Short answer
What is agentic AI and what changed in 2026?
Short answer
Agentic AI is artificial intelligence that does more than answer: it takes a goal, breaks it into steps, uses tools such as your CRM, ERP or email, checks the result and repeats until the job is done. As of October 2026 it is an off-the-shelf product from OpenAI, Google, Microsoft and Anthropic.
Agentic AI is not a new model but a new way of using language models: instead of asking them for a piece of text, you give them a task. An AI agent perceives the situation, plans, acts with tools and checks the result, in a loop. That is the working agentic AI definition I use in this article. For the exact difference between an agent, a chatbot and an automation, see my guide AI agent vs chatbot vs automation; here I focus on the market and on what to do about it.
Agentic AI vs generative AI: what is the difference?
The question behind most searches for agentic AI vs generative AI is simple: what does the agent add? Generative AI produces content that a person then uses; agentic AI uses generative models to decide and carry out the steps itself, within the limits you set.
| Criterion | Generative AI | Agentic AI |
|---|---|---|
| What you give it | A prompt or a question | A goal |
| What it returns | Text, images or code for a person to use | Actions in your systems, then a result |
| Who decides the steps | The person, prompt by prompt | The model, within the limits you set |
| Tools | Optional | Central: it reads and writes in the CRM, ERP, email or other systems |
| What a mistake looks like | A wrong sentence | A wrong action: an email sent, a record deleted |
What has changed over the last ten months is that agentic AI is no longer a demo. The big vendors sell platforms to build, monitor and pay for AI agents, and the industry now shares open standards to connect them. These are the key facts, with dates:
- 9 Dec 2025
The Linux Foundation creates the Agentic AI Foundation (AAIF)
It launches with three donated projects: Anthropic’s Model Context Protocol (MCP), Block’s goose and OpenAI’s AGENTS.md. Its platinum members include AWS, Google, Microsoft, Bloomberg and Cloudflare.
- 5 Feb 2026
OpenAI introduces Frontier
A platform to build, deploy and manage agents in the enterprise, each with its own identity, explicit permissions and built-in evaluation. It started with a limited group of customers.
- 9 Apr 2026
Anthropic’s Claude Cowork opens to everyone
After a preview that began on 12 January, the desktop agent for office work becomes generally available on macOS and Windows.
- Apr 2026
Google launches Gemini Enterprise Agent Platform
At Google Cloud Next ’26 it presents the evolution of Vertex AI to build, scale and govern agents, with a cryptographic identity for each agent and a central registry.
- 1 May 2026
Microsoft Agent 365 becomes generally available
It is a control plane to inventory, monitor and secure a company’s agents. Price: US$15 per user per month as a standalone licence, or included in Microsoft 365 E7 (as of 12 October 2026).
- 25 Aug 2026
McKinsey publishes “The state of AI in 2026”
40% of large organisations are already scaling AI agents, up from 27% a year earlier.
- 8 Oct 2026
Google unveils Gemini agent
A single agent for work that creates sub-agents, combines Gemini and Claude models and records every action in an audit trail. According to 9to5Google, it is in private preview for now.
The common thread is clear: each agent has its own identity, scoped permissions, an action log and spending controls. Google sums it up in one idea: you give the agent “goals, not instructions”. And open standards, such as the protocol I explain in what the Model Context Protocol (MCP) is, are what let you connect those agents to your systems without being locked in to a single vendor.
Agentic AI meaning: is it the same as AI agents?
Use cases
What can AI agents already do in your business?
Today AI agents perform best on repetitive work that requires combining data and applying some judgement, always within a narrow scope. According to McKinsey’s August 2026 report, companies scale them mainly in IT, knowledge management and software engineering; in consumer goods and retail, marketing and sales stand out, and in manufacturing, the supply chain.
AI agents examples: tasks an agent already handles well
These agentic AI examples share one trait: the agent does the searching, cross-checking and drafting, and a person keeps the final say.
- Sorting and prioritising: support emails, tickets or sales enquiries, with a summary and a proposed next action.
- Preparing: weekly reports that combine CRM, ERP and spreadsheet data, or presentations built from that data.
- Chasing: overdue invoices, delayed orders or unanswered quotes, with drafts ready to send.
- Reviewing: contracts, delivery notes or product listings against a checklist.
- Coding: coding agents are the most mature use case; according to McKinsey, around two in ten organisations are already scaling them.
How much autonomy to give it: Gartner’s four levels
The useful question is not “can an agent do it?” but “with how much autonomy?”. On 26 May 2026, Gartner proposed classifying agents into four levels and governing each one in proportion to its risk. Its warning: 40% of companies will downgrade or retire autonomous agents in 2027 because of governance failures that only come to light after an incident.
| Criterion | What the agent does | Example in a small business | Minimum control |
|---|---|---|---|
| 1. Observe | Only reads defined data; the result is seen by whoever asks | Answering “which customers owe us for more than 60 days?” | Scoped data access and a usage log |
| 2. Advise | Recommends or drafts; a person reviews and acts | Proposing the order in which to chase debts and drafting each reminder | Accuracy testing and training for the people who use it |
| 3. Act with approval | Writes data or sends messages, but only with explicit approval of each action | Sending the reminders you approve, one by one or in batches | Approval workflow with a log and an incident plan |
| 4. Act alone | Acts within set limits; people review exceptions and results | Sending standard reminders without prior review | Continuous monitoring, limits, fast rollback and an automatic stop |
My recommendation
Data
How widely is agentic AI really used? The 2026 data
Agentic AI is growing, but mostly in large companies. McKinsey’s survey, published on 25 August 2026, with 1,719 participants from 97 countries, shows that 40% of organisations with more than US$1 billion in revenue are scaling agents in at least one function, up from 27% a year earlier. Among smaller organisations the figure has stalled at 22%.
40%
of large organisations scale AI agents in at least one function
McKinsey, 25 Aug 2026
22%
among smaller organisations, unchanged from 2025
McKinsey, 25 Aug 2026
×5
minimum inference cost of a task handled by an agent compared with a chatbot interaction
Gartner, 17 Aug 2026
70%
of enterprises will abandon agents built by the vendor’s own engineers by 2028
Gartner, 29 Sep 2026
Gartner’s forecasts point the same way: it expects that by the end of 2026, 40% of enterprise applications will include task-specific agents, up from less than 5% in 2025. But it also warns about the noise: many products are being rebranded as “agents” without being agents. If you want to see what companies across the EU are actually doing with AI, with official figures, I cover it in AI for business: what companies really do.
Cost, the forgotten variable
On 17 August 2026, Gartner estimated that sending a task to an agent running a reasoning model costs the provider at least five times more in inference (the model’s computation every time it works) than a chatbot interaction. It also expects that cost per agentic workflow to increase more than fivefold through 2028. The reason: the agent reasons, looks things up and reviews its own work several times.
There is another warning, about how agents are procured. On 29 September 2026, Gartner predicted that by 2028, 70% of enterprises will abandon agents built by the vendor’s own engineers, because costs rise and they cannot evolve them on their own. Its advice: agree knowledge transfer and an exit plan from the outset.
The question is no longer whether agents work. It is how much autonomy you give them, how much each task costs and who is accountable when they get it wrong.
Risks
What are the risks of AI agents? Prompt injection, errors and cost
AI agents carry the risks of any generative AI and one more: because they can act, a mistake is no longer a wrong sentence but an email sent or a record deleted. That is why OWASP, the leading foundation for application security, published its Top 10 of risks specific to agentic applications on 9 December 2025.
Prompt injection: hidden instructions
Prompt injection means hiding instructions in content the agent reads, such as an email, a web page or an attachment. On 22 December 2025 OpenAI described an example of its own: an email with hidden instructions made its agent, while preparing an out-of-office reply, send a resignation letter to the CEO. OpenAI itself admits that, like online scams, it will probably never be fully solved.
The rule I never skip
Compounding errors and uncapped cost
An agent takes several decisions in a row, so an early mistake is carried through to the end. And because every turn of the loop consumes tokens (the units of text the model provider bills for), a poorly scoped task can cost far more than expected. That is why the 2026 platforms include an identity for each agent, a log of every action and spending caps that pause the agent.
| Criterion | What can happen | How to reduce it |
|---|---|---|
| Prompt injection | An email or a web page gives the agent orders and it obeys them | Human approval for external actions, minimal permissions, specific tasks |
| Compounding errors | A data point misread early on contaminates the whole result | Verification steps, a log of every action and sample reviews |
| Uncapped cost | The agent goes round in circles and consumes far more than planned | A spending cap per project, cheaper models for simple tasks |
| Excessive permissions | The agent can touch systems its task does not need | Its own identity for each agent and read-only access at the start |
| Vendor dependency | Nobody on your team knows how to maintain or change the agent | Documentation, internal training and an exit plan agreed in the contract |
Regulation
What does the EU AI Act say about human oversight of AI agents?
The EU AI Act does not regulate agentic AI as a category, but its uses. It also reaches companies outside the EU, including in the UK, when their AI system or its output is used in the EU. If your agent takes part in an Annex III high-risk use case, such as recruiting staff, evaluating workers or assessing a person’s creditworthiness, Article 14 requires that whoever oversees it can understand its limits, monitor for failures, avoid over-relying on it, disregard its output and stop it safely. Article 26 requires the company using it to assign that oversight to people with the training and authority to do it.
Following Regulation (EU) 2026/1744, published in the Official Journal of the EU on 24 July 2026, those obligations apply from 2 December 2027 for Annex III cases. If an agent talks to customers, Article 50 requires you, from 2 August 2026, to tell people they are dealing with an AI unless it is obvious. And Article 4 calls for measures to ensure AI literacy among the people who use it.
My take: even if your agent is not high-risk, Article 14 is the best checklist there is for designing human approval. The legal detail, together with the GDPR, is in my guide the EU AI Act for SMEs.
Impact
What does agentic AI mean for your business?
It means the question shifts from “do we use AI?” to “which decisions do we delegate, and with what controls?”. The impact is not the same in every area:
LeadershipHigh impact
You decide what is delegated and who is accountable. Ask every agentic AI project for a baseline metric, a named owner and a spending cap. Without those three things it is an experiment, not an investment.
Sales and customer serviceHigh impact
This is where it shows first: sorting enquiries, preparing replies and chasing quotes. Keep human approval on everything that reaches the customer, and tell people they are talking to an AI when it is not obvious.
MarketingMedium impact
As well as using agents, your customers will use them to search and compare. If an agent cannot read your website, it will not recommend you: that is the territory of Search & AI Visibility and of GEO (generative engine optimisation).
IT and dataHigh impact
Your job is to give each agent its own identity and minimal permissions, log its actions and connect systems through open standards. Also take stock of the agents that teams are already using on their own.
Plan
What should you do now about agentic AI?
With agentic AI the usual rule applies: process first, technology second. This is the plan I propose, spread across this week and this quarter:
This week: take stock of the agents you already have
Many companies already pay for agents without knowing it: they come built into Copilot Studio in Microsoft 365, Gemini Enterprise or ChatGPT Business. Note who uses them, with what data and with what permissions.
This week: pick two candidate tasks
Run your ideas through this article’s candidate-task test and keep the two that score highest. Park any that make decisions about people until you have checked the EU AI Act.
This month: measure and map the current process
Record the task’s time, errors and cost for two weeks. Then map it step by step with the MAPA method for mapping and automating processes with AI; the template is among my free resources, alongside the checklist “Can AI read your website?”.
This quarter: launch a pilot at level 2 or 3
The agent reads data, proposes and drafts; a person approves. Give it minimal permissions, log every action and set a monthly spending cap. If you outsource it, make sure your team learns to maintain it. This is the approach I take when designing custom AI agents with human approval.
This quarter: compare with the baseline and decide
After a few weeks, measure again. If it is consistently right, move it up one autonomy level; if not, adjust it or retire it. And if you want a second opinion before investing, that is exactly what I offer in my AI consultancy for businesses.
Tool
Is this task a good candidate for an AI agent?
Think of a specific task and tick what it meets. I built this checklist from the criteria I apply as José Galán, Search & AI Visibility and AI agents consultant: three boxes are essential and, if any is missing, the recommendation is to lay the groundwork before building the agent.
Key points
- Agentic AI works with goals, not questions: it plans, uses tools and checks, in a loop.
- In 2026 it is an off-the-shelf product: OpenAI Frontier, Gemini Enterprise, Microsoft Agent 365 and Claude Cowork.
- Real adoption is concentrated in large companies: 40% scale agents, compared with 22% of smaller ones (McKinsey).
- The risks are concrete: prompt injection, compounding errors and a cost per task at least five times that of a chatbot.
- Start at level 2 or 3: the agent proposes, a person approves and autonomy is earned with data.
FAQ
Frequently asked questions about agentic AI and AI agents
What is agentic AI?
Agentic AI is artificial intelligence that acts to achieve a goal: it plans the steps, uses the company’s tools and systems, checks what it has done and corrects course. Unlike a chatbot, it does not stop at drafting an answer. In a business it works best with limited permissions, a log of every action and a person who approves anything sensitive.
What is an AI agent and how is it different from a chatbot?
An AI agent is a system that receives a goal and decides which steps to take to achieve it: it looks up data and acts on tools such as the CRM, the ERP or email. A chatbot converses and answers questions from a knowledge base. The key difference is who decides the path: with a chatbot, the conversation; with an agent, the model itself, within set limits.
What can AI agents do in a business today?
Today AI agents perform best on repetitive tasks that require combining data and applying judgement: prioritising emails or support tickets, preparing reports with data from several systems, chasing overdue invoices or reviewing documents. According to McKinsey (August 2026), they are scaled most in IT, knowledge management and software engineering. Sensitive actions, such as payments or messages to customers, should still go through a person.
Why do so many agentic AI projects fail?
Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. In May 2026 it added that 40% of companies will downgrade or retire autonomous agents in 2027 because of governance failures that surface after an incident. Starting with a measurable task and little autonomy reduces all three risks.
What is prompt injection and why does it affect AI agents?
Prompt injection is an attack that hides malicious instructions in content the agent reads, such as an email or a web page, so that it does something the user did not ask for. It mainly affects AI agents because, as well as reading, they can send, delete or buy. In December 2025 OpenAI admitted it will probably never be fully solved: the defence is to limit permissions and require confirmation.
What does the EU AI Act say about human oversight of AI agents?
The EU AI Act requires effective human oversight of high-risk systems, such as those involved in recruitment or in assessing a person’s creditworthiness: whoever oversees them must understand the system, be able to disregard its output and stop it. Following Regulation (EU) 2026/1744, these obligations apply from 2 December 2027 for Annex III cases. For other agents it is good practice.
Where should my business start with agentic AI?
Start with a task, not a platform. Choose a process that repeats every week, that combines data from several systems and whose result you can measure today. Run a pilot in which the agent proposes and a person approves, with minimal permissions, an action log and a spending cap. If after a few weeks it is consistently right, extend its autonomy step by step.
References
Sources
Checked and verified on 12 October 2026.
- Google Cloud (Thomas Kurian). Gemini at Work 2026: Introducing Gemini agent (8 October 2026).
- 9to5Google. Google Cloud announces “Gemini agent” as universal agent for work (8 October 2026).
- Gartner. Gartner Predicts 70% of Enterprises Will Abandon Agentic AI Built by Vendor Forward-Deployed Engineering by 2028 (29 September 2026).
- McKinsey. The state of AI in 2026: On the road to ROI (25 August 2026).
- Gartner. Gartner Predicts AI Inference Costs Per Agentic Workflow Will Increase More Than Fivefold Through 2028 (17 August 2026).
- EUR-Lex. Regulation (EU) 2026/1744 (Digital Omnibus on AI) (Official Journal of the EU, 24 July 2026).
- Gartner. Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure (26 May 2026).
- Microsoft. Microsoft Agent 365, now generally available, expands capabilities and integrations (1 May 2026).
- Google Cloud. Google Cloud Next 2026 Wrap Up (April 2026).
- Anthropic. Claude release notes: Cowork (12 January and 9 April 2026).
- OpenAI. Introducing OpenAI Frontier (5 February 2026).
- OpenAI. Continuously hardening ChatGPT Atlas against prompt injection attacks (22 December 2025).
- Linux Foundation. Linux Foundation Announces the Formation of the Agentic AI Foundation (9 December 2025).
- OWASP GenAI Security Project. OWASP Top 10 for Agentic Applications for 2026 (9 December 2025).
- Gartner. Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026 (26 August 2025).
- Gartner. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (25 June 2025).
- EUR-Lex. Regulation (EU) 2024/1689 on artificial intelligence (AI Act): Articles 4, 14, 26 and 50, and Annex III (13 June 2024).
This article was created with the help of AI and reviewed by José Galán. I take great care over every post and every translation, but the odd mistake can still slip through. If you find one, write to me: you will be helping me improve.
Custom AI agents
Got a candidate task? Let’s design the pilot.
I design custom AI agents that analyse your data, propose the next action and explain why, with minimal permissions, a log of every step and your approval on what matters. If your case can be solved with an automation, I will tell you that too.
See how I design agents