Which AI line item survives your Q3 budget review
At a Q3 budget review the AI spend that survives is the spend an operator can tie to a P&L or risk outcome and shut off cleanly. Two reports this week show why, and what discipline actually answers the cut question for a Series B.
At a Q3 budget review the AI spend that survives is not the cheapest line. It is the line an operator can attribute to a real outcome and switch off without breaking anything. Two reports out this week reframe the cut decision around trust and cost attribution, and the practical test is three questions per agent: who owns it, what outcome does it move, and can we reverse it.
I spent part of this week sitting with a Series B operator who had printed out their AI spend on actual paper. Highlighter in hand. Going line by line, asking the same question of each one: if I had to defend this to the board in September, could I? About half the lines got a confident yes. The other half got a long pause. And the long pauses, not the dollar amounts, were the real story.
That is the budget review most Series B companies are walking into this quarter. Not a cut for the sake of cutting. A reckoning with the lines nobody can quite explain. Two reports landed this week that, read together, tell you exactly which lines are going to survive and which ones are quietly on the chopping block.
What the AI spend review actually looks like now
The first report came out Wednesday. Kore.ai released its 2026 Agent Productivity Index, a survey of more than 400 IT business leaders, and the headline number is the kind of thing that makes a CFO put down their coffee. As the index put it: “72% say their agents introduce unmanaged financial or compliance risk.” Worth saying plainly that Kore.ai sells agentic-governance tooling, so they have a reason to highlight the scary number. But the supporting figures are harder to wave away, and they paint a specific picture.
The same survey found that 79% of enterprises have already had to reverse an action an AI agent took. 70% have hit a failure their teams could not trace. And the one that turns a productivity conversation into a finance conversation: 42% report lost revenue tied to an AI agent failure. That last figure is the hinge. The moment an AI line item is associated with lost revenue, it stops being a productivity bet and becomes a P&L question. And P&L questions get asked at budget reviews.
Here is the part that explains why this is happening right now. In the first half of 2026, companies handed agents real authority. The index found 41% of agents are running data migrations and system updates, 26% are approving or denying decisions, and 15% are acting on financial transactions. Read that last one twice. Roughly one in seven deployed agents is touching money. The spend grew. The authority grew faster. The ability to account for either did not keep up.
The second report is the more useful one for the actual budget meeting, because it comes from outside the vendor-survey world. SiliconANGLE published theCUBE’s coverage of FinOps X on Thursday, and the through-line is that AI cost is a different animal than cloud cost. Marco Meinardi, a VP Analyst at Gartner, framed it cleanly: “With AI, now we are dealing with spending sources that are even outside of our organization.” That is the quiet problem behind half the unexplained lines on that printed sheet. The spend is dynamic, driven by usage you do not fully control, sometimes by users who are not even your employees.
Where the budget conversation breaks down
So you have two forces meeting in the same room. Spend that is hard to predict, and authority that is hard to account for. The place it breaks is attribution.
The FinOps coverage named it without flinching: there is no universally accepted method today to attribute AI token spend to a business outcome. Think about what that means for the person with the highlighter. They can see the bill. They cannot reliably see what the bill bought. And a cost that cannot be connected to an outcome cannot be defended, which means at a budget review it loses by default to a line that can.
This is the trap a lot of Series B teams built for themselves without noticing. The business case for the first agent focused on token spend and stopped there. I have seen this pattern across enough deployments to recognize it on sight. The pitch was “this will be cheaper than a contractor” or “this will save the team ten hours a week,” and then the thing shipped, the token meter started running, and nobody went back to check whether the ten hours actually showed up anywhere a CFO could find them.
That is the work the long pauses are pointing at, and it is not glamorous work. It is plumbing.
The AI line items that get cut are rarely the most expensive ones. They are the ones nobody can attribute to an outcome. A cheap, unexplained line loses to an expensive, defensible one every time.
There is a more hopeful reading here, and it is the one I gave the operator with the highlighter. The lines getting the long pause are not failures. They are unfinished measurement. The agent might be doing genuinely useful work. The gap is that nobody built the wire from “agent did a thing” to “here is where that shows up in the numbers.” That wire is buildable. It is mostly boring, and boring is good, because boring is what survives a budget review.
Spend follows authority, control follows last
Step back and the two reports describe one shape. Companies gave agents authority over real work faster than they built the means to control or account for it. Kore.ai measured the authority side. FinOps X measured the cost-and-control side. Same gap, two angles.
The FinOps people had a phrase for the discipline that closes it, and it reframes the whole budget question. Parker Nancollas, a Global FinOps practice lead at SoftwareOne, said: “Models need to be looked at as a consumable portion of what we are building that is replaceable.” That is a quietly radical idea for anyone who fell in love with a specific model. The model is not the asset. The workflow and the outcome are the asset. The model is fuel, and fuel is swappable when a cheaper or better one shows up.
Hold that frame and the budget review changes character. The defense is no longer “our spend on Model X.” It becomes “this workflow, which produces this outcome, currently fueled by Model X but portable if the economics shift.” The first is a sunk-cost argument. The second is an operating decision. Boards fund operating decisions and cut sunk-cost arguments. Same money, completely different meeting.
The model is not the asset. The workflow and the outcome are the asset. The model is fuel, and fuel is swappable.
This also explains why the macro spend keeps climbing even as scrutiny rises. Gartner pegged worldwide AI spending at 2.59 trillion dollars for 2026, up 47% on the year, and noted that the predictability of return has to improve before AI scales further. The spending is not the problem the market is solving. The attribution is. Series B companies that build the attribution wire early are not just protecting this quarter’s budget. They are buying the right to keep spending while their peers get frozen.
The 72 percent unmanaged-risk number, and being the exception
For the operator who wants one line to bring to the board, here is the cleanest in-window evidence, attributed honestly as a vendor survey.
"72% say their agents introduce unmanaged financial or compliance risk."
That number is not there to scare anyone. It makes a calm point: the market average is unmanaged risk, and the work in front of us is to be the exception. That is a confident posture, not a fearful one. It says we know where the gap is and we have a plan for our own lines.
The three-question test for every AI line item
The three-question test is the whole thing. For every AI line on the sheet, ask: who owns it, what outcome does it move, and can we shut it off cleanly. A line with a named owner, a defensible outcome, and a clean kill path survives any review I have ever seen. A line missing any one of the three is not necessarily a bad investment. It is an unfinished one, and there is time until September to finish it.
The companies handling this well are not the ones with the most sophisticated AI. They are the ones who treated the budget review as a forcing function instead of a threat. They went line by line before the board did. They turned the long pauses into either a measurement wire or a clean cut. And they walked into the room with a sheet where every line had an answer.
That operator with the highlighter has a few weeks, and so does anyone walking into a September review. The lines that can be explained are safe. The lines that cannot are not doomed, they are just next on the list to finish. Start with the ones touching money, because those are the ones the board will start with too.