By Stuart Kerr, Technology Correspondent, LiveAIWire

On April 22, 2026, OpenAI, Google and Salesforce each announced a competing enterprise AI agent platform within hours of one another, with none of the three appearing to know the others were about to move. That kind of simultaneous, uncoordinated launch almost never happens in enterprise software. It happened because every major technology company has concluded, at roughly the same moment, that agentic AI is the next platform war, and that losing it is not an option.

Agentic AI is software that plans and carries out multi-step tasks on its own, rather than waiting for a person to ask a single question and give a single answer. Gartner expects the share of enterprise applications built around this kind of task-specific agent to jump from under 5 percent in 2025 to 40 percent by the end of 2026, an eightfold increase in a single year that the research firm compares to the early days of cloud computing. That is why OpenAI, Google, Microsoft, Salesforce and Anthropic are all spending heavily to be the platform teams choose, and why the pace of announcements has accelerated so sharply in the past few months.

What Agentic AI Actually Is, and Why It Is Not Just a Better Chatbot

A chatbot answers the question you ask it. An agent is given a goal and works out the steps needed to reach it, often without asking for permission at each stage. Gartner’s own framing is useful here: an AI assistant helps you do a task, while an AI agent does the task. A cybersecurity example the firm uses is an agent that scans network traffic and system logs continuously, identifies a threat pattern, and initiates a response, all without a human approving each individual step. That shift, from answering to acting, is what separates agentic AI from the generative AI wave that preceded it.

This distinction matters because it changes what companies are actually competing to sell. It is no longer just about who has the smartest model. It is about who can be trusted to let software take real actions inside a real business, a payroll system, a customer database, a supply chain, without constant human supervision. That trust question is why the current wave of announcements is arriving with heavy emphasis on governance, permissions and audit trails alongside the capability claims.

The Day Five Tech Giants Moved at Once

OpenAI’s own announcement introduced what it calls workspace agents, shared, cloud-hosted agents built on its Codex model that can keep working after an employee logs off, run on a schedule, and operate inside Slack as well as ChatGPT. The company’s own examples include an agent that handles parts of month-end financial close and one that triages software requests and opens IT tickets automatically. The feature launched in research preview for ChatGPT Business, Enterprise, Edu and Teachers plans, with free access running until May 6, 2026, before usage-based credit pricing began.

Google unveiled its own Gemini Enterprise Agent Platform on the same day, and Salesforce simultaneously expanded its partnership with Google Cloud to connect its Agentforce platform to Gemini. Salesforce’s most recent Agentforce Commerce release shows how far this has already gone commercially: the company reports that AI-influenced sales made up 20 percent of global online spending last holiday season, worth 262 billion dollars, and that retailers running their own shopper agents saw sales grow 59 percent faster than those that had not adopted the technology.

Salesforce’s own investor results back up how quickly this is scaling. In its most recent quarter, the company reported that Agentforce had processed more than 3.2 trillion tokens through its systems, with nearly 9,500 paid Agentforce deals in place and combined Agentforce and Data 360 revenue up 114 percent year over year. Whatever the eventual market-size estimate turns out to be, the deployment numbers from the companies actually selling this technology are large and growing fast.

What This Means for You

If you work in an office job that involves repetitive, multi-step processes, invoice checking, report generation, customer follow-up, ticket triage, some version of your workflow is a live candidate for an agent within the next 12 to 24 months. That does not necessarily mean your role disappears. In most of the deployments reported so far, agents take over the coordination and paperwork around a task while a person still makes the judgement calls and handles exceptions. The practical question worth asking inside your own organisation now is not whether agentic AI is coming, but which of your recurring weekly tasks would be the first sensible one to hand over, and what oversight you would want in place before you did.

Why Every Major Company Is Racing, Not Just Following the Trend

The urgency is coming from a specific warning inside Gartner’s research. The firm’s analysts have told chief information officers they have a window of only three to six months to define their agentic AI strategy or risk being permanently outpaced by faster-moving competitors. Gartner’s best-case projection has agentic AI driving roughly 30 percent of enterprise application software revenue by 2035, more than 450 billion dollars, up from about 2 percent in 2025. For a platform vendor, being the default choice during this narrow adoption window is the difference between owning a market and playing catch-up in it for a decade.

That explains the pattern behind the announcements. Microsoft is rebuilding Copilot around an agent layer called AutoPilot despite acknowledging internally that fewer than 4.5 percent of its 450 million Microsoft 365 users currently pay for Copilot features, a sign that even the companies with the largest distribution still have to prove agentic AI earns its keep. Not every company is winning this race at the same speed. But none of them can afford to sit it out, which is precisely why the announcements keep coming.

Where Agentic AI Is Already Running Quietly

Much of the public conversation about agentic AI focuses on office software, but some of the most mature deployments are happening away from any screen at all. As covered in our reporting on agentic AI in manufacturing, a Jaguar Land Rover body shop now runs a quality inspection process in which AI agents examine every weld, log defects and reroute affected vehicles, all in under four seconds per car, with no human reviewing the result unless the defect severity crosses a defined threshold. Similar patterns are showing up in agentic systems running on edge hardware, where an AI agent embedded in a wind turbine can detect a fault, diagnose the cause and generate a maintenance order in under 200 milliseconds without ever touching the cloud.

These industrial deployments matter for the enterprise software race because they are proof of concept at scale. When a factory floor system is making autonomous decisions with physical consequences and doing so reliably, it becomes a much easier sell to convince a finance department that an agent can safely handle month-end reconciliation. The credibility built in manufacturing and logistics is part of what is now being marketed into back-office software.

The Governance Problem Nobody Has Fully Solved

The honest caveat in almost every one of these announcements, including OpenAI’s own, is governance. Giving software the ability to take real actions inside real systems creates real risk: a misconfigured agent that emails the wrong client, approves the wrong invoice, or is manipulated by malicious content hidden in a document it was asked to read. OpenAI’s own materials describe built-in defences against exactly this kind of prompt injection attack, and Salesforce and Microsoft have both introduced monitoring dashboards specifically so a human can see what an agent has done and pause it if something goes wrong.

This is also the space where the AI vendors themselves openly compete on more than capability. As our comparison of ChatGPT, Gemini and Claude found, each of the leading AI providers now positions itself differently on trust, precision and integration depth, and the same differentiation is playing out one level up, at the agent-platform layer, where the pitch to a chief information officer is as much about auditability as it is about raw performance.

The Money and the People Behind the Race

None of this is cheap, and the competition for the engineers who can build it has become its own story. As we detailed in our coverage of the AI talent arms race, top AI researchers are commanding compensation that rivals senior Wall Street pay, and companies have been willing to make seven-figure counteroffers to retain a single engineer being poached by a rival. That spending sits alongside the tens of billions being committed to compute infrastructure that our reporting on the broader AI infrastructure race has tracked across Google, Amazon and Meta. Agentic AI does not run without both the people to build it and the data centres to serve it, and the same handful of companies are trying to win all three fights simultaneously.

What Happens Next

Gartner’s own roadmap suggests the current wave of task-specific agents is only the second of five stages, with agents that collaborate with each other inside a single application expected by 2027, agent ecosystems working across different applications by 2028, and a stage the firm calls the new normal by 2029, in which at least half of knowledge workers are expected to build, govern or direct agents as a routine part of their job. Whether that timeline holds exactly is less important than the direction it describes. The tools available to office workers a year from now are likely to look meaningfully different from the ones available today, and the companies racing hardest right now are the ones betting they can define what that difference looks like.

About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.