Vendors are buying the last mile | Mon Jun 22, 2026

Vendors Are Buying The Last Mile

Four enterprise software deals in three weeks. None about a model. All about the machinery that lets an agent actually do something.

Asana paid $75 million for a no-code agent builder most of its customers had never heard of. Coupa is buying a document-reading engine. Salesforce signed for Contentful at a reported $1 billion-plus. The conversation in 2025 was whose model is smartest. The buying in June 2026 is about whose plumbing lets that model touch your ERP, your contracts, and your content without a human in the loop. If you are budgeting for AI this quarter, the question has quietly moved from which model to which execution layer, and the vendors have already answered with their checkbooks.

Enterprise software is buying the verb, not the noun

Coupa is acquiring Rossum, whose transactional model has been trained on tens of millions of documents, to push intelligent document processing into its source-to-pay stack. Salesforce signed a definitive agreement for Contentful, the composable content platform used by more than 4,800 brands, at a reported $1 billion to $1.5 billion. Vertice is buying Vendr to fold SaaS-purchasing data into its agents.

None of these is a model deal. Not one buyer is acquiring a frontier lab. They are all buying the layer between intelligence and action: document parsing, workflow orchestration, content delivery, transactional data. The reason is the part of agentic AI that does not demo well. A model can read your invoice and tell you what it says. Getting it to post the entry, route the approval, and update three downstream systems without breaking anything is a different problem, and it is the one enterprises actually pay to solve.

This is why the build-versus-buy calculus just changed. For two years the assumption was that the model was the scarce asset and integration was commodity glue. The June buying spree inverts that. Models are now abundant, increasingly interchangeable, and dropping in price every quarter. The scarce asset is the validated, governed, audit-logged path from a model's output to a committed action inside a system of record.

The bottleneck was never the model. It was the last mile to action, and that last mile just became the product.

There is a second-order read for CIOs. When your incumbent vendors absorb the execution layer, the best-of-breed agent platform you were evaluating becomes a feature of software you already own. The integration work you scoped as a custom build may arrive as a release note. That is good for speed and bad for leverage: the more the verb gets bundled into the suite, the harder it is to switch suites later. The vendors understand this perfectly. Bundling the execution layer is how you turn an AI experiment into a renewal.

N ^o ###  Sources: The AI Insider · erp.today · Salesforce

The Obvious & The Overlooked

Moving Pieces

The rest of what moved, with the analysis.

Talent

The transformer's co-inventor leaves Google for OpenAI

Noam Shazeer, co-author of "Attention Is All You Need" and a VP at Google DeepMind, has joined OpenAI as Lead for Architecture Research. The sting is the price tag: Google spent an estimated $2.7 billion in 2024 on a Character.AI deal largely to bring him back and install him as a Gemini co-lead. Less than two years later, that investment walks out the door. The enterprise read is not about one researcher. It is that the people who can change model architecture number in the hundreds globally, and they are concentrating at the labs racing to IPO. When you bet on a vendor's roadmap, you bet on its ability to keep names like this.

Sources: The AI Insider · MLQ

Deals

OpenAI follows Anthropic into the channel

OpenAI launched a formal Partner Network on June 14, backed by $150 million, aiming to certify 300,000 consultants by year-end. It is a direct answer to Anthropic's Claude Partner Network, which had already certified more than 10,000 consultants. Both labs now treat ecosystem control as a priority on par with model development, because the bottleneck to revenue is no longer capability, it is delivery capacity. For buyers this is mixed: more certified integrators means easier sourcing, but a certification badge measures vendor loyalty, not implementation skill. Vet the team, not the tier.

Sources: TechTimes · Anthropic

Policy

Trump's AI order trades mandates for a 30-day preview

The June 2 executive order, "Promoting Advanced Artificial Intelligence Innovation and Security," leans into voluntary cooperation over regulation. Frontier developers can give the government up to 30 days of early access to new models before wider release, and the order stands up an AI cybersecurity clearinghouse. The catch for enterprises is in the timelines: many directives carry 30-to-60-day implementation clocks, and they cascade onto anyone who contracts with the federal government. If you sell to or through a federal customer, your AI deployment terms may shift inside a quarter, on the government's schedule, not yours.

Sources: White House · NPR

Infrastructure

Nvidia turns the agent stack into a product line

Nvidia shipped a full agentic stack: the Vera CPU for agent orchestration, Nemotron 3 Ultra (a 500-billion-parameter open-weights model it claims runs about 30% cheaper than comparable frontier models), and NemoClaw, an orchestration framework with templates for multi-agent delegation and tool calls with error recovery. Seventeen software leaders, including Adobe, Salesforce, SAP, and CrowdStrike, are adopting the Agent Toolkit. The strategic move is the open-weights model: by giving away a capable model that runs best on its own silicon, Nvidia keeps the value in compute and orchestration. For buyers, open plus 30% cheaper to run is a real lever against per-token API pricing.

Sources: VentureBeat · Nvidia

Governance

Ten days later, Fable 5 is still dark

Anthropic's Claude Fable 5 and Mythos 5 remain suspended worldwide as of June 21, more than a week after a June 12 US export-control directive ordered access cut for all foreign nationals, which Anthropic could only meet by disabling the models for everyone. Set the policy debate aside and look at the operational fact: a model that enterprises had embedded in production went dark at 5:21pm on a government order, with no notice and no fallback. That is single-vendor concentration risk made concrete. If one provider's model is load-bearing in your stack, you do not have an architecture, you have a dependency.

Sources: Anthropic · Al Jazeera

On the Radar

$725B

The combined 2026 capital-expenditure guidance of Google, Amazon, Microsoft, and Meta, up roughly 77% from last year's record.

Roughly three-quarters of it targets AI infrastructure. Four companies are spending more on data centers and chips this year than the annual GDP of most countries. The execution-layer deals in today's Big Story are measured in the millions and low billions. The compute they ultimately run on is measured in the hundreds of billions. That gap is the real shape of the AI economy.

Source: Tom's Hardware

Everyone is buying execution. Most execution still fails.

The whole industry is racing to own the layer that lets agents act. Here is the inconvenient data underneath the shopping spree. A widely cited RAND meta-analysis of 65 documented enterprise AI initiatives put the failure rate at roughly 80%, with about a third abandoned before production and another quarter reaching production but never delivering the promised value. Gartner's April report found only 28% of AI projects in infrastructure and operations delivering their expected return.

Buying a polished execution layer does not close that gap, because the gap is rarely the technology. It is integration debt, missing governance, and data that was never clean enough for an autonomous system to act on safely. A vendor can sell you the verb. It cannot sell you the organizational readiness to let a machine use it.

So as the bundled execution layers arrive over the next year, treat them as accelerants, not cures. The 20% of programs that succeed share something no acquisition provides: clear ownership, a human who can halt a bad action fast, and data the agent can be trusted to touch. Buy the plumbing if you must. Just do not mistake it for the readiness.

Sources: Gartner · RAND, as reported

For two years the industry sold nouns. The best model, the biggest context window, the highest benchmark. This month it started buying verbs, and that is a more honest accounting of where the work actually is.

A model that summarizes your contract is a party trick your associates already perform. A system that reads the contract, flags the off-market clause, routes it to the right reviewer, and updates the obligation tracker without anyone retyping a thing is a business outcome. The distance between those two sentences is the entire job. It is also where most AI programs quietly die, somewhere between the impressive demo and the procurement system that will not accept the agent's output.

§

The vendors absorbing the execution layer understand this. They are not betting that their model is smartest. They are betting that owning the path from intelligence to action is what enterprises will renew on. They are probably right. But owning the plumbing is not the same as owning the outcome, and a CIO who confuses the two will buy a beautiful execution layer and still land in the 80% that never ships.

§

The unglamorous truth we keep relearning in client work is that the last mile is the whole race. The model was never the hard part. Getting an organization to let a machine act, with the right human watching and the right brake within reach, is the hard part. The market just put a price on that mile. The teams that win this year will be the ones who already did the boring work the price tag assumes.

The verb is for sale now. Whether it works is still on you.

Let's get to production,
AK

The Agentic Enterprise

A daily AI intelligence briefing from Spearhead. Know more about AI than 95% of your peers. Every morning.

Spearhead  ·  Contact  ·  Get The Agentic Enterprise