The AI bill is huge and the bottleneck is now your process

Two threads run through today. Enterprises are committing enormous sums to deploying AI while the heaviest spenders appear to have stopped accelerating, and the limit on value is no longer how fast a model runs but how fast the work around it moves.

For a business choosing tools, both threads point the same way. The price on a licence is a small part of what you will actually spend, and the slow part of any AI rollout is rarely the model.

Where the AI budget actually comes from

Menlo Ventures released its fourth annual State of Enterprise AI report today, and the top-line number is large. Enterprises will spend an estimated $176.8 billion deploying AI in 2026, more than five times the $33 billion spent the year before [1]. The report frames this as deployment spend, which is a wider category than software subscriptions.

The detail worth sitting with is in the title of the piece: only about a fifth of that spend came out of software budgets [1]. That tells you something important about how AI costs behave. The subscription you sign is the visible part. The larger part is the work around it: people to configure it, data to feed it, integration with the systems you already run, and time spent checking its output.

For a buyer, the lesson is to budget the whole job, not the sticker. When you compare two tools, the question is not only which licence costs less. It is which one needs less surrounding work to reach a result you trust. A cheaper tool that demands months of integration can cost more than a dearer one that fits your existing process. Treat the licence as the opening line of the invoice, not the total.

The heaviest spenders have stopped accelerating

Two separate numbers landed on the same point today. Ramp's latest AI Index, built from transactions at more than 85,000 US businesses, shows the heaviest AI spenders peaking in July [2]. On the same morning, Anthropic's amended S-1 showed its quarterly dollar add shrinking for the first time [2].

A plateau at the top is not the same as a decline. It usually means the obvious wins have been captured and the next gains are harder to find. The firms that were going to spend aggressively have largely done so. Growth from here comes from new buyers entering, not from the early adopters doubling again.

That changes your position if you are choosing tools now. You are no longer early. The patterns that work have started to settle, pricing is maturing, and there is a body of other buyers who have already made the mistakes you can now avoid. The upside of arriving later is that you can ask vendors for evidence rather than promises. The trade-off is that the pricing advantages of being an early, hungry customer have mostly passed. Weigh both before you assume that waiting costs you nothing.

Adoption that gets counted

Salesforce reported that more than 185,000 users are now putting its Agentforce Coworker to work, naming The Adecco Group, Fulton Bank, and BACA Systems among the companies expanding adoption [3]. The company describes usage rising 21 times in three months as enterprises delegate real work to AI [3].

Usage figures like this are worth reading carefully, because a user count and an outcome count are not the same thing. A tool that 185,000 people have opened has cleared the first hurdle, which is that people are willing to use it. It has not, by that number alone, told you how much of their work it actually finished or how much of that work was checked afterwards.

When a vendor leads with adoption, the useful follow-up question is what the agent is allowed to decide on its own. Delegation is the right word for it: you are handing a task to something that will act. The line you draw around that task matters more than the headline number. We have written before about the decisions automation should never make, and the arrival of agents that act across a business makes that line more important, not less.

The bottleneck moved from the model to the process

Salesforce published a piece today with a sharp image. Picture a race car flying down a country highway, only to find itself stuck behind a tractor [4]. The car's engine is spectacular, but it cannot go any faster than the tractor [4].

That is the shape of the problem most buyers now face. The model is fast. The approval step in front of it is not. The agent can draft a reply in a second, but the reply still waits two days for a manager to sign off. The agent can pull a quote together instantly, but the quote still sits in a queue because that is how the queue has always worked.

The teaching point is that speeding up one stage of a chain does nothing if the slow stage is somewhere else. Before you buy a faster tool, map the actual path a piece of work takes from start to finish and find where it waits. Often the wait is a handoff, an approval, or a step that exists for a reason no one remembers. Fixing that costs little and frees the speed you already paid for. This is a recurring theme in how we read these announcements, and it ran through yesterday's briefing as well.

When the tool comes to you

Zapier published an explainer today on Claude Duet, which it describes as Anthropic's answer to Meta's Muse and OpenAI's dots [5]. The common thread across these products is a shift from pull to push. Instead of you opening a tool and typing a prompt, the tool is always on and reaches out to you.

This is a real change in what "using a tool" means. A pull tool does nothing until you ask. A push tool decides when to speak and what to raise. That can save attention, because the morning summary arrives without you going to fetch it. It also moves a decision to the software: what is worth interrupting you for, and what is not.

The trade-off to watch is scope. An agent that only answers when asked can do little harm by staying quiet. An agent that acts on its own needs a clear boundary around what it may touch and a clear record of what it did. If you adopt one, decide up front which actions it can take without you and which it must bring to you first. The convenience is genuine. So is the need to keep the boundary tight.

What to take from the day

The spend figures and the slowdown read differently at first, but they agree. Money is flowing into AI, the early rush is maturing, and the gains left on the table are in the work around the model rather than the model itself. A buyer who maps their process, budgets the whole job, and draws a firm line around what an agent may decide is reading the day correctly. A good tool rewards all three. 360REV is built so that the record of what was decided and by whom stays in one place, which is where most of these questions end up pointing.