Google announced a lot at Google I/O 2026. New models, new glasses, a redesigned Search, a 24/7 personal AI agent, a developer stack rebuilt around agents, and an entirely new way to shop. Most of the coverage was framed for developers, consumers, or creators.
We watched the whole keynote through two lenses: procurement, and enterprise adoption. Here are the six Google I/O 2026 procurement implications that matter for CPOs: what's real, what's hype, what your team should do this quarter, and where the Workspace data-protection story holds up versus where it doesn't.
The short version. Google did not announce a procurement product. Google announced the operating system that the next generation of procurement products will be built on, plus a handful of consumer agents that will start doing procurement work whether your P2P vendor is ready or not. Agentic AI just moved from demo to default. The enterprise version of that operating system is safe to deploy. The consumer version, used by employees on personal accounts, is where the shadow-AI and shadow-procurement risk lives.
The model price-performance curve just bent again
Google released two new models at I/O 2026. Gemini 3.5 Flash is the new flagship for speed and intelligence. Google claims it outperforms Gemini 3.1 Pro across most benchmarks while running roughly four times faster than other frontier models. Gemini Omni is a multimodal "world model" built around video understanding and editing. Gemini 3.5 Flash is the engine now powering Search, Spark, Antigravity. All of it.
Pricing came in at $1.50 per million input tokens and $9.00 per million output tokens. Roughly 25% cheaper than Gemini 3.1 Pro on coding workloads. To put that in context, here is how Google's lineup compares against Anthropic and OpenAI:
| Model | Input | Output |
|---|---|---|
| Fast / mid-tier (the bracket 3.5 Flash sits in) | ||
| Gemini 3.5 Flash | $1.50 | $9.00 |
| Claude Sonnet 4.6 | $3.00 | $15.00 |
| GPT-5.4 | $2.50 | $15.00 |
| Flagship tier | ||
| Gemini 3.1 Pro | $2.00 | $12.00 |
| Claude Opus 4.7 | $5.00 | $25.00 |
| GPT-5.5 | $5.00 | $30.00 |
Two things stand out. At the mid-tier, Gemini 3.5 Flash is roughly half the price of Sonnet 4.6 and GPT-5.4 across input and output combined. At the flagship tier, Gemini 3.1 Pro is the cheapest frontier model on the market. Less than half the cost of Opus 4.7 or GPT-5.5 on output, where most production workloads accumulate spend. Whether quality matches the price is a different question. We will publish a deeper procurement-workflow benchmark of Gemini 3.5 Flash and Omni in our upcoming Handshake series.
Two practical implications for CPOs and buyers of procurement tech.
One: the "AI premium" in your P2P vendor's renewal pricing is now mostly margin, not cost. The model running under their AI features is materially cheaper than it was at your last renewal. If their per-seat price went up while their underlying model costs went down, that is a renewal conversation worth having before you sign.
Two: specialist procurement tools running directly on these models will undercut suite "AI features" on price, often by half. Ask your account manager which model powers their AI features and what it costs them per unit. If they cannot answer, or the answer is a model generation behind, factor it into the renewal. This is one of the recurring patterns we cover in why most procurement AI projects fail.
Workspace data protection is real. The consumer-version gap is bigger.
The thing Google did not say loudly on the keynote stage, but matters most for enterprise adoption, is this: every consumer-facing announcement has a Workspace-bound enterprise version, and the enterprise versions come with materially different data protections. Before any of this is worth piloting, your CISO and your DPO need to understand the gap.
What Google commits to on enterprise data
Your data is not used to train Gemini models outside your domain. This applies to Workspace Business, Enterprise, and Frontline plans, and to Gemini for Google Cloud. Prompts, responses, files, and generated content are treated as your customer content under standard Workspace terms. It is the answer to the question your CISO will ask first.
The standalone Gemini add-on is gone. Gemini is now bundled into Workspace Business Standard, Plus, and Enterprise plans. I/O 2026 confirmed the bundling continues. Spark in Workspace, voice in Gmail/Docs, and Pics roll out inside the existing per-seat envelope rather than as paid extras. There is no longer a separate "should we buy the Gemini add-on" decision.
Antigravity for enterprises runs inside your Google Cloud boundary. Cloud customers can now connect Antigravity directly to their Google Cloud projects, with "customer data in your control and agent activity within your secure cloud boundary by default." Standard Cloud Terms apply, including data residency. The consumer Antigravity is a different product with different terms.
Gemini Spark in Workspace inherits Workspace protections. A Workspace business preview is coming "soon" via the Gemini app. It asks for confirmation before sending emails, scheduling meetings, or completing purchases. Admins control what it can access through existing Workspace controls. The Workspace Spark and the consumer Spark are not the same product, and the distinction matters more than the coverage suggests.
The compliance posture is solid. Workspace Gemini holds ISO 42001, BSI C5, FedRAMP High, ISO 27001/27017/27018, SOC 2, and HIPAA. DLP, IRM, client-side encryption, audit logging, and Drive trust rules are all available.
Where the real shadow-AI risk sits
Consumer Gemini is a different product. Employees using personal Google accounts (the free Gemini app, or AI Plus / Pro / Ultra) on company data are operating outside your domain protections. Google may use that data to improve consumer models. Your IT team has zero admin visibility. Every consumer-facing agent demoed at Google I/O 2026 is now a new vector for sensitive enterprise data to leave your control.
Universal Cart has no enterprise version. Employees will use it for work purchases the moment it ships, bypassing preferred suppliers, spend controls, and audit trail. No Amazon Business equivalent (yet), no enterprise terms, no admin console. Shadow procurement waiting to happen.
Personal Spark and Information Agents can connect to whatever the user lets them connect to. An employee granting personal Gemini Spark OAuth access to a work calendar is an exfiltration path that does not show up in existing security tooling.
Three questions go to your IT and security counterparts this quarter. One: are we on a Workspace plan that includes Gemini, and do we have usage telemetry, meaning do we know who is using what, on what data? Two: have we blocked or warned employees about consumer Gemini accounts touching company data? Three: do we have a position on consumer agentic shopping tools? If not, expect business users to start using them, and to discover the exposure in spend audits months later.
The Workspace story is solid. The risk is not Workspace itself. It is the gap between the Workspace version and the consumer version, and the speed at which employees will adopt the consumer version while IT is still evaluating the enterprise rollout. Our AI Readiness Assessment covers this kind of consumer-vs-enterprise gap analysis for procurement functions.
Personal AI agents will route around procurement
Spark turns Gemini from a chatbot you ask into an agent that does. Always-on inside the Gemini app, reasoning across your connected apps and taking action on your behalf. Managing information, following up on threads, completing tasks while you sleep. In beta now, rolling out first to the $100/month AI Ultra tier.
Every business user in your company is about to get a personal agent that can browse the web, send emails, and complete tasks for them. Some of those tasks will be procurement tasks. Sourcing a vendor. Getting a quote. Placing an order. Expensing a subscription.
A category manager asks Spark to "find three suppliers for a 40,000-unit injection-moulding run, get indicative quotes, and book a 30-minute call with the most responsive one this week." Spark identifies suppliers from the open web, sends introductory emails, parses the responses, scores them, and surfaces a shortlist with calendar invites attached. Nothing in that workflow touched your preferred-supplier list or your spend-control thresholds, because neither was available in a form Spark could read.
Your preferred supplier catalogs and approval workflows were designed for a world where a human read them. They were not designed for a world where an agent makes the decision on the user's behalf. If you do not have machine-readable procurement policies, or an agent of your own enforcing them, Spark and its competitors will quietly route around you.
The right move is not to block it. It is to publish your policies in a format an agent can read, and to deploy a governed procurement agent that talks to theirs.
Spark in Workspace is in preview "soon" for business customers, inherits Workspace controls, and asks for confirmation before high-stakes actions. Spark on the consumer AI Ultra plan does not. And that is the version your employees will reach first. Get ahead of the consumer rollout with a clear policy on personal Gemini accounts and company data.
Supplier intelligence just became a $20/month subscription
Information Agents are persistent agents that run 24/7 in the background, scanning the web and Google's real-time data feeds (finance, shopping, sports), notifying you when something material changes. Launching this summer for AI Pro and Ultra subscribers in the U.S.
The consumer use cases Google led with (apartment listings, sneaker drops, sports score alerts) are unimpressive. The underlying capability is not.
This is the most underrated Google I/O 2026 announcement for buyers. Information Agents are, functionally, a free supplier-intelligence and market-monitoring system. You can configure agents to track commodity prices, freight indices, supplier news, regulatory changes, force majeure events, and contract expiry windows. Capabilities your spend analytics vendor currently charges six figures for are about to become a $20-per-month consumer subscription.
The work is not going away. The question of who does it, and how much it costs, just changed. Two moves this quarter: identify the three supplier-intelligence reports your team currently buys or builds manually, and prototype the same workflow on an Information Agent. Then renegotiate the vendor contract before renewal. Or replace it.
Agentic commerce arrived, disguised as a consumer feature
Universal Cart is an AI-powered shopping cart that watches prices, tracks deals across apps, and assembles complex purchases across multiple retailers. The on-stage demo built a custom PC by sourcing parts from different vendors automatically. Google explicitly framed Universal Cart as "the foundation for agentic commerce." Meaning: this is the start, not the destination.
Tail spend, roughly 20% of dollars but 80% of transactions in most organizations, is exactly the workflow Universal Cart was built for. Coupa's own published data says best-in-class companies achieve only 55.3% structured spend. The other 45% is exactly where Universal Cart will land first.
Amazon Business reached roughly $83B in GMV by attacking this same problem with a marketplace. Universal Cart attacks it with an agent that does not need to host the inventory. It shops the open web on your behalf, which is a much harder thing for incumbents to defend against. Your P2P vendor's tail-spend module is now the most exposed surface in your procurement stack. We covered the broader pattern in 12 AI use cases in procurement that actually work.
Universal Cart has no Workspace or enterprise variant. There is no admin console, no spend governance, no preferred-supplier enforcement, no audit trail. Employees will use it on personal accounts for work purchases, and you will not see it until reconciliation. Pilot a governed procurement agent on one tail category this quarter. Both to capture the savings, and to give business users a sanctioned alternative before the consumer version trains them otherwise.
Building procurement agents just got cheaper for everyone
Google's agent-development platform got a major upgrade. Antigravity 2.0 is a standalone desktop app for orchestrating multiple agents in parallel. Antigravity CLI is a keyboard-first agent surface for developers. Antigravity SDK lets you customise and self-host the same agent harness. Managed Agents in the Gemini API spin up an isolated Linux sandbox with a single call. An agent that can reason, run code, browse the web, and manage files.
The enterprise version layers on top: Cloud customers can connect Antigravity directly to their Google Cloud projects, deploy managed agents inside existing infrastructure, and govern them centrally. This is the developer-side counterpart to everything above. It is the infrastructure that lets other companies (including procurement software companies) build agents on Google's stack.
Two implications for procurement buyers
One: every procurement tech vendor on the market now has a credible path to ship agentic features quickly. The next 12 months will bring a flood of "AI agent" announcements from incumbents and challengers alike. Most will be cosmetic. Some will be transformative. The test is simple: can the agent actually complete a multi-step procurement task without a human in the loop?
Two: the cost of building procurement agents just dropped for everyone, including you. Expect newer, smaller vendors to ship more capable products faster than the incumbents can refactor their legacy stacks. This is the dynamic we covered in the retrofit trap.
If your company already runs on Google Cloud, the path to a procurement-agent stack just got shorter. The platform handles the messy parts (identity, isolation, governance, audit logging) that legal and infosec will ask about. The procurement-specific work (policies, workflows, supplier knowledge, category strategy) is what you build on top. If you are on Microsoft, the equivalent is Copilot Studio. On AWS, Bedrock Agents. The same architectural pattern applies.
1. End-to-end task completion: Show me the agent completing a procurement task from intake to PO without a human in the loop. Where does it stop, and why?
2. Policy enforcement: How does the agent know our supplier preferences, payment terms, and approval thresholds? Where do those policies live, and how are they updated?
3. Audit and reversibility: If the agent makes a bad call, how do we see the reasoning, roll it back, and prevent the same mistake next time?
4. Cost and lock-in: What is the marginal cost per task, and what happens if we want to move to a different model in 12 months?
If we were running a procurement team today: five moves to make this quarter
Strategy without action is decoration. Here is what we would do in the next 90 days if we were sitting in your seat.
- Audit your tail spend honestly. Universal Cart's first home is in your unmanaged spend. If you cannot say with confidence what percentage of your indirect spend is unmanaged, that is the first project.
- Get aligned with IT on the consumer-vs-Workspace gap. The Workspace versions of these tools are safe to deploy. The consumer versions, used by employees on personal accounts, are where the shadow-AI and shadow-procurement risk lives. Set a position on personal Gemini accounts and company data this quarter, before the audit findings force you to.
- Make your procurement policies machine-readable. Spark and its competitors will route around you if you do not. Convert preferred-supplier rules, approval thresholds, and category guardrails into a format an agent can consume.
- Pilot a procurement agent on one category where failure is cheap. Tail-spend, low-risk subscriptions, or routine MRO are good starting points. The goal is organisational muscle, not headline savings. Our implementation guide covers the operating cadence.
- Stop assuming your P2P vendor will solve this for you. The incumbents have a roadmap; you do not have time to wait for it. Start the parallel path now and re-evaluate in 12 months.
Read Google I/O 2026 as a forward indicator, not a product launch. The announcements preview where the whole stack is heading over the next 12 to 18 months: faster and cheaper models on a tighter release cadence, agents moving from chat to action, agentic commerce going mainstream. The procurement functions that come out ahead will not be the ones that picked the right model this quarter. They will be the ones built for model portability, able to swap Gemini for Claude for GPT, or for whatever ships next, without rebuilding the workflows around any one provider. Lock-in is becoming the expensive bet.
If you want help turning that read into a 90-day plan, our AI procurement consulting team works with CPOs on exactly this kind of operating cadence. If the bottleneck is team capability rather than strategy, our AI training for procurement teams is built for it. And if you need to start with a diagnosis, the free AI Readiness Assessment is the easiest way in.