Enterprise & Work

Andus Labs Index: AI Pilot Sprawl Tops Enterprise Risk Ranking

Andus Labs' third Ground Truth Index, released November 18, ranks runaway AI experiment buildout and shallow assistant adoption as the leading enterprise barriers, drawing on 2,100 conversations across 700 organizations.

Andus Labs Index Puts AI Pilot Sprawl and Workflow Gaps at the Top of Enterprise Risks
Andus Labs Index Puts AI Pilot Sprawl and Workflow Gaps at the Top of Enterprise Risksgeralt / pixabay
By James Calloway5 min read

Why it matters

  • Andus Labs released the third edition of its Ground Truth Index on November 18, identifying 25 priority patterns.
  • The index draws on 2,100 meetings and conversations across 700 organizations and a library of 332 patterns.
  • Four of the top eight ranked patterns concern institutional coherence — whether structures and spending support intended changes.
  • Nine of the 25 patterns are new; 16 appeared in previous editions.
  • Five AI analyst agents independently score patterns on gravity, velocity, proximity, progression, and scale probability.

Andus Labs' third Ground Truth Index, released November 18, ranks runaway AI experimentation and shallow assistant adoption as the two leading barriers to enterprise AI value — ahead of any technical limitation. The ranking names 25 priority patterns, nine of them new, drawn from a library built on 2,100 meetings and conversations across 700 organizations.

The report's sharpest claim is a distinction most adoption dashboards obscure: access to AI tools is not the same thing as an operating model that can use them. "Tool access alone does not create an operating model that can use AI effectively," the index states in its key points. Enterprise AI can look busy, Andus warns, without changing how a company actually operates.

What tops the ranking?

Andus ranks "Runaway Buildout" first and "Copilot Ceiling" second. Runaway Buildout describes experiments launched without a common basis for assessing results — pilots with no agreed question, baseline, or decision rule. Copilot Ceiling describes organizations where AI assistants become familiar tools for drafting and summarization but do not substantially alter business workflows.

Four of the top eight patterns concern what Andus calls institutional coherence: whether organizational structures and spending actually support the changes a company says it wants. Beyond the top two, that category includes "Artifact Inflation," ranked sixth — review overload caused by polished-looking AI output — and "Swim Lane Sprint," ranked eighth, in which disconnected teams independently build similar capabilities.

The practical issue is not that experimentation is wasteful. A pilot can answer a valuable question even when the answer is to stop. The problem emerges when teams run experiments with no agreed measure of success, because high usage may demonstrate curiosity or convenience without showing that a process became faster, more reliable, or less expensive once review and integration costs are counted.

Andus offers a concrete illustration: two departments testing separate document assistants could each report strong engagement. That still leaves unanswered whether the tools duplicate one another, whether either reduces end-to-end turnaround time, and who would own a shared deployment. Those are organizational decisions, not model-selection problems.

Where does accountability fit?

The next three entries in the ranking are "Empty Chairs," "Learning While Drowning," and "Shadow Agency." Andus uses these labels for, respectively:

  • missing oversight of AI decisions;
  • training added to already crowded workloads;
  • unapproved AI use involving sensitive business information.

These concerns connect through ownership. A team may have permission to adopt a tool while lacking the authority to redesign the process around it. Employees may be expected to learn new methods while still being measured against an unchanged workload. A system may be allowed to make recommendations without a named person responsible for reviewing exceptions.

The index cautions that fixing this mismatch requires more than another approval checkpoint. A useful control should specify the decision, the accountable owner, the evidence that owner needs, and the circumstances requiring intervention. Without that, "oversight itself can become another queue that absorbs the time an AI tool was supposed to save."

The takeaway: permission to use AI without authority to change the work produces risk without returns.

How is the index built?

The methodology is worth examining before treating the ranking as data. Andus says its pattern library draws on those 2,100 meetings and conversations across 700 organizations and now contains 332 patterns. Five AI analyst agents independently score the patterns from different professional perspectives across five dimensions:

  • gravity
  • velocity
  • proximity
  • progression
  • scale probability

A median score forms the baseline, after which human researchers review the top 50 and adjudicate disagreements before publishing the final 25 priorities.

The index also incorporates English-language news coverage, large US companies' financial filings, and regulatory signals. Andus is explicit about the limits here: the findings are not a representative survey of enterprise adoption and not a measurement of financial losses. They are a diagnostic framework assembled from field observations, AI-assisted scoring, and researcher judgment.

Two caveats follow directly from that design. The underlying conversations are not a random sample, so the ranking should be read as Andus' assessment of priorities rather than the share of businesses experiencing each problem. And taking the median of several AI assessments does not establish that those assessments are independently accurate — agreement among five agents is not the same thing as correctness.

Why does this matter for buyers and boards?

For enterprise leaders, the index offers a way to interrogate an AI portfolio without assuming every deployment is failing. The central question Andus poses is what changed in the work: which handoff disappeared, which decision improved, which review burden fell, or which new responsibility became clear.

That evidentiary standard cuts both ways. A credible scale-up decision needs operational evidence alongside adoption figures — and a credible shutdown decision needs the same. An index that reframes the question from "how many experiments are we funding" to "can we compare them, learn from them, and take responsibility for what happens next" addresses the governance gap that pure usage metrics leave open.

The stakes are straightforward. Companies now measure AI engagement in licenses, tokens, and assistant sessions, while the Ground Truth Index argues the relevant metrics sit in process redesign, ownership, and decision quality. Vendors selling copilots and enterprises buying them both have incentives to report adoption; almost nobody has an incentive to report that the work itself stayed the same.

Andus' next editions will show whether the pattern library — 332 patterns and growing — turns into a durable benchmark or remains one consultancy's structured judgment. For now, the index's value lies less in its rankings than in the discipline it imposes: every AI deployment should be able to name the question it answers and the person accountable for the answer.

Source: unite.ai

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. Only 10% of Enterprises Let AI Agents Decide Without Human Approval
  2. 67.8% of Workers Now Use AI Weekly — But 56.4% Get No Time to Learn It
  3. AI Assistants Now Write the B2B Shortlist — and 60% of the Lists They Trust Are Competitors' Own Posts
  4. 51% of AI projects miss targets, yet 85% of firms raise budgets
  5. AI Agents Cheat on Their Own Tests, Benchmark Shows

« Previous articleNext article »