Pilots by invitation · The Digital Foundry is the build studio of Digital Edge · Pricing on request

Home / Insights / Why agent pilots stall before production

insights · delivery

Why agent pilots stall before production

Most agent pilots do not fail on the model. They fail on everything around it: scope, ownership, integration and a decision nobody agreed to make.

Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from under 5% in 2025. The same firm predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, on escalating cost, unclear business value or inadequate risk controls. Both forecasts can be true at once. Agents are arriving; many of the projects built around them will not survive.

BCG’s 2025 research points the same way: 60% of companies report little or no value from AI, and only 5% achieve value at scale. The difference is rarely the technology. It is what happens between a promising demo and a system people rely on.

Five reasons pilots stall

1. The pilot proved the demo, not the workflow

A pilot that runs on clean sample data in a sandbox proves the model can do the task. It does not prove the agent can do the task inside your workflow, on your data, with your exceptions. When the real cases arrive, the results change and confidence goes with them.

2. Nobody owned the outcome

Innovation teams run pilots. Operations teams run workflows. If the person who owns the workflow was not part of the pilot, nobody on the business side is accountable for taking it into production, and nobody fights for the budget.

3. Success was never defined

“Let’s see what it can do” is not a success criterion. Without a baseline, a target and a decision date agreed before the pilot starts, the end of the pilot becomes a debate about whether it worked. Debates rarely end in funding.

4. Integration and run cost were left for later

The pilot read from an export and wrote to a spreadsheet. Production needs live integration, access control, monitoring and someone on call. When those costs appear after the pilot, the business case built on pilot numbers no longer holds.

5. Governance arrived at the end

Security, privacy and risk teams see the build for the first time when it is ready to go live. They ask reasonable questions: what data does it touch, who approved this decision, where is the log, what happens under POPIA section 71? If the answers require a rebuild, the project stalls.

What we do instead

  • Write the decision first. Before we build, we agree what the pilot must show for you to fund production, and when you will decide.
  • Pilot inside the real workflow. Real cases, real systems, least-privilege access, a person at every checkpoint that matters.
  • Price production before the pilot ends. Integration, run cost and support are estimated while the evidence is fresh.
  • Bring governance in on day one. Logging, oversight and data boundaries are part of the pilot design, so the review confirms rather than discovers.
  • Treat a stop as a result. A pilot that shows a workflow is not worth automating has saved you a production budget.
the short version

An agent pilot is a test of a decision, not a test of a model. Define the decision, and the pilot has somewhere to go.

That is the shape of every Digital Foundry build: scope, proof of value, production with governance in place. See how a build goes.

Sources

  1. Gartner press release, 26 August 2025: 40% of enterprise apps will feature task-specific AI agents by 2026, up from less than 5% in 2025.
  2. Gartner press release, 25 June 2025: over 40% of agentic AI projects will be cancelled by end of 2027.
  3. BCG, The Widening AI Value Gap, 2025.
  4. Protection of Personal Information Act 4 of 2013, section 71.
more notes

Keep reading.

insights/governance

Designing agents for POPIA section 71

Section 71 limits decisions made solely by automated processing. Build the human checkpoint in from the first design, not after the security review.

Read the note
insights/scoping

Choosing the first workflow to automate

The first agent sets the tone for everything after it. Pick the workflow that can prove value quickly, safely and in numbers your finance team accepts.

Read the note
next step

Have a workflow worth proving?

Tell us about one workflow. If it is a good candidate for a proof-of-value pilot, we will say so, and say why.