Before the H2 budget call: real AI win or expensive pilot?

A split editorial illustration contrasting a polished AI product demo glowing on a screen with a quiet office whiteboard showing a single measured workflow metric, rendered in muted navy and gold tones.

AI budgets keep climbing while measured returns lag. Before the H2 budget call, here is how to tell a real AI win from an expensive pilot: a named owner, a number a CFO can see, and honest running costs including the humans still in the loop.

TLDR

AI budgets keep climbing while measured returns mostly have not arrived, and the Q3 budget call is where that gap finally gets settled. The test for a real win is not a great demo or a big agent count. It is a named owner, a number a CFO can see, and honest running costs. The most disciplined move right now is refusing the circular bet of funding next year on savings nobody has actually banked.

I spent part of this week reading a Bain survey of 951 companies, each pulling in more than 100 million dollars a year. One finding stopped me cold. Budgets keep climbing. The returns mostly have not shown up. Bain’s own summary of the moment was about as blunt as a consulting firm gets: the technology worked, the value didn’t arrive.

That sentence is going to sit in a lot of Q3 budget meetings. Because right now, in most companies, a finance lead is building the second-year AI number, and a CEO is about to defend it to a board that has stopped clapping for pilots. The question on the table is no longer whether to invest in AI. It is which of last year’s AI bets earned another year of money, and which were expensive demos wearing a production badge.

4%
of companies achieved AI savings above 30%, in Bain's survey of 951 firms

How most AI pilots reached the H2 budget call

For about eighteen months, the entire market told them the same story: get agents into production, and the value follows. This past week alone showed how loud that message has gotten. Aible wired NVIDIA’s Nemotron 3 Ultra into its agents to, in the vendor’s own words, improve first-run planning quality and reduce costly retries. Cognizant and Snowflake expanded a partnership built to compress delivery timelines for enterprise workflows. Buzzy shipped a builder update it says shortens the path from prototype to production. Meta launched a whole Business Agent platform. Every one of those announcements is real, and every one is optimistic by construction, because that is what announcements are for.

Sitting above all of it, the headline this week was Anthropic, whose IPO CNBC called the first real test of whether the AI boom’s valuations hold, on a reported 47 billion dollar revenue run rate, up from about 10 billion a year ago. So the capability is not in doubt. The models work. The platforms ship. The tooling to get something live has genuinely never been better.

And that is exactly the trap. When shipping gets easy, shipping stops being the proof. A year ago, getting an agent into production was an achievement. Today it is a Tuesday. Which means the thing a board used to treat as evidence of progress, we put it in production, now tells the board almost nothing about whether it actually works.

Key Insight

When getting an agent into production becomes easy, "we shipped it" stops being evidence of value. The board needs a different test, and so do you.


Where AI pilots actually break before production

The break is more subtle than a failed pilot. Bain found that 44 percent of companies funding their next wave of AI spend are drawing on savings from prior automation programs. Savings that, for many of them, never actually landed. Here is the number underneath that pattern, in Bain’s exact words.

"Nearly 40% of companies that measured AI cost savings landed below 10%, despite targeting 11% to 20%."

Bain & Company, global AI survey of 951 firms, June 2026

Read that twice. Companies are pre-spending returns they have not earned, to buy the next round of tools that are supposed to earn them. Bain’s phrase for the pattern is the most useful thing I have read about AI budgets all year.

Self-funding the next wave from past returns sounds like discipline. In reality, it is a circular bet with a structural leak.

The other half of the breakage is quieter. For all the talk of autonomous agents, Bain found only 7 percent of companies run fully autonomous agents in production today, and 38 percent still require a human to approve every decision. That is not a failure. It is reality colliding with the slide deck. The agents that work are usually narrow, supervised, and pointed at one expensive workflow, not the company-wide digital coworker the demo promised.

So what actually separates a real win from an expensive pilot? Not demo quality. Not how many agents are running. I have watched teams brag about forty agents in production and then fail to name a single number that moved. The real wins share three boring traits: a named owner accountable for one workflow, a measured before-and-after that shows up somewhere a CFO can see it, and honesty about the running cost, including the human supervision Bain just reminded everyone is still there. The expensive pilots have a champion, a great demo, and a budget line that quietly assumes savings nobody has banked yet.

Real win vs expensive pilot: what to check before the re-fund
SignalReal winExpensive pilot
OwnershipOne named owner per workflowA champion and a committee
EvidenceBefore-and-after a CFO can seeA demo and a testimonial
Cost honestyCounts human supervision and retriesAssumes savings not yet banked
ScopeOne workflow, measured"Company-wide transformation"

The pattern: AI capability sprints, value lags

AI adoption has quietly split into two timelines. There is the capability timeline, which is sprinting. Models, agents, platforms, all of it improving faster than most teams can absorb. And there is the value timeline, which moves at the speed of how fast a company can redesign a workflow, assign accountability, and measure the result. Those two timelines are not the same, and 2026 is the year the gap between them turned into a budget problem instead of a research curiosity.

Gartner expects 40 percent of organizations to demote, limit, or retire AI agents because of governance challenges. That sounds grim until the reframe lands. Pruning is not failure. It is what a portfolio is supposed to do. The companies that look smart at the end of this year will not be the ones with the most agents. They will be the ones who quietly retired the agents that were not paying rent and concentrated the budget on the two or three that were.

That is also the calmer way to walk into a board meeting. A short list of bets that each survive scrutiny beats a long list that collectively impresses and individually crumbles. Boards can feel the difference between a leader who is defending momentum and a leader who is defending results.


What I’d tell you before the H2 budget call

If we were getting coffee before your board meeting, here is what I would say. Walk in with a short list, not a long one. For every AI initiative asking for second-year money, be ready to name the owner, the one number that moved, and the real cost including the humans still in the loop. If a project cannot survive those three questions, that is not a tragedy. That is the budget call doing its job.

And resist the circular bet. Do not fund next year on savings that have not actually shown up on the P&L. The capability is real and it is not going anywhere, which means there is time to be disciplined instead of frantic. The companies getting this right are not moving faster than everyone else. They are just refusing to confuse motion with progress. That turns out to be a calmer place to run a company from, and it happens to be the place where the returns finally arrive.

Sources

  1. Your AI Budget Is Growing. Your Returns Aren't. Here's Why. - Bain & Company, 2026-06-01
  2. The Tech Download: Anthropic's IPO sets up first big test of AI boom valuations - CNBC, 2026-06-05
  3. AI Update, June 5, 2026: AI News and Views From the Past Week - MarketingProfs, 2026-06-05
  4. Daily AI Agent News - This Week - AIAgentStore.ai, 2026-06-07

Back to all insights