Four in ten agencies now have at least one AI agent in production. That sounds like the dam has broken. The less comfortable detail is that a quarter of those teams are losing money on the work.
Digital Applied’s April 2026 survey of 250 marketing and development agencies puts the split in plain numbers. Production use jumped from 9% in Q1 2025 to 41% today. Median reported ROI is 3.2x. The top decile claims 11x. The bottom quartile sits at 0.7x.
Same technology. Very different outcome.
The difference is not usually the model. It is the operating discipline around the model. Teams that treat agents as a new class of worker ask boring, useful questions before they scale anything: what does the manual process cost, what counts as a pass, how often does the agent fail, what does each successful task cost, and who can stop it when it goes sideways?
That last question matters more than most demos admit.
WRITER’s April 2026 enterprise survey found that 97% of executives say their company deployed AI agents in the past year, but only 23% report significant ROI. It also found 67% believe their company has had a data leak or breach due to unapproved AI tools. Worse, 35% could not immediately pull the plug on a rogue agent.
That is not an AI strategy. That is a permissions problem wearing a blazer.
The strongest returns are coming from narrow workflows with clean inputs and obvious pass or fail tests. Digital Applied found SEO audit agents delivering 11.4x ROI. Client-report drafting sat at 1.6x. That makes sense. An SEO audit has defined checks, repeatable output, and easy comparison against a baseline. A client report is more political. It needs judgement, account context, and tone. Automating the first saves hours. Automating the second often creates review debt.
This is where mid-market firms should be careful. Copying an enterprise agent program is usually a waste. The better pattern is smaller and harder to fake.
Pick one repetitive workflow. Measure the current manual cost. Build a short eval set from real examples, not synthetic happy paths. Define what the agent can access, what it can change, and what must go through a person. Then run the numbers as cost per successful task, not tokens spent or seats purchased.
The token bill is still useful. Digital Applied found a median monthly spend of $1,800 per active agent, with the top quartile at $4,200 and the bottom quartile at $420. Agencies that shared billing data had overstated ROI by about 18% and understated token spend by about 24%. That gap is where bad business cases live.
A production agent needs three things before it earns more scope: an eval harness, a human review model that matches the risk, and a rollback plan. Without those, more agents just means more places for quiet failure to hide.
For StormView clients, the practical move is not to deploy agents everywhere. It is to prove one agent is cheaper, safer, and better than the old process, then use that proof to choose the next workflow.
Leave a Reply