Four AI agent platforms launched in a month. None agree what it is

Article Brief

Key Takeaways

4 Points24s Read

1. 01Four in a month–OpenAI Presence, Meta Business Agent Platform, NVIDIA/ServiceNow Project Arc, and Google Gemini Enterprise all launched within roughly a month.
2. 02No shared definition–They optimize different control points — distribution, containment, managed services, and data gravity — so the category is four adjacent markets sharing a buzzword.
3. 03Governance is table stakes–Every platform now leads with guardrails and audit; control language stopped being a differentiator and became the entry fee.
4. 04Read the pricing–Each commercial model — volume, services margin, license expansion, consumption — reveals the real strategy and the intended buyer.

Four of the largest companies in technology have stood up enterprise AI agent platforms in roughly a month, and the striking thing is that no two of them agree on what the product even is. OpenAI shipped a governed deployment service. Meta shipped a distribution channel. NVIDIA and ServiceNow shipped a secured desktop runtime. Google Cloud shipped a model-and-data layer. Same category name, four incompatible theories of where the value actually sits.

That disagreement is the most useful signal a buyer will get this year. When incumbents converge on one design, the market has decided what matters. When they scatter like this, it means the control point of enterprise agents — the place where the money and the risk concentrate — is still genuinely unsettled. Reading the four bets tells you more than any single product demo.

Four platforms, four theories

Start with what each company actually built, because the architecture is the argument.

Meta’s bet is distribution. The Meta Business Agent, whose platform opened on July 1, does not ask a business to adopt a new surface — it puts the agent inside WhatsApp and Messenger, where Meta says over a million businesses already run agents across a billion active daily threads. It connects to “hundreds of systems like Shopify, Zendesk, and Shopee,” sets up “within minutes,” and starts free before moving to subscriptions. The thesis is blunt: the model matters less than being where the customer already is. Reach is the moat.

NVIDIA and ServiceNow bet the opposite way — on containment. Project Arc is a long-running desktop agent that can touch local file systems, terminals and installed applications, wrapped in three governance layers: an AI Control Tower for auditability, Action Fabric for workflow connection, and OpenShell, an open-source runtime that lets enterprises “define what agents can see, which tools they can access, and how actions are contained.” ServiceNow’s Jon Sigler framed it as “delivering the governance and security that enterprise AI requires.” The thesis here is that capability is assumed and control is the product.

OpenAI’s Presence, which TECHi examined as a buyer’s checklist last week, is a third theory: managed accountability. It packages policies, guardrails and graded simulations, but ships only through Forward Deployed Engineers and integrators — no self-service. The bet is that enterprises will pay for turnkey governance installed by the vendor’s own people, a services engagement with software inside.

Google Cloud’s expanded Gemini Enterprise is the fourth, and the most conventional: lead with the model and the data gravity of the cloud the enterprise already runs on. The bet is that proximity to the data, and a frontier model attached to it, pulls agent workloads along.

Why the scatter matters more than any winner

It is tempting to ask which of these wins. That is the wrong question this early, and the four-way split is exactly why.

Each platform is strongest for a different buyer, because each optimizes a different failure mode. A retailer whose customers live in WhatsApp threads has a distribution problem, and Meta’s answer removes it. A bank that cannot let an agent touch a terminal without an audit trail has a containment problem, and Project Arc’s runtime is built for precisely that. A mid-market firm with no platform team has an integration problem, and Presence rents it the engineers. An enterprise already all-in on one cloud has a data-gravity reality, and Gemini Enterprise meets it there.

None of these is a general answer, and the vendors’ own designs admit it. When four sophisticated companies solve four different problems and call the results the same thing, the honest read is that “enterprise agent platform” is not yet one market. It is four adjacent markets sharing a buzzword, and they will stay adjacent until one control point proves decisive enough that the others reorganize around it.

The governance thread they all pull

For all the divergence, one element shows up in every pitch: control language. Guardrails, audit, approved actions, containment, measurement. A year ago the enterprise-agent conversation was about capability — what the agent could do. In July 2026 every serious platform leads with what the agent is prevented from doing.

That convergence is the real maturation signal, and it did not come from the platform vendors first. It came from the tooling layer beneath them — the world of execution hooks that gate an agent mid-run and toolkits that wrap production controls around agent actions. What the platforms are doing now is absorbing those controls as defaults rather than leaving them for the buyer to assemble. OpenShell productizes containment; Presence productizes policy simulation; Meta productizes rule-setting; Control Tower productizes the audit trail. The controls stopped being a project and became a feature.

For buyers, that is genuinely good news with a catch. The good news is that the floor for safe deployment rose across the whole category at once. The catch is that “has guardrails” is now table stakes, which means it tells you nothing when comparing platforms — every box is checked. The real comparison moved underneath, to questions the marketing pages do not answer: whose containment survives an adversarial user, whose audit trail satisfies your actual regulator, and whose controls you can export when you leave.

The commercial model is the strategy, spoken aloud

If the architecture is the argument, the pricing is the confession. How each company plans to charge reveals which theory it actually believes, stripped of the launch language.

Meta starts free and moves to subscriptions “for businesses of every size.” That is a volume play — the economics only work at the scale of a billion daily threads, which is exactly the asset Meta already owns and no rival can rent. Free-to-paid is how you monetize distribution you have rather than capability you sell. OpenAI’s Presence, by contrast, ships exclusively through Forward Deployed Engineers with undisclosed pricing; that is not a product price, it is a services quote, and the opacity is a feature of a high-touch, high-margin motion aimed at large accounts that expect a deployment team. Two companies, two commercial models, two entirely different definitions of who the customer is.

NVIDIA and ServiceNow sit inside a third model again: Project Arc extends platforms enterprises already license, so the agent is less a new purchase than an expansion of an existing governance relationship — sold to the CIO who already bought the control plane. Google’s Gemini Enterprise rides cloud commitments, monetizing agent workloads as consumption on infrastructure the customer is already metered against. Distribution rent, services margin, platform expansion, consumption pull — four revenue theories, and each one predicts a different sales conversation, a different buyer in the room, and a different reason the deal closes or dies.

The lesson for a buyer is to read the price model as a statement of intent. A vendor whose commercial motion depends on locking workflow into a channel it controls will optimize for that lock-in whatever the demo shows. A vendor selling services margin will resist the self-service transparency that would let you compare it cleanly. The pricing page, not the product page, tells you what the platform is really built to do to your account over three years.

What a buyer should take from a fragmented market

The practical response to a four-way split is not to wait for a winner. It is to diagnose your own control point first, then buy the platform that optimizes it.

• Name your dominant failure mode before shopping. Distribution, containment, integration capacity, or data gravity — one of them is your real constraint. The platform that removes your constraint beats the platform with the best benchmark, every time.
• Treat “guardrails” as a floor, not a feature. Every entrant has them. Push each vendor to demonstrate a control on a scenario your team writes, because that is where the identical-sounding pitches diverge.
• Price the exit alongside the entry. Meta’s agents live inside Meta’s channels; Presence’s policies live inside OpenAI’s system; Project Arc’s containment lives in ServiceNow’s platform. Each is sticky by design. Ask what you keep if you leave, in what format, before you commit the workflow.
• Expect consolidation, and buy for two years, not ten. Four theories will not all survive. Favor platforms whose lock-in you can afford to unwind, because at least two of these bets will be reorganized around whichever control point wins.

The honest bottom line

The enterprise agent platform is being invented in public by four companies that disagree about its nature, and that disagreement is the most valuable disclosure of the season. Meta is betting on reach, NVIDIA and ServiceNow on containment, OpenAI on managed accountability, Google on data gravity. Each is right for a specific buyer and wrong as a universal answer, which is why the category still has four shapes instead of one.

The convergence on control language underneath the divergence is the signal that the market is maturing — governance is no longer the differentiator, it is the entry fee. The differentiation moved to the control point, and the control point is exactly what has not been settled. Buyers who internalize that will stop asking which platform is best and start asking which one is built for the problem they actually have. That is the only version of the question the market can currently answer, and it is enough to buy well in a market that has not finished deciding what it is.