TL;DR: according to a Ness Digital Engineering report published on August 19, 2026, 99% of companies plan to deploy AI agents in production, but only 9 to 14% have actually done so.
The gap fits in two numbers, and it is dizzying. While almost every leadership team puts AI agents on its roadmap, fewer than one company in seven runs them day to day. That gap now has a name, a documented explanation, and above all some concrete ways around it.
Key takeaways
- 99% of companies plan to put AI agents into production, but only 9 to 14% have managed it, according to the Ness Digital Engineering report published on August 19, 2026.
- The report identifies two main failure causes: loss of trust in probabilistic answers and the lack of visible change in the teams' daily work.
- The agentic AI market in financial services alone is projected at $33.26bn by 2030.
- That same fortnight, NVIDIA and six financial institutions announced platforms designed to mobilize over $500bn for AI infrastructure (August 10, 2026).
A documented gap between intent and production
Engineering firm Ness Digital Engineering published a study on agentic AI adoption on August 19, 2026, picked up by the business press. The finding: 99% of the companies surveyed plan to deploy AI agents in production, but only 9 to 14% have crossed that line. The authors of the detailed analysis, Shankara Ramanathan and Mihai Hosu, call it a "death valley" between proof of concept and production: the zone where projects stall after a successful demo.
Two causes dominate, per the report. First, loss of trust in probabilistic systems: an agent built on a large language model never answers exactly the same way twice, and business teams end up doubting its output. Second, the absence of visible difference in daily operations: many pilots automate an isolated task without anyone seeing their work actually change. The project then loses its internal sponsors, then its budget.
The financial stakes remain very real: the same study projects the agentic AI market in financial services at $33.26bn by 2030.
Why do projects die between pilot and production?
Because most companies treat an AI agent as a technology upgrade, when it is really a workflow transformation. That is Ness's central thesis, and it matches what practitioners have observed for two years. A successful pilot proves a model can execute a task. Production demands everything else: error handling, human oversight, integration with existing systems, access governance and cost tracking.
The report offers concrete recommendations. Assess before building: the business domain, the infrastructure, the availability of data and APIs. Put the agent inside an interface teams already use, the internal messaging tool for instance, rather than adding yet another window. Measure success on the overall experience of a process, not on the automation of one isolated task. And track real costs, tokens, subscriptions and cloud, at the application level, so they can be compared against the expected return.
One part of the diagnosis deserves emphasis: the balance between probabilistic and deterministic logic. A company's business rules are deterministic; a language model is not. Agents that survive in production keep the two separate: the model understands and proposes, coded rules validate and execute.
Meanwhile, infrastructure is raising hundreds of billions
The contrast with the same fortnight's news is striking. On August 10, 2026, NVIDIA signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create financing platforms designed to mobilize over $500bn of third-party capital for AI infrastructure. On August 11, IBM and Together AI announced a $240m agreement for an NVIDIA inference cluster on IBM Cloud, expected in the first quarter of 2027. On August 12, Ryanair signed a five-year partnership with Google Cloud that will roll out Gemini Enterprise and Google Workspace to its 35,000 employees.
In other words: compute supply, models and platforms are moving far faster than organizations' ability to do something with them. That lag is not new in the history of technology, but its scale is. When 99% of companies declare an intention and fewer than 15% follow through, the bottleneck is neither the model nor the server: it is implementation.
What does this change for a small business?
First, a counter-intuitive piece of good news: on this terrain, a small business starts with an advantage. The two failure causes Ness identified, trust and visible change, are far easier to handle in a company of 10 to 100 people than in a large group. There are fewer processes, decision chains are shorter, and the result of an automation shows within weeks, not quarters.
Then a warning: the death valley does not spare small structures. An agent assembled in three days on an impressive demo, never wired into the company's actual tools, will die the same way a corporate pilot does, only faster. The table below sums up what separates, at small-business scale, a doomed pilot from an agent that lasts.
| Criterion | Pilot that dies | Agent that lasts in production |
|---|---|---|
| Scope | "AI everywhere" | One process, one measurable result |
| Measurement | An impressive demo | Time saved, quantified every week |
| Errors | Ignored until the incident | Error path planned, human in the loop |
| Costs | Discovered at month end | Tokens and subscriptions tracked from day one |
| Interface | One more window to open | Embedded in the tool already in use (messaging, CRM) |
Finally, judgment: not every process calls for an agent. When the task is repetitive and unambiguous, deterministic automation, an n8n workflow for instance, costs less, is easier to fix and never loses anyone's trust, precisely because it always answers the same way. An agent is justified when you need to understand language, triage cases or decide between several actions. Start with simple automation and reserve the agent for the steps that genuinely need it: that is often the shortest road to those famous 9 to 14%.
What I see in the field
The systems I keep running in production confirm this diagnosis point by point. The IA Brew newsletter, 93 n8n nodes, publishes every week without intervention: what keeps it alive is not the model, it is the error paths and the monitoring. The automated press monitoring for Fromagerie Ermitage lasts because it lives inside a tool the team already checks, not in one more window. And on AI agent projects, the first question is never "which model?" but "which process, measured how, at what cost per run?". It is less spectacular than a demo; it is what separates the 99% of intentions from the 9 to 14% of realities.
The number to remember is not 99, it is 9. The difference between the two cannot be bought from NVIDIA or Google: it is built process by process, with a narrow scope, honest measurement and tracked costs. One question to close: in your company, which process would show visible results from its very first automated week?
Frequently asked questions
Why do so few AI agents reach production?
According to the Ness Digital Engineering report of August 19, 2026, two causes dominate: loss of trust in probabilistic answers and the lack of visible change in the teams' daily work. Projects treated as a simple technology upgrade, without rethinking the workflow, stay stuck at the pilot stage.
Does a small business have better odds than a large company at shipping an AI agent?
On one specific point, yes. A small business has fewer processes, shorter decision chains and results that show within weeks, which neutralizes the two failure causes Ness identified. Its constraint is budget, which is why a narrow, measurable scope matters from day one.
Do you always need an AI agent to automate a process?
No. A repetitive, unambiguous task is better handled by deterministic automation, an n8n workflow for instance, which is cheaper and more predictable. An AI agent is justified when the process requires understanding language, triaging cases or choosing between several actions.