I/O 2026: Google has officially declared the agentic era. Here is what it changes for your company tomorrow morning
Google has officially declared the agentic era. Meanwhile most French SMBs still have not mapped their own processes.
By Patrick de Carvalho, CEO Apps Velocity
Sundar Pichai showed a wall of tokens, a wall of money and an agent that never sleeps. Meanwhile, most French SMBs still have not mapped their own processes. We have a problem, and it is not a tooling problem.
In short — At I/O 2026 Google announced the shift to the agentic era: systems that carry out tasks end to end instead of answering questions. The orders of magnitude show the balance of power, 3.2 quadrillion tokens processed per month and close to $185 billion of annual investment. For a small or mid-sized company the bottleneck is not access to the tool: it is knowing which processes the company actually runs.
Act 1: last night, on my sofa, with the Kindle Scribe on my lap
Last night, 19 May 2026, I watched Sundar Pichai open Google I/O from my sofa outside Paris, the Kindle Scribe on my lap and a coffee going cold beside me.
I always take notes by hand when something changes.
I took notes three times.
First note: "3.2 quadrillion tokens per month". That is seven times more than last year on the same stage. A token is the unit models use to read and write, roughly a short word or a fragment of one. When Google says it processes 3.2 million billion of them per month, that is no longer a demonstration. That is heavy industry.
Second note: "capex $180 to $190 billion". Capex, short for capital expenditure, is long-term investment in infrastructure, as opposed to opex, day-to-day running costs. Google plans to invest around $185 billion this year in servers, chips and data centres. In 2022 the figure was $31 billion. Six times more in four years. For scale, that is the equivalent of the annual budget of the French national education system. One private company. One year.
Third note: "Gemini Spark, a personal agent running around the clock on a dedicated virtual machine".
That third note is the one that made me put the stylus down.
Because the first describes a ramp-up. The second describes a wall of money. The third describes something else entirely: a change in kind.
And that change in kind is going to knock, tomorrow morning, on the door of every company that filed AI under "we will look at it later".
Which is to say, most of them.
Act 2: what Pichai actually announced, and what you miss if you skim
I will spare you the exhaustive list. You can find it everywhere this morning. I will tell you what genuinely made me react.
Pichai announced two things, not ten.
First: the agentic era is officially declared.
Agentic is the stage where AI no longer merely answers questions. It acts. It carries out multi-step tasks. It decides. It queries tools. It comes back. An AI agent is no longer an assistant dictating an email, it is a software colleague that takes the email, checks your calendar, consults your CRM (customer relationship management software), sends a contextual reply, schedules a follow-up and briefs you the next morning on what it did.
Pichai gave that shift an official name: Gemini Spark. A personal agent running twenty-four hours a day on a dedicated virtual machine in Google's cloud. It reaches your tools through MCP (Model Context Protocol), an open standard that lets AI models talk to any application. You can reach it from the Gemini app, by email, by chat, through Chrome.
You read that right: by email. You write to your agent. It answers like a colleague who spent the night working.
This is not two years out. It is this week for early testers in the United States. This summer for Pro and Ultra subscribers.
Second: Google laid the infrastructure that makes it irreversible.
The new TPU 8t and 8i chips (eighth generation Tensor Processing Units, processors dedicated to AI training and inference, designed in-house by Google) are close to three times more powerful than the previous generation, and twice as efficient per watt. A million TPUs distributed across the planet train in parallel. Antigravity 2.0, the harness that lets developers program these agents to run autonomously, opened yesterday. Gemini Omni Flash, the new multimodal model, generates text, image and video from any kind of input.
And behind all of it, the detail that settles the matter: Gemini 3.5 Flash, freshly released, costs less than half of equivalent competing models and runs four times faster.
Translation for a company director: AI agents have just become affordable at scale.
Which does not mean they are free. Nor that they will install themselves in your business. We will get to that.
Act 3: why I am not surprised, and why you should not be either
Let me be direct.
Agentic AI is not a 2026 novelty. It is a promise we have been hearing for as long as we have been hearing about AI.
What is new is that the pieces of the puzzle have finally come together at the same moment: models fast enough, models good enough, models cheap enough, a standardised communication protocol between AI and applications (that same MCP), and planet-scale cloud infrastructure to run all of it continuously.
When I founded Beyond Coder with Pascal Roche in late 2013, we were already telling clients that we would eventually reach code generated from a business intention. Not code from detailed technical specifications. Code from an intention expressed in plain language.
At the time, people thought I was dreaming.
Pascal and I had been partners since 2006 and had already spent seven years looking for a way to cut the entry cost of custom software for smaller companies. Not for the elegance of it. Because we were tired of watching six-figure projects sink while consultancies delivered PDFs.
The term vibe coding came out of Andrej Karpathy's mouth in February 2025. More than ten years after our Beyond Coder. Vibe coding is the idea of programming by describing what you want in plain language and letting the AI generate the code from that intention.
I tell this story for one reason, and it is not to award myself a medal.
Major technological shifts are never a surprise to those who watch them coming, and always a shock to those who were not looking.
Yesterday's I/O is not a surprise. It is a confirmation.
And a confirmation is more dangerous than a surprise. Because everyone will say "yes, we knew", and almost nobody will move.
Act 4: what will actually happen in your company over the next twelve months
Now that you have the context, let us talk about you.
If you run a small or mid-sized company, here is what is coming, whether you decide anything or not.
Within three to six months, your people will start using personal AI agents (Google's Spark, but equally the equivalents from ChatGPT, Claude or Mistral) for their own individual work. They will hand them things: their calendar, their email, pieces of client files, sometimes sensitive information. Sometimes without realising they are transmitting data protected by the GDPR (General Data Protection Regulation).
They will not ask your permission. Not out of malice. Out of habit. Exactly as they started using ChatGPT in 2023 without asking anyone.
Within six to twelve months, your SaaS vendors (the software you rent online by subscription, Software as a Service) will send you enthusiastic emails: "Discover our new built-in AI assistant". Your CRM will get an agent. So will your accounting software. So will your project management tool.
You will find yourself with five, ten, fifteen AI agents inside your company. All different. All well-intentioned. None of them genuinely talking to the others, whatever the marketing says.
Within twelve to eighteen months, you will start receiving client tenders that explicitly ask how you use AI in your processes. Not to look good on a slide. Because your own client is being squeezed by their own market.
At which point, two scenarios.
Scenario A: you mapped what actually happens in your company before the machine got involved. You know who does what, with which tools, on which data, under which rules. You can plug an agent in cleanly, because you have a plan. You move faster. You negotiate your agent-enabled SaaS contracts from a position of real strength, because you know what you want.
Scenario B: you did not map anything. You stack agents on top of an estate nobody ever inventoried. You accumulate software dependencies nobody on the team understands. You pay three subscriptions for three agents doing roughly the same job. And when one of them makes an absurd decision, and they do, nobody on your team can trace back why.
I have watched scenario B play out fifty times in the last two years. Not with agents: with SaaS tools stacked up over the years without a single inventory. The result is identical. Invoices that keep growing and nobody able to explain what is being paid for.
Agentic AI will simply accelerate the phenomenon.
What separates scenario A from scenario B is not company size. Nor sector. Nor AI budget. It is one thing: have you mapped what actually happens inside your company, before the machine gets involved?
A fast agent that gets things wrong quickly is precisely what you do not need.
Act 5: the technical detail nobody will explain to you, and it changes everything
You may have heard of MCP. If not, listen carefully, because it is the detail that makes the agentic era real rather than theoretical.
MCP, Model Context Protocol, is an open standard published by Anthropic in November 2024 and since adopted by Google, OpenAI and most major AI players.
The idea is simple.
Before MCP, every AI talked to every application through a bespoke integration. Your CRM wanted to talk to your AI model? A custom integration. Your calendar wanted to talk to the model? Another one. Multiply that by fifty tools in your company. It is unmanageable, and you pay fifty vendors to maintain fifty bridges that break at every update.
MCP is a universal socket. Like a USB port for peripherals. Any AI agent can plug into any MCP-compatible tool. And any tool can expose its functions to any agent. One socket, many devices.
Pichai confirmed yesterday that Gemini Spark will integrate with third-party tools through MCP "in the coming weeks".
What that means for you, concretely: within months, Spark will be able to open your HubSpot CRM, read your Outlook calendar, amend a quote in your ERP (Enterprise Resource Planning, the integrated business management software), send an email from Gmail and brief you the next morning on what it did. On its own. With no human step in between.
Do you see what that implies?
It means the question is no longer "is my company ready for AI?". The question has become: are my company's processes ready for an AI to operate inside them autonomously?
If the answer is no, do not plug in an agent. Not yet. First understand what your teams do, in what order, on what data, under what rules. Map before you connect.
Otherwise you will have an extremely capable agent creating chaos at very high speed inside processes nobody understood well to begin with.
Act 6: what Pichai's numbers really say about the balance of power
Let me spend two minutes on the keynote figures. Not to impress you. To help you read what comes next.
Gemini app: 900 million monthly active users. In one year it went from 400 million to 900 million. The curve is vertical, and it is not flattening.
Search AI Overviews: 2.5 billion monthly users. An AI Overview is the AI-generated summary shown at the top of Google results. You now see it on close to half of all queries. It is the fastest-moving change in the history of the search engine.
Search AI Mode: more than a billion monthly users. AI Mode is the full conversational experience inside Google: you ask a question, it holds a dialogue with you, it shows dynamic visualisations and dashboards generated on the fly. Not ten blue links at all.
Here is a figure absent from the keynote but appearing in independent analysis. The overlap between the sources cited by AI systems (Google AI Overviews, ChatGPT, Perplexity, Claude) and the top classic Google results has fallen from roughly 70 % two years ago to under 20 % today (Brandlight measurements cross-checked against several analyses, May 2026). In practice: ranking first on Google no longer guarantees being cited by AI at all. Two worlds are quietly separating.
Organic website traffic: between 15 % and 35 % down depending on sector since AI Overviews rolled out at scale. The click is disappearing. Citations are rising, but they do not carry the same economic value.
Why am I telling you this in an article about agentic AI?
Because the two phenomena are the same story seen from two angles.
On the demand side: the end user no longer runs ten Google searches and three clicks through to your site. They ask an agent. The agent searches on their behalf. The agent decides which sources to cite. The user never sees your site, unless the agent decides it is worth citing.
On the supply side: SaaS vendors who fail to put an agentic layer on top of their products will lose. Not tomorrow. In eighteen to twenty-four months. Forrester projects $576 billion in SaaS spending by 2029, up 81 % on 2024. SaaS is not dying. But the SaaS that survives to 2029 is not the one you rent today.
And you, running a smaller company, are caught between the two. You are simultaneously a user of agents, on the demand side, and a supplier of something to your own clients, on the supply side.
There is no scenario in which you can afford not to understand what is happening in front of you.
Act 7: a story from 1998 you will find familiar
One last thing, and I promise it is not old-soldier nostalgia.
In 1998 I was twenty-four. With my partners I launched what we called at the time an online newsstand: planetepresse.com. The idea: buy and read French magazines and newspapers through a web browser, ten years before the Kindle went mainstream and fifteen before consumer e-readers.
We were early. Far too early, in truth.
But above all we had a problem nobody could solve for us: nobody could find us on the internet.
The reference engine in 1998 was AltaVista. Yahoo too, as a human-curated directory. Google had existed for a few months and was still a project by two Stanford students. The term SEO, search engine optimisation, was not in common use. People said "referencing", and it was a mush of opaque tricks sold by intermediaries in ties.
One day a salesman came to see us. Briefcase. Printed deck. He sold us, on an invoice, an "internet referencing campaign" for 30,000 francs, roughly €7,000 at the time, for a company doing its first million in revenue.
He sold us fear. His words: "Without us, you will not exist on the web".
We paid. We lost six months. We eventually worked out how the web really functioned on our own, and we came through it. The cheque was never refunded.
Ever since, I can spot that salesman from ten kilometres away.
He is still around in 2026, in another form. Today he sells you "turnkey AI solutions", two-day "express vibe coding courses", €15,000 "AI audits" that produce an eighty-slide deck nobody will ever reopen.
It is not that AI is a scam. AI is the opportunity of a decade, perhaps of our generation.
It is that fear is being sold exactly the way it was in 1998. And the antidote to fear is never a reflex purchase. It is clear thinking, done cold, by you, inside your company, with your teams.
Act 8: what you can do on Monday morning, at no cost
Let me finish on something concrete. No disguised sales pitch. No seven-point call to action with tick boxes.
Here is what you can do yourself, from Monday morning, without spending a euro:
Get your team leads in a room for ninety minutes. No longer. Ask one question: "If I asked you tomorrow to tell me what your team actually does day to day, process by process, could you?" If the answer is "well, roughly", you have your starting point. You cannot plug an AI agent cleanly onto "roughly".
List your ten most time-consuming business processes. Not your hundred. Your ten. The ones that eat your teams' weeks. Invoicing, chasing payment, quotes, scheduling, reporting, hiring, support, whatever they are. But ten, no more, no fewer.
For each one, ask this question: "Would I be comfortable letting an agentic AI run this process in my place, fully autonomously, for a whole month?" If the answer is no, understand why. That is exactly where your room for manoeuvre sits for the next twelve months.
Pick one single process to trial over thirty days. Not three. Not five. One. With a numeric target, for instance "cut the time spent on this task by 40 %, at equal quality". With a clear end date so the result can be judged without indulgence.
Subscribe to at least two independent AI information sources. Not Google. Not OpenAI. Not vendor marketing blogs. Analysts, researchers, practitioners. So you build your own filter and stop depending on the vendors' official narrative.
That is all.
You can take those five steps yourself, with nobody from outside. No consultancy. No €5,000 training course. And you will already be further along than 80 % of the company directors who believe they are "keeping an eye on AI".
Act 9: the conclusion that is not a conclusion
In yesterday's keynote Sundar Pichai said something that made me smile. "We're now in the part of the AI cycle where people want to see the value in the products they use every day."
He is right. And he is wrong at the same time.
He is right because the enthusiasm of early adopters is no longer enough. What is needed now is tangible, measurable value that survives a board meeting. The era of magical demos is over.
He is wrong because for value to exist inside a company, there first has to be clarity about what the company actually does. And clarity does not come with a $19.99 Ultra subscription.
Clarity is a leadership decision. It means sitting down with your teams and looking at your processes as they exist, not as they are described. It means accepting that foundations are not glamorous, but that everything else rests on them. It means testing small before deploying wide. It means stopping what does not work and amplifying what does.
Google declared the agentic era yesterday. Fine.
The channel changes every decade. The contract has never changed: understand what you do before you accelerate what you do.
I never lose. Either I win, or I learn.
Patrick de Carvalho
FAQ
What is the agentic era?
It is the shift from an AI that answers to an AI that executes. An agent pursues a goal by chaining decisions and actions together, without a human validating each step. The difference with a chatbot is not one of degree, it is one of kind: one produces text, the other produces an outcome.
What is a token, in practice?
A token is the unit models use to read and write, roughly a short word or a fragment of one. Google announced it processes 3.2 quadrillion of them per month, seven times more than a year earlier on the same stage. That is no longer a demonstration, it is heavy industry.
How much is Google investing, and why does the figure matter?
Around $185 billion of capex in 2026, against $31 billion in 2022. Capex, capital expenditure, is long-term investment in infrastructure as opposed to day-to-day running costs. Six times more in four years, for a single private company: that is the order of magnitude of the annual budget of the French national education system.
Do you need to change tools to benefit from this?
No, and that is the most common misreading. The bottleneck is not access to the technology, which is already inside the tools you use. The bottleneck is knowing which processes your company actually runs, in what order, on what data. An agent can only automate what has been described.
Where should a smaller company start?
With mapping, not with buying. Take a week of your activity and note what repeats, what gets passed along by email, what waits for an approval. You end up with the list of automation candidates before spending a euro.
Is search optimisation dead now that answers are generated?
No, it is mutating. What is dying is the equation rank equals click: AI-generated answers respond without sending a visit. The practical consequence is that you now have to aim for the citation inside the answer as much as the position in the list, which requires content that is structured, sourced and dated.
Sources
- Sundar Pichai, I/O 2026: Welcome to the agentic Gemini era, official Google blog, 19 May 2026. blog.google/innovation-and-ai/sundar-pichai-io-2026/
- Anthropic, Model Context Protocol specification, initial version November 2024 with 2025-2026 updates.
- Andrej Karpathy, public remarks on "vibe coding", February 2025.
- Forrester, SaaS Spending Forecast 2024-2029, projecting $576 billion by 2029.
- Brandlight and cross-checked independent analyses, AI citation overlap with Google top results, consolidated May 2026.
- Similarweb, Zero-click search trends, measurements May 2024 and May 2025.
- BrightEdge, AI Overviews coverage of search queries, May 2026 data.
- Growth Memo, AI Overview citation position, April 2026.