Salesforce Invents New Metrics as 71% of Enterprises Get Locked In

So, What Actually Happened?

Thursday I got stuck on something small and could not put it down. Nasdaq's economists published a piece arguing that ARR no longer measures an AI business, and buried in it was the detail that Salesforce has started reporting something it calls an Agentic Work Unit, a metric it invented because none existed. We scanned 190,000 articles this week so you don't have to. Inside the same forty-eight hours, the Linux Foundation quietly stood up a standards body for AI cost, and an analysis landed saying 71% of enterprises cannot easily switch AI vendors. Three different rooms, one complaint. Two years into buying this stuff, there is still no agreed way to count it, price it, or leave it.

The Bottom Line: You cannot negotiate what you cannot count, and right now every vendor is handing you a different ruler.

The Tracks That Matter

1. Software Companies Are Inventing Their Own Units of Work

Nasdaq's economic institute put the problem plainly: the old formula was sell software to more people and grow revenue, and AI breaks that formula. So the vendors started building their own rulers. Salesforce now reports an Agentic Work Unit. ServiceNow says its first-line agent closes 80-85% of service requests without a human. Intuit counts hours returned, twelve a month from its accounting agent. Every one of those numbers is honest. None of them convert into each other, which leaves your procurement team comparing three vendors on three scoreboards with no arithmetic between them. One layer down it gets worse, because the engineers running model gateways now warn that a model name is not a stable unit of quality: the same model behaves differently depending on who hosts it.

Here's what works: Before your next renewal, ask each AI vendor what one completed task costs. Whoever cannot answer is selling you seats with new labels.

2. AI Cost Just Got Its First Standards Body

The Linux Foundation launched the Tokenomics Foundation this week with 30 founding member organizations and one unglamorous job: agree on how AI spend gets measured. Standards bodies are boring until you remember what forming one signals. Nobody builds one while a market is small, and nobody builds one while everybody already agrees. They appear when the bills got large enough that finance started arguing with engineering about what a token actually costs, and the honest answer was that unit prices keep falling while the bills keep climbing. That gap is not a billing error. It is what consumption looks like when it scales faster than anyone's ability to attribute it.

Here's what works: Split your AI invoice into model spend and business outcome, on two lines. Any vendor blending them is hiding the ratio between them.

3. Seventy-One Percent of Enterprises Cannot Leave Their AI Vendor

An analysis landed the same week EU AI Act enforcement began: 71% of enterprises cannot easily switch AI vendors. Taken alone, that reads like a procurement grumble. Put the two together and it is a compliance exposure. The Act asks you to evidence how your systems behave, and if you cannot realistically move off a provider, you also cannot credibly threaten to, so you are asking for that evidence from a position with no leverage. Healthcare teams are meeting this first, where high-risk classification reaches ordinary software that nobody bought as a regulated system. Lock-in used to be a cost problem. It just became a filing problem.

Here's what works: Cost out a full provider switch for one live AI system this quarter. That number is your actual negotiating position, whether you use it or not.

Quick hits:

• Consulting bought its way into the agent business. AlixPartners acquired agentic AI consultancy Artium, which tells you the delivery layer is now valuable enough that restructuring advisors want to own it rather than subcontract it.
• Zero trust arrived for agents, not just networks. DXC and Primary launched an AI-native zero-trust platform, pushing controls down to the agent layer, where most enterprise security still quietly assumes a human is doing the typing.
• Voice AI got ninety-five percent cheaper. Smallest.ai raised $13 million on a cost collapse that took synthetic speech from 20 cents to 1 cent per minute, which is the move that turns a demo into a call centre.

Signal vs. Noise

🟢 Signal: security controls for AI agents. Thursday was a buying day rather than a talking day. DXC and Primary shipped a zero-trust platform built for agents, and a working comparison of guardrail platforms across four architectural layers landed for enterprise buyers the same morning. These teams have stopped asking whether agents are safe and started asking which layer catches them.

🔴 Noise: ”agentic AI”. It pulled one of the heaviest volumes of any idea on Thursday while what actually attaches to it kept thinning. The tell is the range the label now covers: a chatbot with a database connection, a fully autonomous procurement system, and a consulting practice. When one word means three things, it has stopped being a category and become a sales word.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

A major software vendor started reporting a unit of work it had to invent, the Linux Foundation stood up a body to standardise what AI costs, and an analysis found 71% of enterprises cannot switch providers. All inside the same forty-eight hours.

Each of those gets filed by a different desk. Enterprise software covers the metrics story. The open-source and FinOps press covers the standards body. Compliance covers the EU AI Act survey. Read them on one page and they describe a single condition: the industry is now buying something it cannot count, cannot benchmark against a reference price, and cannot walk away from. Those three failures are the same failure in different clothes. Markets get a shared unit, a reference price and a working exit at roughly the same moment, because each one depends on the other two.

What changes on Monday is smaller and more useful than it sounds. Any AI contract you sign this quarter is being written in the window before those three exist, which means the defaults in the paper are the vendor's defaults. You are not going to fix the industry's measurement problem this year. You can decline to commit multiple years to a price expressed in a unit only the seller defines.

By The Numbers

• ServiceNow's first-line AI agent closes 80-85% of service requests without human intervention — the company's own disclosure, and the clearest sign yet of a vendor measuring itself in work finished rather than seats sold.
• 71% of enterprises cannot easily switch AI vendors — published as EU AI Act enforcement started, which turns a procurement inconvenience into a documented compliance risk.
• The Tokenomics Foundation launched with 30 founding member organizations — the Linux Foundation's answer to the fact that nobody can agree what a single unit of AI spend is.
• Smallest.ai cut text-to-speech costs from $0.20 to $0.01 per minute — a 95% drop that moves real-time voice out of pilot budgets and into thin-margin operations.
• A school district signed a $57,000 AI contract for a tool third-graders used ten minutes in two weeks — the distance between AI procurement and AI usage, in one line item.
• See what's rising in our 190K-article corpus this week →

Deep Dive: Nobody Agrees What One Beat Is

When I started mixing, everything ran on BPM. Beats per minute, counted off the first bar on your fingers, and once you had it you knew exactly where you were inside a record you had never heard before. Two DJs from two countries could argue about taste all night. They never argued about the count. The count was the one thing everybody shared.

Everyone brought their own counter
Salesforce reports Agentic Work Units. ServiceNow reports the share of tickets closed without a human. Intuit reports hours handed back per month. Each is internally consistent and probably true. None of them converts into another, so a buyer running all three through one procurement cycle is holding three scoreboards and no exchange rate.

The bill arrives in a unit you did not pick
The invoice, meanwhile, is denominated in tokens, and tokens keep getting cheaper while totals keep climbing. That is why thirty organisations sat down together this week. Not because AI cost is hard to calculate, but because there is no agreed thing to calculate it in. Finance is auditing a number that engineering defines.

And you cannot walk out to check the price
Price discovery needs a door. When most buyers say they cannot realistically move providers, the numbers on the table stop being prices and start being invoices. The EU AI Act will make that visible fast, because evidence you cannot source anywhere else is evidence you will pay whatever is asked.

What Actually Works

1. Make each vendor name one unit: what does a single completed task cost, in writing, held to at renewal.
2. Separate model spend from business outcome: two lines on the invoice, always. The ratio between them is the whole story.
3. Rehearse one exit a year: cost a real migration for one live system. You do not have to do it, you have to know the number.
4. Give AI spend a named owner in finance: the models have owners already. The invoice usually does not.

You can dance to a track without knowing the BPM. You cannot mix two of them.

What's Coming

Your Renewal Turns Into an Exit Test

The EU AI Act's high-risk classification is reaching tools that were bought as ordinary software, and the enforcement calendar does not care when you signed. Expect legal to start asking for portability terms procurement has never written before. The first renewal where a large buyer actually walks will be reported as a vendor loss. It will really be a pricing event.

Public Compute Becomes a Credible Second Quote

New York brought Empire AI's beta fully online with half a billion dollars behind it and the federal government taking notes. State-funded compute is not going to run your production workload. It does something more useful: it gives research-heavy buyers a real alternative to name out loud in a negotiation.

Reliability Overtakes Capability as the Buying Question

Analysts are already circling the AI reliability trap, where systems demo beautifully and fail in the seams between them. Once cost has a unit and exit has a price, the only question left is whether the thing holds up on a bad day. That is where the 2027 budget arguments get decided.

For Your Team

Monday's meeting prompt: ”Take our three largest AI vendors. What unit does each one bill us in, and what unit does each one report value in? Wherever those two differ, who benefits from the gap?”

Share-worthy stat: A school district signed a $57,000 contract for an AI platform that third-graders ended up using for roughly ten minutes across two weeks. The distance between AI procurement and AI usage is not a big-company problem. It belongs to everybody.

Go deeper: Track where AI spending is moving and what it is actually buying →

The Track of the Day

”For decades, the formula was simple: sell software to more people, grow revenue. Artificial intelligence breaks that formula.”
Nasdaq Economic Institute

Nobody has replaced the formula yet. That vacuum is the story this quarter, not the model releases.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: August 7, 2026 | Curated by Yves Mulkers @ Ins7ghts