The Customer Is a Machine Now

agents coordination enterprise governance

On November 3, Menlo Ventures published its annual State of Enterprise AI report, and the headline number is not the one most coverage will lead with. Enterprise spending on AI deployment reached $176.8 billion by September — 5.4 times the year before. Companies spent, on average, two and a half times what they had budgeted. But adoption — the share of companies using AI at all — rose only from 55 percent to 65 percent.

Read those two numbers together and a different story appears. This was not a year in which many more organizations joined. It was a year in which the ones already inside went much, much deeper. And the report names what drove the depth: agents. On OpenRouter, one of the few places agent traffic is directly observable, agents now generate 80 percent of all tokens — about seven times human-driven usage, up thirty-seven-fold since agent traffic overtook human traffic in February.

The customer changed before anyone rewrote the contract. For two decades, software economics assumed a human at the end of every transaction: a person who reads the invoice, notices the overage, gets tired, goes home. An agent does none of those things. It does not experience a budget as a constraint; it experiences it as a parameter, if it experiences it at all. It will happily burn at 3 a.m. what a team would have debated at 3 p.m. The two-and-a-half-times budget overrun is not evidence that companies were careless. It is evidence that their cost controls were built for a customer who sleeps.

This is why the most telling product launch of the same day came from a security company, not a model lab. At its Fal.Con Europe conference, also on November 3, CrowdStrike shipped an "AI Kill Switch" — a control that lets an administrator halt an internal agent instantly, block it from restarting, and cut its network access. In the same breath it agreed to acquire Fiddler, which monitors agent behavior, and Dreadnode, which tests AI systems before deployment, for a reported $495 million combined. Governance is arriving, as it usually does: not as a principle first, but as a product category with a price tag, bought after the invoice made the abstract problem concrete.

It would be easy to read the kill switch as an admission of failure — we built agents we now need an emergency brake for. That reading is too cheap. Every mature technology gets a brake eventually, and the brake is not the opposite of the accelerator; it is what makes sustained speed survivable. The honest question is not whether agents should have kill switches. It is why the brake is being sold separately, by a third party, a year after the spending curve went vertical — rather than arriving as a standard fitting on the agents themselves. A car whose brakes are an aftermarket security product would tell you something about the car industry. It tells you the same thing here.

There is a quieter disclosure buried in the same news cycle, and it completes the picture. On November 2, Anthropic amended its IPO filing to say that its future models may imminently meet its own threshold for "dramatically accelerated progress attributable to AI" — and that the company may not notice the acceleration until after it has happened. Set aside the capability question. Notice the shape of the admission: the builder is telling investors, in a legal document, that its detection may lag its product. Spending that outruns budgets, agents that outrun their reviewers, progress that may outrun its own measurement — these are not three stories. They are one story told at three altitudes: our instruments were calibrated for a slower, more human system, and the system has changed speed.

None of this argues for stopping. The same week's other ledger is real: India's renewable build-out saved the country more than $100 billion in fuel imports over fourteen years, a reminder that a technology adopted at scale, with boring persistence, quietly rewrites national accounts. Scale is not the villain. Unmeasured scale is. The fix is not a smaller ambition; it is instrumentation that matches the customer we actually have now. Budgets that bind at execution, not at the quarterly review. Approvals that expire when the context does, instead of riding along in a queue. Verification priced into the deployment instead of invoiced after the incident. A kill switch on every agent — not because we expect to use it daily, but because a system you cannot stop is not a system you control, whatever the dashboard says.

The machines are the customers now. They will be excellent customers in every way except one: they will never, on their own, ask whether the bill makes sense. That question remains ours — which means it has to be built into the machinery, asked automatically, at machine speed, every time. The organizations that learn to ask it in the architecture, rather than in the postmortem, are the ones whose 5.4x year will look, in retrospect, like a beginning instead of a warning.

agents coordination enterprise governance

Where this sits

Step 9 of 15 in Agents in the Wild.

Concepts: Alignment as infrastructure

Comments

Loading…

Leave a comment

Comments are moderated — yours appears after approval.