The newest phase of the artificial intelligence market is no longer defined by chat alone. It is being shaped by systems that can plan, coordinate tools, retain context, and work across multiple data types simultaneously. In practical terms, that means autonomous or semi-autonomous software agents paired with multimodal models that can interpret text, images, audio, video, documents, and, increasingly, operational signals from the physical world. The result is a new product wave: smarter assistants for office work, domain-specific copilots for regulated industries, and what can reasonably be described as vertical “AI factories” built to turn proprietary data into repeatable business output.
That shift is visible in both enterprise architecture and product design. McKinsey argued in June 2025 that the real breakthrough in Agentic AI is not generic conversation but the automation of complex, multi-step business workflows, especially in vertical use cases where agents are aligned to a company’s logic, systems, and value drivers. PwC similarly described agents as digital teammates that can reason across tasks, adapt over time, and use external tools or Application Programming Interfaces (APIs) to complete objectives. Together, those assessments point to a more consequential trend than consumer novelty: companies are starting to build execution systems, not just response systems.
The multimodal side matters because enterprise work rarely arrives in a single format. A financial analyst may need to read a slide deck, parse a table, inspect a chart, and pull figures from filings. A clinician may need to reconcile notes, pathology images, scans, and lab data. A warehouse operator may need to combine video feeds, edge sensors, and service logs. Google Cloud’s description of multimodal Artificial Intelligence (AI) is straightforward: these models can process multiple input types, including text, images, audio, and video, and generate outputs across formats. OpenAI’s developer documentation likewise notes that its latest models support text and image input with vision capabilities. That combination is what makes the new generation of agents materially more useful in real business settings than earlier single-mode assistants.
Finance offers one of the clearest examples of where this is going. Amazon Web Services (AWS) published an architecture in 2025 for an agentic multimodal financial management assistant built with Amazon Nova and Amazon Bedrock Data Automation. The system was designed to help analysts query portfolios, analyze companies, and generate reports while working from mixed inputs such as text, images, and documents, including earnings-call slides. UiPath is pushing a similar direction from the process side, describing agentic orchestration for banking and financial services as a way to combine Artificial Intelligence (AI), automation, and human review across workflows such as fraud review, mortgage origination, and counterparty risk. These are not generic chatbot demos. They are narrowly targeted products for high-stakes, document-heavy decisions.
Healthcare is moving in parallel, though with tighter workflow and governance demands. Microsoft’s healthcare team has described a cancer-care orchestration model in which general reasoning systems and specialized multimodal agents coordinate across imaging, pathology, clinical notes, and genomic information. The company says the healthcare agent orchestrator is available through Azure AI Foundry’s agent catalog, featuring preconfigured agents designed for multidisciplinary, multimodal healthcare workflows, such as tumor boards. AWS has also published a life-sciences toolkit that includes starter agents for research, clinical, and commercial use cases, as well as supervisor agents that can coordinate multi-agent workflows within a controlled enterprise environment. The direction is clear: vendors are packaging agents not merely as productivity helpers, but as workflow components for specialized clinical and research tasks.
Manufacturing and logistics show how the trend expands beyond documents and language into physical operations. UiPath markets agentic automation for manufacturing to streamline production processes while controlling risk. NVIDIA has taken the idea further into infrastructure. Its Enterprise AI Factory is presented as a full-stack validated design for enterprises building on-premises Artificial Intelligence (AI) infrastructure, while its Metropolis platform focuses on visual Artificial Intelligence (AI) agents for factories, warehouses, transport systems, and logistics environments. NVIDIA says these systems combine sensor and visual data to improve safety, worker productivity, and operational efficiency, and it explicitly frames video analytics, automated inspection, and industrial automation as agent-driven use cases. That is where the phrase “AI factory” becomes more than marketing shorthand: it starts to describe a stack built to ingest enterprise data, train or tune models, and produce decisions or actions repeatedly at an industrial scale.
The incumbent software platforms are moving quickly because they do not want to leave the orchestration layer to startups. Microsoft said in May 2025 that it was adding multi-agent orchestration to Copilot Studio and that organizations could tune models and create agents using their own company data, workflows, and processes. It also reported that more than 230,000 organizations were already using Copilot Studio to create and customize agents. On the business-applications side, Dynamics 365 now markets agentic applications for sales, service, finance, and supply chain operations. AWS, for its part, says Amazon Bedrock now powers Generative Artificial Intelligence (Generative AI) applications and agents for more than 100,000 organizations. The pattern is unmistakable: major vendors are racing to become the control plane for enterprise agents.
Startups are not absent from this shift; they are helping define it. In healthcare, Abridge describes its platform as enterprise-grade Artificial Intelligence (AI) for clinical conversations and says it is trusted by large health systems to turn patient-clinician conversations into clinically useful notes at scale. In finance, Hebbia presents itself as a purpose-built Artificial Intelligence (AI) for asset managers, bankers, advisers, and other high-stakes decision-making environments, with automated workflows and financial context at its core. These firms are narrower than hyperscalers, but that is precisely the point. The market is rewarding systems tailored to a domain, data model, and workflow, rather than broad assistants that require the user to assemble everything manually.
The excitement, however, is colliding with an old enterprise reality: execution and governance decide whether the technology scales. McKinsey warns that agentic systems require a different architecture, one built for observability, traceability, and control rather than simple Large Language Model (LLM) interaction. NIST’s Artificial Intelligence Risk Management Framework (AI Risk Management Framework) and its Generative Artificial Intelligence (Generative AI) profile remain relevant here because they give organizations a practical structure for governing trustworthiness, lifecycle risk, and oversight. Deloitte’s current agentic positioning also underscores a familiar tension: spending is rising, but return on investment is still uneven. In other words, the product wave is real, but the winners are more likely to be the firms that can combine domain expertise, multimodal inputs, orchestration, and controls into dependable operating systems for work.
Agentic Artificial Intelligence (Agentic AI) and multimodal modeling are pushing the market into a more operational phase. The first generation of enterprise Generative Artificial Intelligence (Generative AI) proved that language interfaces could be useful. The next generation is proving that systems can also execute, coordinate, observe, and adapt inside business workflows. From finance and healthcare to logistics and manufacturing, the most important products now being built are not broad novelty tools. They are increasingly domain-specific systems designed to combine enterprise data, multimodal understanding, and workflow automation into practical outcomes. The phrase “vertical AI” is no longer just a funding theme. It is becoming a product strategy.
