A year ago, "agentic AI" was a slide in a vendor deck. Today it's a line item. Microsoft, OpenAI, Google, Salesforce and Anthropic have each shipped products in 2026 that do the same fundamental thing: take a goal, break it into steps, and execute those steps across real systems — for hours at a time, with far less human babysitting than the chatbots of 2023–2024 ever needed.
We use two words a lot on this site: agentic and autonomous. They're not the same thing, and the gap between them is exactly what the entire industry is racing to close. Agentic AI handles a bounded task — chase a lead, reconcile an invoice, draft a report — and hands it back for a human to approve. Autonomous AI runs the whole workflow end-to-end, inside guardrails, with a human only in the loop for exceptions. Every vendor below is trying to move its customers from the first category into the second, as fast as it safely can.
The 18-Month Sprint
This didn't happen gradually. It happened in a compressed run of product launches, most of them inside the last six months.
The chatbot era. Copilot, ChatGPT and Gemini answer questions and draft text. Nothing acts on your systems without you copying and pasting the output yourself.
Google rebrands Vertex AI into the Gemini Enterprise Agent Platform at Cloud Next, absorbing Agentspace and pushing the Agent2Agent (A2A) protocol toward version 1.0 in production.
Microsoft's Copilot Agentic Mode goes generally available inside Word, Excel and PowerPoint — Copilot can now take multi-step actions in your documents without step-by-step instructions.
OpenAI launches ChatGPT Work, an agent built on GPT-5.6 that can run unsupervised for up to 24 hours on the enterprise tier — directly answering Anthropic's earlier Claude Cowork launch.
Where the Big Five Actually Stand
Strip away the keynote language and here's what each vendor has actually shipped, and roughly how big it already is.
Microsoft 365 has quietly become an agent platform. Copilot Agentic Mode is GA in Office apps, Copilot Studio is now a full agent-orchestration canvas, and Agent 365 is the new enterprise control plane — governance, lifecycle management and policy enforcement for every agent running in an organisation. Microsoft also shipped "Autopilots," always-on background agents (the first, Microsoft Scout, handles meeting prep and deadline tracking across Teams, Outlook and SharePoint without being asked).
Launched 9 July 2026 on GPT-5.6, ChatGPT Work turns ChatGPT into a task-completing agent: up to 8 hours of unsupervised runtime on the standard tier, 24 hours on Enterprise, browsing the web, writing code, and working across 200+ connected business apps via OpenAI's new Agent API. It's a direct response to Anthropic's Claude Cowork, and the clearest sign yet that the "assistant vs. agent" line has moved.
Google folded Agentspace into a single Gemini Enterprise Agent Platform and is pushing interoperability harder than anyone: its Agent2Agent (A2A) protocol is now at version 1.2, governed by the Linux Foundation's Agentic AI Foundation, and running in production — not pilot — at roughly 150 organisations, with Microsoft, AWS, Salesforce, SAP and ServiceNow all running it too. The pitch is that agents built on different vendors' platforms can hand off tasks to each other automatically.
Agentforce is the fastest-scaling of the group by deal count: from roughly 5,000 deals and $900M in annual recurring revenue at the start of Salesforce's fiscal year to around 29,000 deals and $2.9B in ARR by year end — over 200,000 deployments in Agentforce's first year of general availability.
Anthropic's growth has been the sharpest shift in enterprise spend of the group: its share of combined OpenAI–Anthropic enterprise spending grew from roughly 10% at the start of 2025 to over 65% by February 2026, with 3,000+ enterprises now building agentic applications on the Claude API and Deloitte alone rolling Claude out to around 470,000 employees. Anthropic's Model Context Protocol (MCP) — the standard for connecting an agent to outside tools and data — has become close to a default across the industry, adopted well beyond Anthropic's own products.
What This Looks Like From the UAE
None of this is playing out at arm's length here. The UAE isn't watching the agentic race from the sidelines — by most measures it's ahead of it.
Dubai's Crown Prince, HH Sheikh Hamdan bin Mohammed bin Rashid Al Maktoum, has separately directed the emirate's entire private sector to transition to agentic AI within two years. This is regulatory tailwind most markets don't have — but it also raises the bar. When your government mandates the shift, "we're still evaluating vendors" stops being a viable timeline.
The Part Nobody Puts in the Keynote
Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027 — not because the technology fails, but because of escalating costs, unclear business value, or governance built in after something already went wrong.
Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027"Separately, Gartner forecasts that up to 40% of enterprise applications will feature task-specific agents by 2026, up from under 5% in 2025 — the fastest twelve-month jump we've tracked for any enterprise software category.
Where Neural Horizons AI Fits
We are not Microsoft, OpenAI, Google, Salesforce or Anthropic, and we don't sell any single one of their roadmaps. Every one of the platforms above is genuinely capable — and every one of them is also, currently, the subject of an enormous marketing budget telling you it's the only one you need. It usually isn't. And this is the part we're deliberate about: we don't want to be another Microsoft consultancy.
Microsoft, SAP, Oracle, Dynamics, HubSpot, Salesforce — all of it — is infrastructure. It's the operating system your business runs on. Neural Horizons builds the layer above that: the Intelligence Operating System that sits across whichever tools you already have, orchestrates the agents that act inside them, and is accountable for what those agents actually do.
= Operating System
= Intelligence Operating System
Our work is deliberately model-agnostic. We audit what a business actually has — its existing tools, its data, its department structure — and build the orchestration, governance and decision-making layer on top of it, on whichever underlying platform makes sense for the job, rather than steering every client toward one vendor's stack because it's the one we're incentivised to sell. That's the value an independent implementation partner adds in a market this loud: someone whose recommendation isn't a function of which vendor pays the referral fee.
See the full Zonar OS™ architecture and how it's deployed →
What This Means for Your Business
- The technology gap has closed. Whether you build on Microsoft, OpenAI, Google, Salesforce or Anthropic's stack, the underlying capability to run real autonomous workflows now exists. The differentiator has moved from "can it be done" to "is it governed properly."
- Interoperability is arriving fast. With A2A now running in production across Microsoft, Google, AWS, Salesforce and ServiceNow, agents built on different platforms can increasingly hand off tasks to each other — meaning vendor lock-in matters less than it did twelve months ago.
- Governance is the real project, not the agent itself. Gartner's own numbers say the majority of failed agentic AI projects fail on cost, value clarity or governance — not model quality. Build the guardrails and the audit trail before you scale, not after.
- UAE businesses have less runway than most. With a national 2028 target and a two-year private-sector directive already on the table, "wait and see" is a shorter strategy here than almost anywhere else in the world.
The vendors will keep shipping. The keynotes will keep getting louder. None of that changes the actual question in front of a UAE or MENA business right now: not which AI company to trust, but which task in your business is worth handing to an agent first — and whether you'll know, with full visibility, exactly what it did.