To get value from AI, an organization needs to decide which process should improve, how it will measure that improvement, and when it will stop investing. Model selection matters, but it cannot substitute for those decisions. Without them, even an impressive system can become a recurring expense with benefits that are difficult to prove.
According to Gartner survey findings cited in an Optimum7 publication, only 13% of executives report that their organization has generated significant or substantial value from AI. In our view, this points primarily to a management gap: organizations have learned to launch pilots faster than they have learned to decide which ones deserve to continue.
- 13% report significant value from AI.
- 31% have defined a rule for stopping a pilot.
- 17% have defined a rule for retiring low-value systems.
- 4% have planned to redesign roles.
The figures come from a survey of 2,002 CIOs and senior technology executives, conducted as part of Gartner's 2027 agenda. They are self-reported and cited in a secondary commercial publication, so they should be read as a signal, not a precise diagnosis of the entire market. Nor should we assume that organizations reporting no significant value have generated no value at all.
The problem is not just failed pilots. It is pilots that never end.
A failed experiment can be a good business outcome: the organization learns something, limits spending, and frees up resources. The trouble starts when an experiment with no clear result keeps running because it has already been presented to the executive team, incorporated into a work plan, or associated with a senior sponsor.
The publication describes a correlation: executives who had defined a pilot stop rule were 2.3 times as likely to belong to Gartner's pioneer group. That does not prove that the stop rule caused the success. Organizations with stronger management discipline may both set better rules and achieve better results.
Even so, the practical decision is clear: an AI project needs more than an initial budget. It needs conditions for continued investment and conditions for exit, agreed in advance.
A stop rule is not a vote against innovation. It protects the budget that makes continued innovation possible.
Stopping a pilot, restricting a system to draft-only output, and retiring a production system are different decisions. A decline in quality may justify reducing the system's authority. Unauthorized disclosure of information may require an immediate shutdown, even if the system is profitable.
Measure the process, not how impressive the answer looks
An agent that drafts a quote quickly does not necessarily shorten the quoting process. If an employee still has to recheck prices, find missing terms, and correct commitments, the work may simply have changed shape.
Measurement therefore needs to begin before deployment and cover the process from start to finish. The CFO needs to see total cost. The operations leader needs to see throughput and quality. The process owner needs to know who handles exceptions.
- Sales quotes: Drafting speed is a partial measure; time to an approved quote is a more useful business measure.
- Customer service: Response volume is a partial measure; resolution without repeat contact is more useful.
- Document processing: Extraction accuracy is a partial measure; cost per correctly processed item is more useful.
- Sales activity: Message volume is a partial measure; profit after operating and oversight costs is more useful.
These comparisons show why a strong technical metric is not enough. Costs must include licenses, model usage, integration, maintenance, error handling, and human oversight. Time saved is potential value. It becomes realized value when it produces more output, better service, or an actual reduction in spending.
A comparison between the quarter before deployment and the quarter after it is not sufficient on its own, either. Holidays, external disruptions, changes in customer mix, and seasonal workloads can all affect the result. Where possible, compare performance with a similar group that has not undergone the change, and document external factors.
Human oversight does not require approval of every action
AI can perform tasks that require interpretation and judgment, rather than just follow rigid rules. That is also why it should not be treated as a deterministic mechanism that always produces the same result.
But if every output requires a full manual review, scaling becomes difficult. The goal is not to remove people from the process. It is to move them from routine execution to focused oversight of exceptions.
Research from ChurnZero, mentioned in the source material, suggests a relationship between inconsistent measurement and greater demand for human approvals. Because the sample size and methodology were not provided there, it is not a sound basis for a financial forecast. The operational logic, however, is worth testing in any organization: when you do not know what is reliable, you check everything.
Oversight that supports scale: People should make the decisions that warrant human judgment. Routine actions can proceed within defined authority limits, with monitoring and sampling matched to the risk.
For example, an agent could prepare a draft quote using an approved price list, while routing exceptions to discount limits or contractual commitments to a person. Any expansion of its authority should depend on performance tested in the actual workflow, not just the score the model assigns itself. Action logs, limited permissions, and a way to revert to a manual process are also necessary.
A practical framework before increasing investment
You do not need to start with a massive enterprise-wide program. You can start with a clearly defined process, provided that accountability and measurement are equally clear.
- Choose a process and an accountable owner: Define the business need, the affected users, and a manager with the authority to change the process.
- Establish the baseline: Measure time, cost, quality, and exceptions before introducing the system.
- Set continuation and stop criteria: Agree in advance on performance thresholds, risk limits, and who can suspend operations.
- Design risk-based controls: Distinguish between permitted actions, actions requiring review, and prohibited actions.
- Decide based on results: Scale, revise, or close the project using agreed evidence, not enthusiasm for the demo.
The baseline and the validation of benefits should be established with the process owner and finance or an oversight function, rather than left to the sponsor alone. This does not assume that anyone is acting dishonestly. It recognizes that the person promoting an investment is not the most neutral person to evaluate it.
Two adoption tracks, one organizational responsibility
Organizations need to advance AI literacy and agent development in parallel. Employees need to know how to communicate with models, provide context, and check outputs. At the same time, the organization needs an internal capability to build and manage agents within its workflows.
A simple personal tool may demand a substantial change in habits, while an agent integrated into an existing system may require almost no change to an employee's working environment. Ease of adoption is therefore not determined by technical complexity alone.
As agent activity expands, IT takes on responsibilities resembling workforce management: assigning authority, tracking performance, training, suspending, and retiring agents. That requires shared infrastructure for permissions, monitoring, versions, and costs, not a collection of disconnected automations.
The most important capability is not delivering another demo. It is identifying where AI fits a process and proving that it improves that process. This takes deep AI knowledge, relevant education, and business and operational experience. Combining technical expertise with domain knowledge matters more than familiarity with the latest tool.
At the next budget discussion, the question is not simply which systems to buy. Ask which process has improved, what it now costs to operate, what still requires a person, and which investment should already be stopped. That is where AI adoption becomes something you can manage, not just present.