative of looking for leads by means of scattered notes and folders, gross sales representatives obtain personalised notifications alongside instructed actions and methods primarily based on present occasions out there. If additional data is required, Marketmind affords the choice to dig deeper into the underlying information and experiences utilizing an easy-to-use chat interface. The screenshot under illustrates this shift. Alerts from the market are offered in an actionable, conversational interface inside Microsoft Groups, so Coating’s gross sales group can shift their focus from “What occurred?” to “What ought to I do subsequent?” with out switching instruments.
As proven above, gross sales groups cannot solely ask ad-hoc inquiries to the Marketmind chatbot immediately in Groups, but in addition obtain proactive adaptive playing cards with the newest market developments on a weekly foundation. Customers can discover subjects of curiosity in larger element by clicking the connected URL, which redirects them to the unique information supply. To additional improve the agent’s high quality, we’ve got additionally built-in a voting mechanism that permits customers to shortly give a thumbs up or down, or present extra detailed written suggestions within the backside discipline. This suggestions is captured within the mannequin inference desk and built-in with the present payload information.
“Marketmind turns our discipline interactions into well timed, AI-driven actions—nudging sensible follow-ups, surfacing related alternatives, and connecting friends going through related challenges. The consequence: quicker prep, sharper buyer conversations, and extra time promoting the place it counts.” — Adrian Fierro, Head of World Market Intelligence at BASF Coatings
Why It Labored
Multi-agent structure with Genie as an agent affords a number of important benefits for enterprises like BASF that look to leverage AI successfully of their enterprise contexts. We conclude the important thing power into the next elements:
Specialised agent capabilities with excessive scalability and modularity: inside a multi-agent system, numerous brokers can give attention to their particular domains or duties, enabling deeper experience in dealing with numerous queries and datasets. Furthermore, organisations like BASF can increase their gateway to AI options with an structure that permits every enterprise division to function independently whereas being centrally orchestrated. This modular design helps handle complexity over time.
Enhanced collaboration and improved person expertise: brokers can share data and context with each other, permitting for extra complete responses that combine information from a number of sources. This facilitates smarter, quicker decision-making throughout numerous enterprise capabilities. By integrating AI endpoints to MSFT Groups as a chat interface, we permit customers to work together with brokers utilizing pure language, making it extra accessible to non-technical stakeholders.
Governance and compliance: Defending private and buyer information is the Commented basis of Marketmind and stays our highest precedence. Each interplay is constructed on strict compliance with BASF’s information safety requirements, leveraging Databricks’ enterprise-grade governance capabilities akin to Unity Catalog for fine-grained entry management, lineage monitoring, and auditability. This ensures that whereas Marketmind accelerates insights and actions, it does so inside a safe, clear, and absolutely ruled surroundings.
Shut group work between BASF, Databricks and companions: From challenge begin, BASF Coatings, Databricks account and product groups, and associate Accenture proactively engaged in workshops,. which helped align enterprise targets, technical necessities, and product imaginative and prescient, setting a robust basis for profitable implementation. Proper on time, hands-on periods created fast suggestions loops. Professional steering was repeatedly supplied by Databricks product group, serving to to customise the answer for the complicated, evolving wants of BASF and guaranteeing enterprise-grade high quality.
Wanting Ahead: Multi-Layered Orchestration and Agent Bricks
With the success of the Marketmind multi-agent supervisor answer, the corporate is now increasing the enterprise affect throughout broader operations, together with Provide Chain, Procurement, Chemetall (Floor Expertise subsidiary), and Individuals & Tradition. Along with our product group, we’re exploring a extra scalable multi-layered structure, the place every division operates its personal multi-agent supervisor, whereas a higher-level Coatings-wide orchestrator serves all customers. This hierarchical system – a “supervisor of supervisors” – strikes the precise stability: it allows division-scoped information and gear entry management, preserves flexibility in agent growth, and helps a Coatings-wide “Ask Me Something” functionality.
One among our future enhancement targets is the adoption of Agent Bricks, launched this yr on the Information & AI Summit. Whereas our present Mosaic AI–primarily based answer helps multi-agent orchestration, it stays code-first and requires a extra hands-on method with added complexity in deployment and administration. Agent Bricks affords a streamlined approach to construct and optimize domain-specific, high-quality AI agent programs for widespread use circumstances, together with multi-agent setups. With options akin to automated optimization, price and high quality effectivity, and user-driven suggestions mechanisms, it simplifies agent implementation and permits groups to give attention to core challenges – information, metrics, and problem-solving. Though we’ve got not but been in a position to absolutely take a look at its capabilities as a result of restricted regional availability, we view Agent Bricks as a visionary path and plan to allow integration as soon as it turns into accessible, accelerating division-specific multi-agent supervisor growth.
