The narrative around AI agents in 2026 has been relentlessly optimistic. Deployment numbers are up, investment is up, and you can barely read a tech brief without hearing about “autonomous AI workforces” changing everything.

And then KPMG published its Q2 2026 Global AI Pulse survey, and the data told a more complicated story.

49% of senior leaders at organizations with more than $50 million in annual revenue said they had scaled back AI agent deployments because operating costs outweighed the benefits. Not failed entirely — scaled back. These are businesses that tried, saw the bills arrive, and pulled back.

The survey covered 2,145 leaders across 20 countries. This isn’t fringe pessimism. It’s a mainstream problem that isn’t being talked about clearly enough.

What’s Actually Happening

The cost problem with AI agents isn’t random. It has a specific shape.

When businesses run traditional software, they pay for seats or capacity. The bill is predictable. You know what you’re buying before you buy it. AI agents work differently. They run on usage-based, token pricing — and they use a lot of tokens. A single complex multi-step agent task can cost more than you’d expect, and when you’re running thousands of them a month, the meter adds up fast.

A third of leaders in the KPMG survey said they don’t fully understand AI cost structures, including how token pricing works. That’s the real diagnostic. These aren’t businesses with bad AI strategy — many of them deployed agents to real workflows and saw genuine productivity gains. But they didn’t have a clear picture of what each agent run would cost before they scaled. When the bills arrived, they didn’t know what they were looking at.

The other issue is scope. AI agents are genuinely good at multi-step tasks, which means they naturally expand. An agent deployed to handle one workflow starts getting used for adjacent ones. The use case grows without a corresponding plan for managing the cost of that growth.

Why 79% Are Still Investing

Here’s the part that’s easy to miss in the cost-pullback headline: 79% of leaders in the same survey say AI is still a top investment priority, and spending is holding steady.

These aren’t companies giving up on AI agents. They’re recalibrating how they deploy them. The pullback is a correction, not a retreat — and that distinction matters enormously for how you think about the current moment.

Businesses that went too broad too fast are pulling back to go deep in one place. Businesses that never modeled the economics are now doing the modeling they should have done at the start. And businesses that are watching from the sidelines are either about to enter with more realistic plans, or about to fall further behind the ones that are recalibrating well.

The executives who understand what’s happening here have an advantage. The ones who read “49% pulled back” as confirmation that AI agents aren’t ready yet are reading the wrong signal.

The Deployment Pattern That Works

Looking across organizations that are actually generating returns on AI agents — not just impressive demos — a few patterns show up consistently.

One workflow, one agent, real numbers. The deployments that justify their cost start with a single workflow where the economics are clear. Not “we’ll automate our customer support broadly” but “our agents will handle the 400 identical tier-1 support queries we get each week.” You can model the token cost of those 400 queries before you build. You know what you’re buying.

Measure the offset, not just the output. An AI agent that handles 400 queries a week doesn’t create value in isolation. It creates value by letting people redirect their time to work that actually requires them. The businesses seeing real ROI are tracking both sides: what the agent costs to run and what the team is now doing instead. When you can show that an agent’s monthly cost equals 20% of one staff member’s salary, and the staff member is now doing work that genuinely requires them, the math closes.

Token discipline from day one. The cost overruns that KPMG documented are largely preventable with upfront engineering. Agents that are well-scoped — given clear task boundaries, constrained tool access, and system prompts that don’t invite them to over-explain — use dramatically fewer tokens than agents built without those constraints. The difference between a disciplined agent and an unconstrained one on the same task can be 3-5x in token usage. That’s not a small gap.

Cache what you can. Many agent architectures send the same context (system prompts, knowledge base content, policy documents) on every single call. When that context gets cached at the model layer, the effective per-call cost drops significantly. It’s a technical implementation choice with a direct line to your cost structure.

What Consulting Firms Are Getting Wrong

Consulting organizations have a particular version of this problem because they often deploy agents across client deliverable workflows — proposal generation, research, report writing, data analysis — where the tasks are varied and the token usage is hard to predict in advance.

The common mistake is deploying agents for document-heavy, open-ended tasks without first establishing a cost ceiling. A research agent that synthesizes sources to produce a client briefing does genuinely useful work. But if it’s running against 50 sources per brief and generating 3,000 words of output, and you’re producing 200 briefs a month, the cost math needs to be part of the architecture decision before the build starts.

The other pattern worth examining: agents that are deployed to replace coordination and communication tasks — status updates, meeting recaps, follow-up sequences — often have far better unit economics than agents deployed for open-ended reasoning tasks. They touch the same workflows repeatedly, the inputs are constrained, and the outputs are predictable. These are the places where the 3x improvement in efficiency that AI agents promise actually shows up in the bank account.

What the KPMG Finding Actually Means for Your Strategy

If you’re looking at 49% pulling back and wondering whether now is the right time to start, here’s the honest answer: the cost problem is solvable with the right approach, and the organizations that solve it now are building advantages that will be hard to close later.

The ones who scaled back aren’t going away. They’re rebuilding with better economics. The ones still on the sidelines are watching the rebuilding and concluding the technology doesn’t work. Neither conclusion is accurate.

What’s accurate is this: AI agents create real, measurable value in specific workflows. The cost of running them is real and needs to be modeled. The businesses that combine genuine use case selection with solid cost architecture are the ones generating returns that look nothing like what the pullback statistics suggest.

The gap between organizations that know how to deploy AI agents profitably and those that don’t is widening every quarter. That gap is the opportunity.

The One-Workflow Starting Point

If your organization hasn’t deployed AI agents yet, or if you’ve pulled back and are reassessing, the path forward is simpler than it looks.

Find one workflow that is:

• Repetitive and high-volume (the agent runs many times)
• Has a predictable output format (not open-ended generation)
• Has a clear human equivalent cost you can compare against

Model the token cost before you build. Confirm the economics work on paper. Build narrowly, measure everything, and expand only after you’ve validated the unit economics in production.

This isn’t how the AI agent narrative usually gets told, because “start narrow and model costs first” doesn’t generate good conference keynotes. But it’s what separates the 51% of organizations that are generating returns from the 49% that pulled back wondering why the bills were so high.

The data is clear enough. The question is what you do with it.

---

Enterprise DNA’s Omni Ops service helps organizations identify the right AI agent use cases and build them with cost-conscious architecture from the start. If you’re trying to figure out where to begin — or why a previous deployment didn’t work — book a discovery call with our team.