AI agents are producing measurable returns, but mainly in well-defined workflows with clear rules and human oversight.
In an Anthropic survey of more than 500 technical leaders, 80% of organizations with deployed agents reported measurable returns. Gartner also expects more than 40% of agentic projects to be canceled by the end of 2027.
Both findings can be true. Early returns do not guarantee long-term operating value. The difference comes down to workflow design, governance, and how success is measured.
Value concentrates in work teams can verify
The strongest results appear in repetitive workflows that leave a clear record of each action. A person can quickly review the output, approve it, or handle an exception.
Coding agents have delivered some of the clearest gains. Customer service agents are resolving routine requests at scale. Security teams are using agents for alert and phishing triage, while finance and operations teams are applying them to invoice extraction, document validation, and routing.
These workflows share a practical advantage. The rules are defined, the data is accessible, and mistakes can be caught before they cause lasting harm.
Open-web autonomy and unsupervised agent chains remain far less dependable. They encounter changing websites, ambiguous instructions, credential risks, and errors that compound over time. These uses may justify controlled experiments, but they should stay away from critical operating paths.
Task length creates a reliability cliff
Short demonstrations can give a misleading impression of what agents can handle.
OSWorld 2.0 tested agents on 108 longer computer-use workflows. A skilled person needed about 1.6 hours to complete the average task. The best system completed only 20.6% of them end to end.
Performance fell as tasks grew longer. Agents completed roughly 20% to 24% of work that would take a person less than 45 minutes. Beyond 163 minutes, completion dropped to effectively zero.
The practical response is to redesign the work. If the human version takes more than an hour, divide it into shorter stages. Each stage should produce evidence someone can inspect, with approval required before the agent continues.
Today’s advantage comes from well-placed checkpoints, not from waiting for flawless autonomy.
Production value depends on more than the agent
Most buyers focus on applications, models, and orchestration platforms. The layers beneath them receive less attention, even though they determine whether an agent can operate safely in production.
An agent control plane manages identity, tool access, approvals, evaluation, and audit records. Shared protocol infrastructure connects the stack and makes it easier to change models or vendors later.
Buying an agent does not provide these controls automatically. Organizations that skip them often discover the gap after deployment, when fixing it becomes more expensive.
Model choice should increasingly become a routing decision. The workflow should remain separate from the models used to plan, execute, and review it. That separation protects the investment as model performance and pricing change.
Measure accepted results, not activity
Agent economics should be measured by the cost of an accepted result.
Compare that figure with the human cost of completing the same task, including the time spent reviewing and correcting agent output. Define what counts as an accepted result in the vendor contract.
This detail matters. Vendors may report resolution rates near 71%, while independent estimates place comparable performance between 42% and 50%. Much of the difference comes from what each provider counts as a resolution.
A change in measurement can affect the business case more than a model upgrade.
Start where proof can earn trust
The strongest first projects have clear policies, high volumes, and an existing human fallback. Leaders should be able to change the workflow directly and measure the result without relying on open-ended judgment.
Before writing prompts, define a short operating contract for the agent. State what it may do, what evidence it must produce, when it must escalate, and who owns the result.
If a workflow cannot produce evidence a person can inspect, it is not ready for production.
AI agents can create real value today. The opportunity lies in choosing work the organization can verify, governing it from the start, and expanding autonomy only after the economics and controls have been proved.
Download the full white paper, What AI Agents Actually Deliver
For more information on CloudCollect or our government consulting services, please contact West Coast Consulting Group at info@westcoastconsulting.com or call 415-626-3493.
