Almost two in three technology leaders plan to deploy AI agents in the next two years. Gartner expects more than 40% of those projects to be canceled. The difference is decided before the first line of configuration.

In Gartner’s survey of more than 2,500 CIOs and technology executives behind its agenda for this year, 64% say they will deploy agentic AI within the next 24 months. The conversation in steering committees is no longer whether to use agents, but where and how fast.

The same firm issued a warning in the middle of last year that is worth keeping close: more than 40% of agentic AI projects will be canceled before the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. None of the three causes is technological.

An agent is not just another feature

An assistant answers questions. An agent makes decisions and takes actions inside a process: it reconciles an item, releases an order, proposes a journal entry, replies to a supplier. That difference changes the nature of the risk.

When an assistant gets it wrong, someone reads a wrong answer. When an agent gets it wrong, the error is already posted in the ERP. In its latest survey on the state of AI, McKinsey found that 51% of organizations using it had already experienced at least one negative consequence, with inaccuracy the most frequent. With agents, inaccuracy stops being a reading problem and becomes an accounting problem.

Three questions before going live

In the projects we support, an agent’s return is defined by three answers that are almost never written down before starting:

  1. Which financial outcome does it have to move? Not «improve efficiency», but a concrete metric: days sales outstanding, cost per invoice, days to close, reconciliation hours. Without a baseline, there will be no way to prove the return.
  2. Is the process defined well enough to delegate? An agent executes rules. If the actual process piles up exceptions, duplicate validations and criteria that live in one person’s head, the agent automates that confusion faster.
  3. Who answers when the agent gets it wrong? Every agent needs a business owner, clear limits on its autonomy, such as amounts or transaction types, and a log of what it did and why.

Gartner puts numbers on the first question. Only 33% of CIOs consistently pursue measurable financial outcomes from their technology initiatives, especially AI, and those who do are 25% more likely to excel. The rest measure activity: licenses activated, users, queries. At year end, the executive committee asks about something else.

The trap of the agent that comes in the box

SAP, Oracle and the other ERP vendors are building agents into their platforms. That is good news, because it lowers the technical barrier. It is also a risk, because it plants the idea that switching on the agent is the project.

Gartner also warns about «agent washing»: existing products such as assistants, robotic process automation or chatbots, now presented as agents without being so. For an investment committee, the question is not what the label says, but which decisions the tool makes, within which limits and on which data.

When an agent is not the answer

Not every process needs an agent. If the rules are stable and exceptions are few, traditional automation inside the ERP does the same job at lower cost and risk. An agent adds value where there is variability that calls for judgment: unstructured documents, reconciliations across multiple sources, exception triage. Using one where a rule is enough means paying for autonomy you do not need.

How we approach it at Aunergia

Before switching on an agent inside SAP, Oracle Fusion or whichever ERP the company runs, we map the actual process, not the one in the manual. We define the financial metric it has to move and its baseline. We set the business owner, the autonomy limits and the decision log. Only then do we configure and measure.

The sequence matters: process, outcome, governance, technology. The other way around, agentic AI grows the budget, not the bottom line.

Does your organization know which financial outcome each agent it plans to switch on this year has to capture?

Sources: Gartner, 2026 CIO and Technology Executive Survey (October 2025); Gartner, agentic AI projects prediction (June 2025); McKinsey, The State of AI in 2025.