White paper · Liz Eversoll · September 2026

Keep Your AI Pilot From Failing

Why AI stalls as it scales, and the two steps that keep it running.

Inside this paper

  • Why pilots stallThree causes, none of them the model.
  • The five phasesConnected through to policing, and where it breaks.
  • Two steps that workOne governed home, then one real workflow.
  • The cost of waitingA model you can run on your own numbers.
75%

of AI initiatives have not delivered the expected return, according to CEOs. Only 16% have scaled enterprise-wide, and 64% say the fear of falling behind drives them to invest before they understand the value.IBM Institute for Business Value, 2025 CEO Study, about 2,000 CEOs.

Executives are asked to do three things at once

Accelerate AI, prove the return, and stay in control. Those goals pull against each other, and the failures show up in the pipeline: in S&P Global Market Intelligence’s 2025 survey, 42% of companies had abandoned most of their AI initiatives, up from 17% the year before.

Most companies are well into AI. Claude, ChatGPT or Copilot is connected to the files and teams are getting real work done. Then every team copies company knowledge and rules into its own projects, skills and agents. The copies drift, and IT responds by restricting use.

The fearWhat it looks likeThe outcome
SprawlCopies of the same knowledge and rules in every AI tool, project and skillOne approved version, owned in one place
SlopAnswers built on outdated or conflicting knowledgeEvery agent works from the approved version, and you can see which one it used
SecurityAgents reaching data and taking actions no one approvedTrellis decides what each agent may do as it acts
SpendRising AI budgets, plus the hours lost to duplicate workLess duplicated work, measured in a pilot
Enterprise AI has an authority problem. The tools made company knowledge easy to reach. Nothing in them decides which version is the approved one, who it applies to, or what governed a result.

Five phases most companies move through

Our framework, drawn from conversations with enterprise buyers, not an industry study. At each phase teams fix the problem in front of them and find the next one.

01 Connected

AI reaches the files. The agent finds three versions of a policy and cannot tell which is current.

02 Copied

Teams paste policies into their own projects. Answers improve, and the same rule now lives in dozens of copies.

03 Drifting

A rule changes. The old version stays in every copy that held it, and each change adds more stale ones.

04 Consolidating

IT names authoritative sites and shared skills. The knowledge inside still needs a re-publish each time.

05 Policing

AI is limited to low-risk uses and people check output before it ships. Reviewers become the bottleneck.

Why pilots fail

Rules and knowledge have no single home. One rule can sit in a rules engine, application code, a policy document, an agent prompt, an AI skill and someone’s head at once, with nothing keeping them in sync. The agent uses whichever version it finds first.

RAND’s research on AI project failure found the most common root cause is simpler still: the people involved misunderstand what problem the AI is meant to solve. The third reason is that the process gets automated as it stands, so handoffs that existed only because of old tools are rebuilt inside the new one.

Retrieval can find five conflicting versions of a policy perfectly. That is successful retrieval and failed authority. RAG optimizes relevance. Charter establishes authority.

Two steps keep a pilot running

Step one

Give your knowledge one home

One governed place every agent reads: policies, business rules, standards, processes, product facts, and the code and design standards your engineers use. Each item gets an owner, a version and an effective date. You don’t have to settle everything first — everyone already works from some version today, just different ones.

Step two

Prove it in one real workflow

Pick the step that causes the most pain, record how the people doing it really work, put the rules for that step in Charter, and get it running. Then pick the next, using what you learned. Each improvement stays small, measurable and easy to change.

What the record answers

Identity and access tools decide which systems an agent can reach. They don’t hold your business rules for what it may then do. Charter determines which knowledge and rules apply, makes the governance decision, governs the action, and records the source, version, rule, decision, approval and action.

That record answers the question that arrives months later, when a customer disputes something or an auditor asks. The Air Canada chatbot case showed a company can be held responsible for what its chatbot tells customers. Charter doesn’t expose a model’s reasoning and doesn’t prove a model obeyed a rule. It preserves evidence of what governed the work.

Read the full paper

Twelve pages: the five phases, the sales worked example, the cost-of-waiting model you can run on your own numbers, and what changes for each leader.

Related

Explainer

Prove what governed the work

Replay a decision from last month against the corpus state in force at the time.