By Seth Earley, Founder & CEO, Earley Information Science
────────────────────────────────────
Published: October 27, 2026
Last Updated: November 9, 2026 | Version 1.0
────────────────────────────────────
Who This Is For: AI/ML leaders, CIOs, Chief Data Officers, program and portfolio managers, and procurement and vendor management leaders responsible for evaluating AI. Also valuable for e-commerce and digital commerce executives preparing for AI shopping agents, and for finance, risk, and audit leaders who review AI business cases.
Prerequisites: Basic familiarity with enterprise AI pilots and vendor evaluation.
────────────────────────────────────
Executives are rarely rejecting the technology when they shelve an AI evaluation. They are rejecting the instrument that produced the result.
Picture the meeting. A team presents months of evaluation work. The figures are sound, and everyone in the room also knows that the people presenting them had a stake in the outcome. Leadership thanks the team, sets the findings aside, and the program drifts. This scene recurs across industries and across a decade of selection and assessment work. It is the most common fate of internal AI evaluations, and the least discussed.
Calling this politics misses the point. It is reasonable skepticism, and the evidence supports it.
Every AI number arrives attached to a perspective, and frequently to a bias. A vendor designs its own benchmark, and vendor benchmarks have a habit of ranking the vendor first. Public leaderboards are not immune. A 2025 study of Chatbot Arena, the most closely watched public ranking of AI models, found that certain providers privately tested many model variants and published only the best score. One provider tried 27 variants before a single release. An integrator's assessment tends to recommend more integration. An internal evaluation tends to defend the roadmap the team proposed and the budget it received.
None of this requires bad faith. People can measure honestly and still lean toward their incentives, because they selected the test set, the criteria, and the conditions. Executives discount figures from any party with a stake in the result, and they are right to do so.
The discount compounds. Programs stall when results lack credibility, even good results. Funding pauses. Pilots repeat with new variables and no meaningful controls or comparisons. Meanwhile the organization tries new tools while conditions shift, and leaders switch vendors or approaches without ever isolating what drives results. Each deployment ends up incompletely documented, incompletely understood, or both.
MIT's Project NANDA quantified the pattern in 2025, reporting that roughly 95% of enterprise generative AI pilots produced no measurable effect on profit and loss, a result attributed to integration and evaluation rather than to the models themselves. The pilot-to-production gap usually gets blamed on models and infrastructure. A meaningful share of it is an evidence gap: no one produced a measurement that a skeptical audience could accept.
A biased measurement can often be spotted from its conditions alone, without a detailed audit. Be wary when the test sets changed between evaluation rounds, when success criteria were written after the results arrived, when the methodology is undocumented or lives in one person's head, or when the evaluation ran under demo conditions on curated data with the vendor at the controls. Be wary too when the comparison omitted the alternative that nobody on the team was championing, because then the preferred option never met its strongest competitor. The more of these conditions are present, the less weight the result deserves.
Independence is not an abstraction. It can be specified as a set of engineering choices.
Document the methodology before testing. Stakeholders agree on evaluation criteria, weights, and scoring rubrics, and those are written down and versioned before any evaluation runs. If the criteria change, the change carries a stated rationale and the evaluation is repeated. Medicine offers a natural experiment on this rule. After 2000, large US clinical trials had to register their primary outcome in advance. A later analysis of major cardiovascular trials funded by the NHLBI found that the share reporting a significant benefit dropped from 57% to 8%. The authors call it an association, not proof, but the direction is hard to dismiss: when the success criterion cannot be chosen after seeing the data, far fewer results look like wins.
Run every option through the same process. Use scripted scenarios, identical data, and identical tasks, scored by one methodology under production conditions rather than laboratory ones. A head-to-head in which each vendor demonstrates its own best scenario compares presentations, not systems.
Use instruments that predate the initiative. Take two evaluations of the same support assistant. In the first, the vendor's team writes 20 test questions and reports 94% accuracy. In the second, 200 questions are sampled at random from last quarter's ticket log and scored by the managers who own the queue, against resolution codes already recorded in the ticketing system. Both yield a number. Only the second yields one that leadership cannot easily dispute, because the instrument existed before the project and was not built to flatter it.
Make the result reproducible. A different team given the data, scripts, methodology, and scores should arrive at the same result. An output that cannot be reproduced under the same conditions will not withstand scrutiny.
Count the full cost. Total cost, including operations, data remediation, and change management, is the foundation of an honest ROI. Partial costs erode credibility fast.
This discipline becomes structural once AI agents begin selecting products and completing purchases, because results and costs will face wide scrutiny and anecdote cannot carry the stakes. Every party in such a transaction will publish its own scoreboard. The platform reports engagement. The agent provider reports task completion. The retailer reports conversion. The brand reports its share of selections. Each figure can be true and self-serving at once, because each party counts what suits its objectives and use cases. That is four scoreboards from four perspectives, each serving its own constituency.
Selections made by agents will be challenged, especially in marketplaces and sponsored rankings. When an agent chooses one supplier's product over another's, procurement will ask why. Channel partners will ask why. In regulated categories, auditors will ask why. If the model's decision is opaque, no answer will hold up.
Regulators have begun writing the standard. The EU's Digital Markets Act bars designated gatekeepers from ranking their own products above rivals' and requires transparent, fair, and non-discriminatory ranking conditions, so an unexplainable ranking is already a compliance exposure for the largest platforms before any agent enters the picture. Standards for reproducibility and consistency also exist in other AI application domains. The statistical nature of generative AI makes consistency an inherent challenge, but organizations are building the structures, processes, and frameworks to measure it against objective standards.
Todd Barr, CEO of Axonis, made a related argument on the Earley AI Podcast. The standard for AI-assisted decisions, he said, should include supporting artifacts: records of the data used, the model applied, the prompts issued, and the reasoning followed, so that a decision can be audited later and improved. Commerce has not adopted that bar yet. Regulators, procurement departments, and channel-conflict lawyers will adopt it on its behalf.
That points to a role the market has not filled. No interested party can grade its own performance, and platforms and search vendors cannot referee a game in which they hold a stake. Measurement independence becomes a service worth buying, as it already is wherever numbers carry consequences: financial audit, clinical trials, safety certification. AI results are joining that list. Organizations that operationalize the discipline early will spend their time and resources on ongoing improvement rather than endless pilots or suboptimal deployments.
First, what is the methodology, and when was it created? If the criteria and weights post-date the outcomes, the evaluation was reverse-engineered from a preferred conclusion. Second, who benefits from the number? An agenda does not disqualify a result, but it needs to be disclosed openly. Third, can the results be reproduced? Another team should be able to process the data, scripts, framework, and scores without special guidance. If reproducing the result requires the original team, the approach is not sound.
Teams that can answer well on all three will find their results accepted rather than challenged. A single weak answer leaves a claim without support, and it will be treated that way.
Read the original version of this article on VKTR.