AI agents: Market landscape, use cases and adoption barriers

Editor’s Note

This article was adapted by EUVC from an original insight published by Synaptic.

Read the original article.

AI agents promise a significant change in how knowledge work gets done. Instead of generating an answer and waiting for the next instruction, an agent can analyse a task, make decisions and take actions across multiple steps.

In its March 2025 market overview, drawing largely on activity from 2024, Synaptic argued that agents were moving rapidly from experiments into practical applications. Early adoption was particularly visible in coding, knowledge work and finance, while a growing ecosystem of infrastructure and no-code tools was lowering the barrier to building agentic products.

The opportunity was already attracting substantial investment. Yet the market also faced unresolved problems, from unreliable long-running tasks and compounding errors to enterprise integration, privacy and control.

A large language model, or LLM, processes and generates language. An AI agent uses a model as part of a broader system that can plan, make decisions and act.

The practical difference is execution. An LLM might draft an email. An agent could draft it, determine when to send it, schedule delivery and initiate a follow-up.

Agents can also operate across applications. Depending on their permissions and design, they may browse websites, complete forms, query internal data, update systems or coordinate a sequence of tasks without requiring a human instruction at every step.

This autonomy distinguishes agents from conventional AI assistants. It also creates additional risk: an incorrect answer is one problem, but an incorrect action can have operational consequences.

For a complementary view of how agents could change enterprise software and knowledge work, see the related article AI Agents: The Future of Work?.

Synaptic counted more than 200 startups building agentic products across horizontal platforms and industry-specific applications.

Two factors supported that expansion.

First, many organisations still depend on manual, repetitive workflows involving documents, data transfer and coordination between systems. These processes give agents a clear operational target.

Second, open-source infrastructure such as LangChain, AutoGen and Dify reduced the work required to prototype agentic applications. No-code products lowered the barrier further by allowing non-technical users to describe workflows in natural language.

This accessibility created fertile ground for experimentation and faster product launches. It also lowered barriers to entry, making differentiation and durable distribution increasingly important.

Synaptic found that the most mature agent applications were emerging in domains with repetitive, information-intensive workflows.

Coding agents aim to handle larger parts of the software-development lifecycle, including writing code, debugging, optimisation and routine maintenance.

Products including Devin, Magic.dev, Copilot Workspace, Factory and Cosine were developing variants of the autonomous software engineer. In a published customer case study, Devin reported 12-times efficiency and 20-times cost savings for work completed with Nubank.

These examples illustrated the attraction of coding as an early agent market: tasks are digital, outputs can often be tested, and work can be broken into discrete steps.

Knowledge-work agents focus on searching, analysing and synthesising large volumes of information.

Hebbia was being used across workflows from financial sourcing to due diligence and for complex analysis such as executive succession planning. Synaptic also highlighted Orby AI and Wokelo as products supporting research, compliance and due-diligence workflows.

Fintech offered another concentration of agentic applications. Companies including Sema4.ai and Parcha were developing tools for invoice reconciliation, receivables matching and anti-money-laundering screening.

Norm AI focused on regulatory compliance, including what its investor described as critical regulatory assessments for large companies.

The appeal of these applications lies in their combination of structured processes and high volumes of documentation. However, their regulated nature also increases the need for traceability and human oversight.

In healthcare, Oracle reported that its Clinical AI Agent reduced documentation time by 41%.

In legal work, Spellbook’s Associate was designed to plan and execute projects across multiple documents and applications, review its work and adapt to broader assignments.

These use cases point to the same underlying opportunity: agents can be valuable where professionals spend significant time moving between information, documents and software systems.

By early 2025, major technology companies were developing foundational agent capabilities or integrating them into existing ecosystems.

Anthropic demonstrated agents that could interact with computers by typing, clicking and navigating interfaces. Google’s Project Mariner explored web-based actions such as completing forms and supporting shopping or booking tasks. OpenAI, Microsoft and Salesforce were also developing or integrating agentic capabilities.

For Synaptic, these moves validated agents as an important direction for software. They also suggested that specialist startups would face competition from platforms with existing models, distribution and enterprise relationships.

At the same time, the category could generate partnership and acquisition opportunities where smaller companies offer domain expertise, differentiated workflows or technical capabilities that large platforms lack.

Consumer adoption presents a separate challenge because users must be willing to let software act on their behalf. The related article Consumer AI agents: Ready for take off? provides additional context on trust and adoption in consumer markets.

Synaptic’s analysis recorded more than 150 funding deals involving AI-agent startups in 2024, an increase of 81.4% year on year. It estimated that companies in the category raised more than $1 billion during the year.

Several large rounds illustrated investor interest:

Synaptic also counted more than 130 Y Combinator deals in the category and identified Andreessen Horowitz, Sequoia, Abstract Ventures and Lightspeed Venture Partners among the active investors.

These figures should be read as a snapshot of the market in 2024 rather than a current funding total. They nevertheless show how quickly agents became a major investment theme.

User attention was uneven but substantial among several emerging platforms.

Synaptic reported that Wordware attracted more than 15 million website visits in one month following its August 2024 launch. The no-code platform allows users to create agents with natural-language instructions, making agent development more accessible to non-technical users.

Hebbia recorded approximately 300% quarter-on-quarter growth in website visits, according to Synaptic’s analysis. Dify, an open-source platform combining AI workflows, retrieval-augmented generation, agent capabilities and model management, received more than two million website visits and reached 50,000 GitHub stars in the third quarter of 2024.

Traffic does not establish durable product adoption on its own. However, these signals supported Synaptic’s broader view that both technical and non-technical users were actively exploring agent-building tools.

The market’s growth did not remove the underlying technical and operational challenges.

Agents can perform well on individual steps but become less reliable as tasks grow longer. Errors may compound across a multi-step workflow, particularly when later actions depend on an incorrect earlier decision.

Hallucinations remain a related problem. An agent may produce or act on information that appears plausible but is inaccurate.

Agents that operate through graphical interfaces can struggle with actions that humans find routine, such as navigating complex pages, selecting the correct interface element or avoiding unnecessary scrolling and dragging.

This fragility matters when an agent cannot rely on a stable application programming interface and must interact with software through the user interface.

Enterprise deployment requires agents to connect with existing systems and access organisational data. That creates integration complexity alongside questions about permissions, data handling and security.

The risks increase when an agent can initiate payments, alter records or handle sensitive information. Organisations need safeguards that match the consequences of the actions an agent is authorised to take.

Autonomous execution requires clear boundaries. Enterprises need to determine which actions require approval, how decisions are logged and how an agent can be stopped or corrected.

The value of autonomy therefore depends on control. Removing human involvement from every step is not necessarily the right objective, particularly in regulated or high-consequence workflows.

Synaptic’s analysis pointed to four longer-term developments.

Agents may take over complete workflows rather than support isolated tasks. If their reliability improves, this could alter how organisations allocate work across people and software.

Open-source frameworks and no-code tools make it possible for a wider range of users to design agents. This could extend agent adoption into everyday activities such as calendar management, travel planning and studying.

The fundraising activity recorded in 2024 suggested sustained investor interest in the category. Synaptic expected large rounds to continue, although individual company outcomes remained uncertain.

Agents may become a new layer between software systems, coordinating actions that previously depended on conventional integrations or manual data transfer.

This could affect the API economy and the way internal functions such as sales, operations and compliance interact. Multi-agent systems, in which specialised agents coordinate across workflows, represent a further extension of that possibility.

A related article on how AI agents could redesign retail organisations explores what this operating-model shift might look like in one industry.

The AI-agent market expanded quickly because the proposition is clear: software that can act is potentially more valuable than software that can only answer.

The strongest early applications targeted workflows that were repetitive, digital and expensive to perform manually. Open-source infrastructure, no-code platforms and Big Tech participation accelerated experimentation, while substantial funding supported a growing field of startups.

But autonomy raises the standard agents must meet. Reliability, permissions, integration, privacy and oversight are not secondary considerations. They determine which workflows organisations can responsibly delegate.

The long-term market will therefore depend not only on what agents can do, but on whether users and enterprises can trust them to do it consistently and within clear boundaries.

Synaptic is an AI-powered market intelligence platform for private-market investors. It transforms company signals, including financials, pitch decks, websites, news and alternative metrics, into connected data that helps investors discover, research and monitor companies.

Learn more about Synaptic.

This article was adapted by EUVC from an original article published by Synaptic.

Read the original article.