TL;DR
- Enterprise agentic AI pilots have been running on discounted frontier model pricing — those discounts are expiring, and the real bill is coming due.
- Up to 60% of agentic AI operating costs come from verification and refinement, not generation — per-token pricing is a misleading predictor of true cost.
- Harness design alone can create a 41% variance in cost-per-task on identical models, meaning architecture decisions made today will determine economics tomorrow.
- Enterprises that optimize agent economics at design time — not after production go-live — will be the ones that survive the subsidy cliff.
There's a particular kind of corporate optimism that only exists during a pilot. The dashboards look clean, the demos impress the C-suite, and the invoices feel almost too manageable. Then the bill comes.
For enterprise agentic AI, that moment is arriving in H2 2026 — and it has a name: the subsidy cliff.
The Free Lunch Is Ending
For the past two years, the major frontier model providers — OpenAI, Anthropic, Google, and others — have been aggressively subsidizing enterprise access to their most powerful models. The goal was adoption. The strategy worked. Enterprises across every vertical launched agentic AI pilots, built internal demos, and in many cases, began committing engineering teams to full production architectures built on top of these subsidized cost structures.
Now the discounts are fading.
Progress Software's Chief AI Officer Ed Keisling flagged the issue directly this week, warning that enterprises are "reassessing AI deployments as discounted frontier models fade," exposing the true runtime costs of running autonomous agents at scale. The shift is already accelerating adoption of smaller, cheaper models — but swapping out the model underneath a production agent architecture is not a trivial exercise. It's an excavation project.
The cruel irony? Most enterprises won't feel the pain until they're already committed. The cost problem doesn't surface during pilots. It surfaces during production, six to twelve months after the architecture decisions were locked in.
What "Agentic AI" Actually Costs to Run
Here's where the industry has a serious education gap — and it's not about the technology. It's about the economics.
Most conversations about AI cost still center on per-token pricing. That number is easy to find, easy to model, and almost completely useless for predicting the true cost of running a production agentic system.
According to McKinsey research, approximately 60% of agentic AI operating costs are spent on response verification and refinement — not on generation. Read that again. The majority of your inference bill isn't the model "thinking." It's the scaffolding around the model checking its own work, verifying outputs against constraints, and looping back when something doesn't look right. Context management is the single largest cost driver in a real-world agent deployment. Per-token pricing, in isolation, tells you almost nothing useful.
This is the part that catches engineering teams off guard. They benchmark model performance on task quality and latency. They rarely benchmark the overhead introduced by their own orchestration layer.
The Harness Effect: Your Architecture Is Costing You More Than You Think
This brings us to one of the most underappreciated dynamics in enterprise AI deployment: the harness effect.
Research from Systima consultancy found that harness design — encompassing system prompts, tool definitions, and orchestration mechanics — generates a 38% variance in token overhead on identical tasks running on identical models. More strikingly, cost-per-task can vary by 41%, and execution time by 44%, based on harness design alone. Same model. Same task. Wildly different economics.
Think about that. Two engineering teams could deploy the same frontier model to solve the same business problem, and one team could be paying nearly half again as much as the other — not because of the model, and not because of the task complexity, but because of how their agent is wired together.
The problem is compounded by a visibility gap that plagues most enterprises. Configuration is scattered across teams: the group managing harness selection is rarely the same group doing prompt engineering, which is rarely the same group managing MCP servers. CIOs often have no real-time visibility into whether agents are rewriting cache mid-session, what fan-out costs look like at scale, or how orchestration overhead is accumulating across thousands of concurrent agent runs. It's like running a data center where nobody can see the power bill until the end of the month — except the month lasts two fiscal quarters.
71% of "Agents" Aren't Actually Agents
There's another uncomfortable truth lurking beneath the hype.
The Valorem Reply Agentic AI Framework (published July 2026) offers a useful definition: a true agentic AI system operates through a five-step loop — perception → reasoning → planning → action → learning. It's not a chatbot with a fancier system prompt. It's a system that can pursue a multi-step goal autonomously, adapt when it hits obstacles, and improve its approach over time.
By that definition, 71% of what enterprises are currently calling "agents" are single-prompt wrappers — essentially dressed-up completions endpoints with no genuine planning or feedback loop. They're not wrong to start there. But they are wrong to budget for full agentic deployment costs based on those pilots.
Gartner predicts that 33% of enterprise software will include agentic AI by 2028, up from less than 1% in 2024. That's a staggering growth curve. And it means the vast majority of enterprises are about to make their most consequential agentic architecture decisions — right now, in 2026 — while subsidized pricing is still masking what production will actually cost.
The social conversation reflects exactly this confusion:
"I'm amazed most people still think AI agents are just chatbots. Here's the complete Agentic AI tech stack behind autonomous AI systems."
— @shushant_l"Agentic AI is not a chatbot that responds — it is a system that plans, acts, checks itself and repeats until the task is done. Most developers build agents that fail on the first error."
— @0xNoryxx"Artificial intelligence is entering a new phase. As AI evolves from simple chatbots to autonomous agents capable of reasoning, planning, and executing complex tasks, success will depend on far more than powerful models alone."
— @urdav3
The gap between "chatbot with memory" and "true autonomous agent" isn't just technical. It's a cost multiplier, and most organizations haven't modeled it.
The Race to Capture Enterprise Agent Deployments
The consulting world has noticed. NTT DATA is targeting $2 billion in revenue from agentic AI services by 2027. Cognizant, LTM, and other major IT consultants are scaling forward-deployed engineer programs specifically for enterprise agent rollouts. The market is moving fast — arguably faster than the economics have been properly stress-tested.
As Analytics India Magazine reported this week, the end of AI subsidies is forcing a fundamental rethink of how enterprises should approach agent deployment economics. The article notes that Progress Software's Keisling sees the expiry of discounted pricing as an accelerant pushing companies toward smaller, purpose-built models — a strategic shift that only makes sense if it's planned into the architecture from the start.
And that's the crux of the challenge. You can't bolt on cost-efficiency after the fact. Not when 38% of your token overhead is baked into your harness design. Not when your orchestration mechanics were built for a pricing model that no longer exists.
Optimizing for the Cliff Before You Fall Off It
So what does smart enterprise agentic AI economics actually look like in practice? A few principles are emerging from the teams getting this right:
1. Design for cost, not just capability. The temptation during the design phase is to optimize for what the agent can do. Smart teams also model what it will cost to run at 10x and 100x scale. Context window management, tool call frequency, and verification loop design all need economic constraints baked in from day one.
2. Treat harness design as a first-class engineering concern. System prompts and orchestration logic aren't afterthoughts. They're cost levers. A well-designed harness isn't just more reliable — it's meaningfully cheaper to operate. The 41% cost variance identified in Systima's research isn't noise. It's signal.
3. Build real-time cost visibility into the architecture. If your CIO can't see harness-level cost metrics in a live dashboard, you're flying blind into the subsidy cliff. Instrumentation needs to be part of the initial build, not a bolt-on when the invoices get scary.
4. Model the shift away from frontier models now. The move toward smaller, cheaper, purpose-built models is coming whether enterprises plan for it or not. The ones that plan for it will architect their agent layers with model-agnostic interfaces. The ones that don't will face expensive re-engineering cycles in 2027.
5. Benchmark verification overhead, not just generation. If you're only measuring token costs at the generation layer, you're missing the majority of your actual cost structure. Build test harnesses that measure the full inference loop — including all verification and refinement steps — before you commit to a production architecture.
The Inflection Point Is Now
The enterprises moving from pilots to production in H2 2026 are doing so at a peculiar moment: the last window in which subsidized pricing can obscure architectural cost problems before they become budget emergencies.
This isn't a prediction about AI's potential — it's a reckoning about how that potential gets deployed. Gartner's 33% enterprise software penetration by 2028 is almost certainly going to happen. The question is how many of those deployments will be economically sustainable, and how many will become costly lessons in the difference between a pilot that impressed a boardroom and a production system that runs at scale.
The subsidy cliff isn't a cliff you can see coming from far away. By the time it's visible, you're already committed to the architecture. The engineering teams working on agent design today are making decisions that will determine cost structures in 2027. The ones who treat economics as a design-phase concern — not a deployment-phase surprise — are the ones who will make it to the other side intact.
And for the 71% still building single-prompt wrappers and calling them agents? The cliff will feel less like a wall and more like a very expensive wake-up call. Let's hope the snooze button isn't priced per token.
Published in Stream · Dispatch #458 · July 20, 2026 · 9 min read.
Reply to paolo@mont3.ch - every email gets a human answer within 24h.