AI budgets are growing, and finance teams reject AI requests more often

AI vendors took 8.1% of enterprise software spending in the latest year, up from 1.4% the year before, according to purchasing data Zip published on September 22. Zip runs procurement software, and the numbers come from about $18 billion in approved purchase requests across its customers.

The line that caught my attention was a different one. In the same period, 21% of purchase requests for AI vendors were rejected or canceled. For every other kind of vendor, the rate was 12%.

So companies are spending far more on AI, and their finance teams are saying no to AI requests almost twice as often as everything else. Both things are true at once. That combination tells you where most companies are on returns right now.

DataPurchase requests rejected or canceled

• AI vendors21%
• All other vendors12%

The money is concentrating in a few tools

Zip’s data shows a few more things worth knowing if you sign off on software.

84% of the companies it observed raised the share of software spend going to AI. The average active adopter now works with 5.9 AI vendors, up from 2.0 two years earlier. Yet the spending is concentrated: a company’s largest AI vendor gets 68% of its AI spend on average, and its top three get 95%. Across the whole group, Anthropic, Cursor and OpenAI took 74% of AI vendor spend.

Read that as a pattern. Companies try a lot of AI tools, then put nearly all their money behind two or three. The rest are trials and seat licenses that someone requested and few people use. Zip says its data shows which requests got rejected, not why, so I won’t pretend to know the reasons. My guess is that finance teams are getting better at spotting the fifth AI tool that does what the first four already do.

DataAI vendors' share of software spend

• Prior year1.4%
• Latest year8.1%

Most companies still can’t find AI in their profits

Spending data tells you what companies bought. It doesn’t tell you what they got. For that, the best recent number comes from McKinsey.

In its 2026 State of AI survey, published in August from 1,719 responses, 37% said AI had contributed positively to their organization’s EBIT. That figure barely moved from the year before. The group McKinsey calls high performers, who attribute at least 5% of EBIT to AI, held at 6% of respondents.

What those high performers do differently is specific. Nearly three-quarters of them have fundamentally redesigned workflows around AI, compared with about a quarter of everyone else. They are also twice as likely to have defined processes for measuring impact.

There is some longer-run evidence that the effort pays. An August NBER working paper by Tania Babina and co-authors, summarized by ITIF on September 8, found that firms growing their share of AI workers saw roughly one extra percentage point of productivity growth per year from 2018 to 2024. The mechanism the authors point to is what they call organization capital: the processes and systems a company builds around the technology. In their data, 58% of AI jobs built that kind of capital, against 31% of other jobs.

Put the three sources together and the story is consistent. Companies are buying AI fast, and finance teams are getting pickier about each new tool. Returns show up when a company changes how the work is done.

By the numbers

37%

of companies say AI has added to EBIT, essentially unchanged from a year earlier

A tool you bought is not a workflow you changed

Say a company buys an AI assistant for 40 seats. People use it to rewrite emails and summarize meetings. That is useful, and it saves people a few minutes here and there. But nobody’s job changes, nothing gets measured, and when the renewal comes up, nobody can say what it returned. That is the request a finance team learns to reject.

The other version starts from the workflow. Someone decides that a specific piece of work, say the weekly ad refresh or the first draft of every product fix, will now be done by an agent, with a person approving the output. The job changes. The hours change. You can count them.

My newest company, FinalPeace, runs this way. AI agents do the work of a ten-person team across development and marketing, covering ads, emails, landing pages and product fixes. They send finished work to an approvals queue, and I approve what ships in about half an hour a day. The software isn’t what makes that work. The decision to redesign the work around an approval step is.

That is also the answer to the finance team. “We bought an AI tool” gets rejected. “This workflow used to take this many hours, and now it takes this many” is the request that gets approved.

Four questions to ask before approving the next AI tool

If you are the person who signs, the rejection numbers suggest your finance team already has instincts here. Give them a short list so the instincts become a standard. These are the four I would ask of any AI purchase request, large or small.

Which workflow does it change? The answer should be a specific piece of work with a name, like “first-pass replies to warranty claims,” and not a department or a vague goal.

How long does that work take today? If nobody has timed it, the request waits until someone does. It takes a day of logging, and it is the only way anyone will know later whether the tool paid for itself.

Who reviews what it produces? Output nobody checks is a risk. When a named person already reviews the work, the tool slots into a step that exists, and that is where a redesigned workflow starts.

What does it replace? If the company already pays for an AI tool that could do the job, the new request should explain why that one won’t. Zip’s data says most companies end up spending almost everything on two or three vendors anyway.

A request that answers all four is easy to approve. A request that can’t answer the first one deserves to land in that 21%.

What to do this week

1. Pull every AI subscription and seat license your company pays for. Your card statements and expense reports will catch the ones procurement never saw.
2. Next to each one, write the specific workflow it changed. If you can’t name one, mark it.
3. For the marked tools, ask the person who requested each one what they use it for. Cancel or consolidate where two tools do the same job.
4. For the tools that did change a workflow, write down the before and after hours. That is your evidence for the next budget conversation.
5. Pick one workflow that should be redesigned around an agent and a review step, and give it an owner and a date.

Sources

• Where the AI budget is actually going (Zip, September 22, 2026)
• The state of AI in 2026: On the road to ROI (McKinsey & Company, August 2026)
• Fact of the Week: Firms Investing More in AI Experienced Roughly One Additional Percentage Point of Productivity Growth Per Year (ITIF, September 8, 2026)