By Mitch Edwards · · 6 min read

97% of companies deployed AI agents this year. Half wish they hadn't.

The 2026 reports from KPMG, Deloitte, Writer and Forrester all tell the same story: high AI agent deployment, modest ROI, and a widening gap between spend and value. A practical guide to what to fix before your next AI rollout.

Nearly every company surveyed by Writer in Q1 2026 had deployed AI agents in the past year. Almost half called the result a disappointment. The bigger enterprise reports back the same picture.

That gap is the most interesting number in enterprise software right now.

The tech is rarely where the failure lives. Most companies skipped the operations work.

The pattern I hear most often goes like this. A founder is being pitched an AI tool. The vendor’s plausible, the demo’s clean, the price feels manageable. When I ask what success looks like at day 30, the answer isn’t there yet. The tool isn’t where the failure is going to live. The operations work hasn’t been done before the procurement decision is on the table.

What the data actually says

I’ve been reading through the major 2026 enterprise AI reports. KPMG’s Q1 AI Pulse Survey, Deloitte’s State of AI in the Enterprise, Writer’s 2026 AI Adoption Survey, Forrester’s work on agent ROI. They tell the same story from different angles.

KPMG, Q1 2026 AI Pulse Survey. 54% of organisations have AI agents deployed in at least one function; 80% in AI-specific domains. Deployment is uneven across functions, and the gap between investment and measurable value is the consistent theme.

Deloitte, State of AI in the Enterprise 2026. Investment continues to climb but enterprise-wide value capture lags. Governance and operating-model gaps are repeatedly named as the binding constraint.

Writer, 2026 AI Adoption Survey (2,400 knowledge workers across the US, UK and Europe, fielded Dec 2025 to Jan 2026). The most striking headline:

  • 97% of companies say they deployed AI agents in the past year
  • 52% of employees are already using them
  • 48% call adoption a massive disappointment, up from 34% last year
  • Only 29% see significant ROI from generative AI, and just 23% from AI agents specifically

Forrester. The most-cited Forrester figure in 2026 reporting is that around 88% of AI agent pilots never reach production. The root-cause analysis points consistently at unclear success criteria, insufficient tool or data access, and drift in evaluation coverage. Different language, same underlying gap: companies invest in agents before the workflows, owners, and metrics are in place.

The numbers shift depending on whose survey you read. The direction is the same across all of them: high deployment, low scaled ROI. The companies winning aren’t the ones that bought the best tools. They’re the ones that did the boring work first.

Quick definitions, because the labels are slippery

Most “AI agent” headlines aren’t really about agents. Three things get lumped together:

  • AI assistant. A chat tool you ask things and that drafts things back. Copilot, Claude, ChatGPT used inside a workflow. Human stays in the loop on every step.
  • AI workflow automation. A scripted sequence triggered by an event, with AI doing one or two specific steps inside it (classify, summarise, draft). Predictable inputs and outputs.
  • AI agent. A system given a goal that decides which steps to take, in what order, calling tools or other AIs as it goes. Less predictable, harder to test, easier to get wrong at scale.

Most of what gets counted in that 97% looks more like assistants and workflow automation than fully autonomous agents in the technical sense. That’s my read, not something Writer spells out. The disappointment rate tracks closely with founders skipping straight to the third bucket before they’ve got the first two right.

Why most agent deployments fail

When I talk to founders and CEOs about AI rollouts that haven’t delivered, the same patterns keep coming up. The ones below are composite observations from those conversations, not specific clients.

The process wasn’t documented before the agent was built. If a human can’t write down what they do, an AI can’t do it either. The agent inherits the chaos and adds latency.

No owner. Engineering built it, then handed it to “the business.” The business didn’t know it existed, and nobody updated the prompts when policy changed. By month three the agent was confidently wrong about pricing tiers that had changed in February.

The wrong starting unit. Founders ask “what should our AI agent do?” The better question is “what’s the slowest, most error-prone, most repetitive workflow in this team?” Start there. The technology choice comes last.

No KPI. I asked one founder how he’d know the agent was working. He said “we’ll see if people use it.” That’s a hope dressed up as a metric.

Five questions to ask before you commission an agent

If you’re about to spend money on an AI agent or platform, sit with the answers to these before you sign anything.

  1. What’s the human-run baseline for this workflow? Time per ticket, error rate, cost. If you can’t measure today, you can’t prove improvement tomorrow.
  2. Who owns the agent after launch? Name the person. Their job is to maintain prompts, monitor failures, and own the metric.
  3. What’s the fallback when it fails? Hand off to a human, escalate, retry? Document the path before launch, not after.
  4. How will we know it’s working in 30 days? Pick one metric. Cycle time, deflection rate, first-contact resolution, cost per task. One.
  5. What’s our policy for the kinds of decisions the agent isn’t allowed to make? (Refunds above £X, hires, dismissals, contract terms.) Write it down.

If you can’t answer four out of five, you’ve got more scoping to do before you sign anything.

The operations playbook

This is what I run with clients when an AI deployment is the right answer.

Week 1. Map the workflow as it exists today. Time it. Find the failure modes. Most teams discover the problem is somewhere other than where they thought it was.

Week 2. Decide what stays human, what gets automated, what gets assisted. Build the metric.

Weeks 3 to 6. Build small. Pilot on 10% of volume. Measure against the baseline.

Week 7 onward. Scale only if the metric moves. If it doesn’t, fix the process. A different model rarely changes the answer.

This is how you join the minority getting real ROI from AI agents rather than the growing group calling adoption a disappointment.

Want help?

If you’ve got an AI rollout in motion and you’re not sure which group you’re heading for, the Ops Audit gives you an outside read on the operational fundamentals. Two weeks, fixed fee from £5,000. You walk away with a ranked list of workflows worth automating, the risks to clear first, and the next steps to take. Or start smaller: book a free 30-minute discovery call and we’ll work out whether it’s the right fit.

Sources

Book a Discovery Call