Every AI Platform Now Lets Outside Services In. What That Changes for Companies That Sell Services.

Last updated: July 2026

In July 2026, Google began letting people connect third-party apps directly inside AI Mode, its conversational layer in Google Search, starting with Instacart, Canva, and YouTube Music for US users (Google). Weeks earlier it had added apps to Gemini Spark, its agentic assistant, over MCP (Google). With that, the last major AI company joined a pattern the others already lived in: outside services now run inside the AI surface, right where the customer is asking. ChatGPT does it. Claude does it. Google now does it in both its assistant and its search box.

For a developer-tools company, this is a plumbing update. For a services vendor, a company selling a quote, a policy, a loan, a plan, a booking, it is something else. It is the moment the channel where their customers ask for advice became singular, transactable, and measurable, all at once. The rest of this piece is about what that means for them, not for the platforms.

Start with the stakes. On the day the first quoting insurance apps went live in ChatGPT, broker stocks had their worst session in years. Willis Towers Watson fell around 12%, its steepest drop since 2008; Aon and Arthur J. Gallagher each shed roughly 9 to 10%; and the S&P 500 Insurance Index posted its biggest single-day fall since October (Bloomberg, Insurance Journal). That is tens of billions in broker value gone in a single day. Not because the apps were clever, but because they proved a service could be sold at the exact moment a customer asks, inside the assistant, before that customer ever reached a broker’s website. That mechanic now works on platforms reaching more than two billion people.

What Actually Changed for Services Vendors

The mechanism is the same everywhere. The assistant can call a vendor’s live systems in real time, reach the pricing engine, check real availability, generate a quote, and act on it inside the same conversation, instead of describing the company from stale training data (primer here). Most of this runs on a shared open standard, MCP, that Anthropic released in 2024 and OpenAI, Google, and Microsoft adopted. But the label matters less than the outcome: every major platform now has a way to let outside services in, right where the customer is asking.

Three things follow, and they are the whole story for a services vendor.

AI stopped being a place a brand gets mentioned and became a place it gets sold. Monitoring tools report whether an assistant recommends a company. That is the old game, top of funnel. The new game is bottom of funnel: inside the conversation, the customer describes their situation, gets a personalized quote, and completes the step that matters, a lead captured, a quote accepted, an application started, or the sale itself where the product and regulations allow. One recent headline is easy to misread. OpenAI built its own platform-level checkout rail with Stripe, the Agentic Commerce Protocol, then in March 2026 pulled back from it after in-chat purchases of retail goods converted roughly three times worse than a click-through to the retailer’s site (Forbes). That was OpenAI’s retail payment button underperforming. It says nothing about whether a vendor’s own app can convert. A vendor’s own app runs its own funnel and owns its own conversion event, independent of any platform checkout, and the vendor decides what, if anything, hands off to existing systems. Being recommended is nice. Owning the conversion inside the conversation is revenue.

One build reaches everyone. A year ago this was barely a channel at all: ChatGPT's app platform did not launch until October 2025, and Google's arrived only in mid-2026. Now all three surfaces exist and share a protocol, so a single build can travel across them instead of being rebuilt for each, and it can serve all three buyer motions: the consumer asking for advice, the small-business owner sanity-checking a broker's price, and the broker's own procurement agent fetching quotes across carriers. One build, deployed where the customers already ask.

The app is now the least valuable piece a vendor owns. This is the part most companies get backwards. With the tools open and every platform offering a way to connect, standing up a working app is the easy day.

Which Platform for Which Service

Because every platform can now do roughly the same thing, the question that matters is distribution: which assistant puts a vendor in front of the right buyer. Each made a different bet.

ChatGPT’s edge is reach plus commerce ambition. It is furthest down the road on agentic payments, having co-built the Agentic Commerce Protocol with Stripe, and it also hit the wall first: in March 2026 it scaled back that platform-level checkout for retail goods and started routing those purchases to merchants’ own sites (The Keyword). That is about OpenAI’s own retail payment rail, not about whether a vendor’s app can convert. A vendor’s own app runs its own funnel and captures its own conversion inside the conversation regardless. So ChatGPT’s real pull is the sheer size of the audience arriving with intent. Claude’s edge is quieter and, for regulated B2B, maybe more valuable: it is where the machines talk to each other, so when a broker’s procurement agent compares carriers, that conversation is more likely to run on Claude than in a consumer window. Gemini’s edge is ambient reach, the assistant already sitting inside the products billions of people open without thinking.

That last point is where Google is easy to misread, because it shipped apps on two different surfaces in mid-2026 and they are not equal. Gemini Spark, its agentic assistant, connects outside apps but is gated to premium Google AI Ultra subscribers and absent from regular Gemini chat (Google). Connected Apps in AI Mode, inside Google Search, began rolling out to US users on July 16 with Instacart, Canva, and YouTube Music, so a shopper can build a cart straight from a prompt like everything needed to host a barbecue (Google). The second surface is the one with real reach, because AI Mode sits inside Search’s one-to-two-billion-person audience. The catch for a services vendor is that AI Mode’s Connected Apps are curated and partner-by-partner today. Google says more apps and developers are coming soon but has not committed to open, self-serve access, so the door will likely widen the way ChatGPT’s already has, without a date to bank on. Instacart is the tell either way. It already runs inside ChatGPT, Claude, and Gemini, and now Google Search too (Instacart). One service, every surface.

One more caution: do not over-index on any single platform’s current identity. In June 2026 OpenAI put its Codex agent inside ChatGPT and shipped six business plugins for sales, analytics, and investment banking (9to5Mac). It is chasing the same high-value business work as Claude and Google. The platforms are converging, so a build tied to one of them is a bet that has to be made again next year.

Under the Hood, the Plumbing Is Standardizing

A services vendor does not need to track every acronym. It helps to know the plumbing under all of this is standardizing into a stack, each layer removing a different piece of friction between the moment a customer asks and the moment the vendor gets paid. These are not shopped for one at a time. They snap together.

Sources: MCP Registry, WebMCP spec, A2A, AP2.

Read the stack top to bottom and it maps the entire sell-through motion for a service: get discovered (Registry), be callable (MCP, or a company’s own site through WebMCP), coordinate across companies (A2A), and get paid with proof (AP2). A year ago almost none of this existed as a standard. It is being built now, by a coalition of rivals, on protocols no single vendor controls. That is the difference between AI as a demo and AI as a channel a business can run on.

Two honest notes on timing. MCP and A2A are production; the discovery, browser, and payment layers are still early, preview specs and experimental browser trials, so anyone promising “publish once and be instantly buyable everywhere” is ahead of the facts. And for regulated vendors, the payment layer matters for a non-obvious reason: AP2’s signed authorization trail, its record of exactly what the customer approved, is the audit evidence that lets an agent move money at all. Vendors build on what is solid today and position for the layers snapping in above it. The ones doing that are not waiting for the stack to finish. They are compounding presence while it assembles.

Where the Real Work Is

If the app is easy, where does a services vendor actually win or lose? We have built more than 20 live distribution MCPs and tested more than 20 apps in production, and the answer is consistent. Three things, none of which is the app itself.

Getting suggested. Across the apps we tested, the assistant treats a vendor's tool response as raw material, not a final answer, and decides on its own whether and when to surface it. Every platform's discovery engine is still primitive and improving fast. The vendors live and optimized now are the ones the assistants learn to recommend as those engines mature, the same compounding that rewarded early SEO. Being late is not neutral. It means a competitor is the one the assistant already knows.

Measuring it. Default web analytics capture only a sliver of AI-sourced traffic. An AI app with no visibility into who used it, at what point they dropped, and what converted is a website with no analytics: technically live, commercially blind. A channel no one can see cannot be optimized.

Staying compliant. For a vendor selling insurance, banking, or credit, a quote generated inside an assistant is still a regulated interaction. Rules move, and they move per jurisdiction. The compliance layer is not a one-time review, it is continuous, and it is table stakes in regulated verticals from the first customer, not the hundredth.

That is why the app is a commodity and the layer around it is the moat: discovery, measurement, and compliance are what turn a live app into a channel a company can trust with real customers.

The Bottom Line for Services Vendors

Three rivals who agree on nothing all built the same thing: a way for outside companies to sell to their users. That settles the debate about whether AI is a distribution channel. What is left for a services vendor is narrower and more urgent: treat AI as one channel, not three platforms; build the storefront once and instrument it everywhere; get live before competitors become the default the assistant reaches for. The app is the ticket in. The advantage is being early, measured, and compliant when a customer asks. For the wider pattern across industries, see how AI retail distribution predicts the future of services.

FAQ

What does it mean that AI platforms now let outside services in?
It means a company can plug its live systems into the assistant, answer questions, generate a personalized quote, and capture or close a sale where the customer is already asking. For services vendors it turns AI from a place a brand gets mentioned into a place it gets bought.

Do ChatGPT, Claude, and Gemini all let outside services connect?
Yes. ChatGPT has its Apps SDK, Claude has Connectors, and Google connects apps in both Gemini and AI Mode in Search. Several run on a shared open standard, MCP, but the point for a vendor is simpler: every major platform now has a way in, and one build can serve all of them. Together they reach well over two billion people.

Which AI platform should a services vendor build for first?
Sequence by where the buyers already ask. High-intent consumer services lean toward ChatGPT, which has the widest reach and the most developed commerce infrastructure. B2B and agent-driven sales lean toward Claude. Mass-market, discovery-heavy services lean toward Google, where reach comes from AI Mode in Search, both its answers and its new Connected Apps. Because the platforms increasingly work the same way, a vendor extends rather than rebuilds.

How does a service actually reach people across Google?
The surface with real reach is AI Mode in Google Search, where Connected Apps rolled out broadly to US users in July 2026 with partners like Instacart, Canva, and YouTube Music, on top of organic suggestion in AI Overviews. That reaches the one to two billion people Google serves. Gemini Spark, the June 2026 MCP-app release, is separate and gated to premium Google AI Ultra subscribers. One nuance: AI Mode’s Connected Apps are curated and partner-by-partner today. Google has signaled that more apps and developers are coming, but has not confirmed open, self-serve access yet.

If every platform can connect outside services, what is left to compete on?
The app is becoming a commodity. The durable advantages are getting the assistant to suggest the service, measuring the funnel from discovery to quote to lead, and staying compliant as platforms and regulators change. Those are where services vendors win or lose, not in standing up the app.

Is this only relevant for insurance?
No. Insurance moved first because its products are quote-based and fully digital, but the same mechanic applies to any service with personalized pricing or consultation: banking, credit, wealth, utilities, travel, and B2B SaaS with a self-serve motion.