Views: 22
Every major software cycle starts with a new interface.
The web gave us the browser. Mobile gave us the app. SaaS gave us the dashboard. Generative AI gave us the chat box.
AI agents are something different. They are not just another interface to software. They are the beginning of software that does the work.
That is why the agent conversation is moving so quickly from demos to budgets. Gartner predicts that by the end of 2026, 40% of enterprise applications will include task-specific AI agents, up from less than 5% in 2025. McKinsey’s 2025 global AI survey found that 88% of organizations are already using AI in at least one business function, while 23% are scaling at least one agentic AI system and another 39% are experimenting. Deloitte expects close to three-quarters of companies to deploy agentic AI within two years, but only 21% say they have mature governance for agents.
The first phase of AI adoption was about productivity. Write this email. Summarise this document. Draft this report. The agent phase is about delegation. Resolve this IT ticket. Investigate this security alert. Update the CRM. Review this pull request. Qualify this lead. Process this insurance claim.
The opportunity is enormous. But so is the fragility of many startups building in the category.
What is an AI agent?
The term “AI agent” is already being abused.
A chatbot that answers questions is not necessarily an agent. A copilot that suggests the next step is not necessarily an agent. An AI feature with a shiny button is not necessarily an agent.
A real agent has four ingredients: a goal, context, tools, and some degree of agency. According to AWS:
An AI (Artificial Intelligence) agent is an autonomous software program that perceives its environment, makes decisions, and takes actions to achieve specific goals set by a human. Unlike standard chatbots that only respond to prompts, agents plan, use tools, and iterate independently.
That last part matters. A model gives an answer. An agent takes action.
It can access systems, reason through steps, use tools, ask for approval, and complete tasks. That is why agents feel so powerful. It is also why they introduce new risks.
Why SMBs are adopting agents quickly
In theory, large enterprises should adopt agents first. They have bigger budgets, more data, and larger IT teams.
In practice, SMBs often have the sharper pain.
A 20-person company does not have enough people for every back-office function. The founder is often the CFO, the sales manager, the HR team, and sometimes the IT help desk. A managed service provider does not want more dashboards. It wants fewer tickets. A small ecommerce company does not want a “generative AI solution.” It wants returns processed, customers answered, stock issues flagged, and invoices reconciled.
This is why agent adoption in SMBs will not look like traditional enterprise software adoption. Many SMBs will not buy an “agent platform.” They will discover agents inside the tools they already use: CRM, accounting, customer support, IT management, marketing automation, scheduling, and vertical SaaS.
Atera is a good example. The Israeli-founded company positions its AI agents around “Autonomous IT,” helping IT teams identify issues, diagnose root causes, and execute fixes. The pitch is simple: reduce ticket volume and automate routine support without adding headcount.
Wonderful AI is another example. The Israeli-founded company is building multilingual AI agents for customer support across voice, chat, and email. Its wedge is not AI support in the abstract, but localised customer support agents for non-English markets, with cultural, linguistic, and workflow adaptation.
That is the SMB lesson: agents become valuable when they map to a job, not a technology trend.
Enterprises want agents, but they want control
Large organizations already have workflows. The problem is that those workflows are fragmented across Salesforce, ServiceNow, Jira, Slack, Microsoft 365, Google Workspace, Snowflake, GitHub, Zendesk, SAP, Workday, and dozens of internal systems.
The promise of agents is to sit across those systems and execute work.
But that is also the risk.
An agent that can read, reason, and act across systems is not just a productivity tool. It is a new non-human identity inside the organization.
This is why Google, Microsoft, Salesforce, ServiceNow, Atlassian, and others are moving aggressively into the category. They already own the systems where work happens. They have the data, permissions, distribution, and procurement paths. For enterprise buyers, that matters. According to a recent Deloitte survey of 3,235 information technology and business leaders from 24 countries,
For startups, it raises a harder question: why should this be a standalone company rather than a feature inside the customer’s existing stack?
Israel’s agentic AI opportunity
Israel has natural strengths in this market: cybersecurity, developer tools, infrastructure, enterprise software, and deep technical talent from elite military units and global R&D centers.
The Israeli agentic AI ecosystem is already forming across software development, IT, customer support, cyber, sales, marketing, and orchestration.
In software development, Pandorian is tackling one of the biggest second-order problems of AI coding: governance. As AI generates more code, engineering teams need validation, review, governance, and context-aware checks. Tabnine, another Israeli-founded AI coding company, has also moved from autocomplete into more agentic workflows, including agents that can work from Jira tickets and help turn requirements into code.
In IT and customer support, Atera and Wonderful represent two important wedges. Atera is focused on autonomous IT for internal teams and MSPs. Wonderful is focused on multilingual customer service agents for enterprises and underserved language markets. Both show that strong agent companies often look less like “AI platforms” and more like labour leverage in a specific workflow.
In cybersecurity, Israel’s advantage is especially clear. Torq is building an AI SOC platform to help teams triage, investigate, and respond to threats faster. Twine Security is creating AI digital cybersecurity employees, starting with identity and access management. Noma Security focuses on securing AI and agents across the enterprise, including discovery, governance, runtime protection, red teaming, and monitoring. Lasso Security and Zenity are also operating in the AI security and governance layer.
This is a very Israeli pattern. When a new computing surface emerges, Israel often finds the security problem before the rest of the market fully understands the productivity opportunity.
The uncomfortable truth: agent startups are vulnerable
The agent opportunity is large, but many agent startups are more exposed than they appear.
The first risk is LLM platform dependency. OpenAI, Anthropic, Google, and others are adding more agentic capabilities directly into their platforms: tool use, browsing, file search, memory, computer use, evaluations, and orchestration. These tools make it easier to build agents. They also make many agent startups easier to copy.
If the model provider gives you planning, tool use, file search, browser control, memory, evaluations, tracing, and deployment primitives, what exactly is your startup’s moat?
The second risk is incumbent distribution. Agents are most valuable when they live where the work already happens. That gives Microsoft, Google, Salesforce, ServiceNow, Atlassian, GitHub, and other system-of-record companies a massive advantage.
The third risk is orchestration commoditization. A lot of early AI startups looked differentiated because they wrapped LLMs with clever prompts, chains, and tool calls. That is no longer enough. Agent orchestration is becoming infrastructure. The durable value is shifting to domain context, workflow ownership, proprietary data, evals, trust, distribution, and outcomes.
The fourth risk is security and liability. A bad chatbot gives a bad answer. A bad agent can send the wrong email, approve the wrong payment, expose the wrong file, delete the wrong record, or escalate privileges across systems.
That means agent startups cannot treat security as an enterprise checklist. Security is part of the product.
The LLMs are not just suppliers
The most important strategic question for agent startups is not “which model should we use?”
It is: what happens when the model company builds our product?
OpenAI is moving deeper into agent infrastructure. Google is moving from models to enterprise agents, marketplaces, no-code creation, and agent-to-agent communication. Microsoft is embedding agents into the productivity stack. Salesforce is turning agents into a CRM and workflow layer. Anthropic’s Claude Code shows how quickly model companies can move into high-value developer workflows.
This does not mean startups cannot win. It means they need to be brutally honest about where they sit in the stack.
If you are a thin wrapper, you are vulnerable.
If you own a painful workflow, integrate deeply into messy systems, learn from proprietary customer feedback, handle governance and edge cases, and deliver a measurable business outcome, you have a shot.
What durable agent startups will look like
The best agent startups will not sell “agents.” They will sell completed work.
They will say: we reduce Tier 1 IT tickets by 40%. We close IAM access reviews in days instead of weeks. We review every pull request against company standards. We resolve customer support issues in Arabic, Hebrew, French, and Dutch. We investigate every SOC alert before a human touches it.
The winners will likely share a few traits:
- Workflow depth over model novelty – TThe model will change every few months. The workflow will not. The startup must understand the job better than the model provider or horizontal platform.
- Proprietary context – The moat is not the prompt. It is the operational data, feedback loops, exceptions, policies, integrations, and human corrections that improve the agent over time.
- Permissioning and auditability – Agents need identity, least privilege, logs, approvals, rollback, and clear boundaries. The more autonomy they get, the more trust infrastructure they require.
- Model flexibility – A startup that depends entirely on one model provider is fragile. Customers will want choice across OpenAI, Anthropic, Google, open-source models, and private deployments.
- Outcome-based packaging – Seat-based SaaS pricing may not fit agents. If the agent does work, customers will increasingly expect pricing tied to usage, resolution, throughput, or business outcome
- Distribution into the workflow – The hardest part will not be building an impressive demo. It will be getting into the daily workflow. Agents need to live inside Slack, Teams, Jira, GitHub, Salesforce, ServiceNow, Zendesk, email, browser, and the systems of record where work happens.
The danger is funding beautiful demos that become features. The opportunity is backing companies that turn agents into trusted digital workers for specific, expensive, recurring jobs.
The winners will own the work
AI agents are not hype in the same way chatbots were hype. They represent a real shift from software that helps humans do work to software that performs work with humans in the loop. For VCs, the agent wave is tempting because the demos are magical. But demos are not moats.
SMBs will adopt them because they are short on people. Enterprises will adopt them because their workflows are too complex and fragmented. Many VCs aren’t very used to investing in the SMB market – we see it with our own portfolio companies at Remagine Ventures. SMB has different pricing, churn, conversion rate and service needs.
But agent startups are entering a brutally competitive market. The LLM companies are moving up the stack. The SaaS incumbents own distribution. The security burden is high. The primitives are commoditising quickly. As I wrote in a recent post, the LLMs are moving from being simply the infrastructure for Agentic AI to developing the end product for the clients or giving them the tools to easily create them alone. This competes with startups and organisations continue to debate build vs. buy.
The next great agent companies will not be the ones with the best demo. They will own the workflow and become the system of record. They will wrap around the customer’s data and create proprietary data. They will improve with usage and be liable when things go wrong.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google’s first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.