AI Solved Procurement's Information Problem. It Created a Judgment Problem.

A category manager once needed hours, sometimes days, to pull together a supplier review: contract language, old offers, market movements, meeting notes, cost drivers and internal constraints. AI can now compress much of that work into minutes. The improvement is real, but it changes where the difficult part of procurement sits.

When information becomes cheap to produce, the scarce capability is deciding whether a polished answer deserves to influence a commercial decision.

A 2026 study on trust calibration in human-AI decision making gives this problem a useful frame. Igbodor and colleagues combined evidence from 68 primary studies with survey data from 200 professionals and 30 interviews across several domains. Their findings are not procurement benchmarks, and the domains should not be collapsed into one, but the central idea travels well: strong human-AI performance depends on calibrated reliance rather than maximum trust.

October evidence: the context gap is not hypothetical

Icertis's The Context Imperative: Procurement's Million-Dollar Blind Spot (5 October 2026) surveyed 500 US procurement leaders. Nearly four in five respondents said they entered negotiations without all the information they needed, and a similar proportion said information discovered after a supplier decision would have changed their approach. Fifty-eight percent reported discovering payments above contract terms, while 72 percent reported learning of supplier underperformance only after business impact. One-third estimated annual losses exceeding $1 million from unrealised supplier commitments. These are vendor-sponsored, self-reported survey findings, not audited loss estimates.

This is a critical distinction for negotiation AI. A fluent recommendation cannot compensate for missing contract commitments, supplier history, qualification constraints or stakeholder knowledge. A transcript of the spoken negotiation is likewise evidence of what was said, not a complete account of the commercial mandate. Before recommending a concession or judging a buyer's move, the system must identify what context it has, what it lacks and which missing fact could reverse the conclusion. Preparation and post-conversation analysis can make that uncertainty visible without burdening a buyer with live prompts under pressure.

The dangerous answer is often plausible, not absurd

Procurement people are rightly alert to hallucinations, but fabricated facts are only the obvious failure mode. A system can produce a recommendation in which every sentence sounds defensible and the conclusion is still commercially weak because the frame is incomplete.

An AI might correctly note that a commodity index rose six percent and still mishandle a supplier increase because the supplier's actual exposure is different, productivity changed, the relevant period is wrong or the contract already allocates the risk. It can summarize a termination clause accurately while missing that switching suppliers requires nine months of qualification. It can build a strong negotiation plan around an alternative source that nobody has actually qualified.

These errors are harder to spot because the output looks professional. The uncertainty has not disappeared; it has been turned into coherent prose.

Good users interrogate the structure of the answer

A buyer using AI well therefore asks more than whether the answer sounds correct. Which parts came from facts supplied by the user, which came from sources, which are inferences and which assumptions were introduced because information was missing? Which single assumption would change the recommendation most if it were wrong?

That discipline matters because procurement information is distributed. Engineering knows qualification risk, finance knows budget limits, operations knows service failures and the buyer knows supplier history. The system sees only what reaches it. Our article on AI prompts for negotiation preparation makes the practical version of this point: before asking for a strategy, make the model identify the missing information that could change the strategy.

The purpose is not to distrust everything. It is to make reliance proportional to the evidence, the uncertainty and the consequence of being wrong.

AI is widening the buyer's job before the title changes

A buyer can now ask AI to interpret a liability clause, sanity-check a should-cost model, examine a supplier dataset or draft a financial case without handing each task to legal, analytics or finance. That is useful. It also means the boundary of the job can move faster than the boundary of the buyer's expertise.

OpenAI Economic Research published a useful signal of this in September 2026. Across more than 1.5 million sampled work-related ChatGPT messages from April through July, the researchers tracked how people used AI for tasks associated with occupations other than their own. Among roughly 6,200 consistently observed workers, previously used cross-occupation tasks rose from 13.1% of occupation-specific AI activity in April to 25.9% in July. In a separate matched comparison, workers returned to a cross-occupation task they had used the previous month 23.6% of the time, compared with 8.4% among similar workers with no observed prior use of that task.

This is not evidence that jobs formally changed, or that the work was done well. The sample is drawn from selected ChatGPT users, and the study measures observed use rather than output quality. What it does show is that some cross-functional use is recurring rather than a one-off experiment.

The way people prompted outside their usual occupation is just as interesting. They were less likely to ask for explanations or step-by-step instruction and more likely to provide context, examples or ask for verification. One plausible reading offered by the researchers is that people are borrowing expertise rather than trying to become experts in the adjacent field.

For procurement, that creates an awkward governance question. If a buyer can now perform work that looks partly legal, financial or analytical, who owns the judgment when the answer is plausible but wrong? AI can extend a buyer's reach without transferring the accountability, context or professional depth that used to sit with the specialist.

The judgment gap is now visible in workforce data

A global IBM Institute for Business Value study published on 21 September 2026 makes the judgment issue much more concrete. IBM and Oxford Economics surveyed 1,500 CHROs and equivalent workforce executives across 21 geographies and 23 industries, alongside 8,800 full-time employees across 28 countries.

Seventy-one percent of CHROs identified the ability to supervise, validate and override AI outputs as the workforce's most essential skill. Only 29 percent of employees ranked judgment as important. Sixty percent of employees said they worried AI was eroding their skills, while 41 percent of CHROs believed employees might not feel safe challenging or overriding AI outputs.

These are survey responses, not measured tests of employee judgment. They should not be read as proof that AI has already reduced critical-thinking ability by a particular amount. The useful signal is the mismatch: organisations increasingly say that judgment over AI outputs matters, while the people expected to exercise that judgment may not prioritise it to the same degree.

IBM also reports that 80 percent of CHROs see AI creating invisible work such as validating recommendations, fixing mistakes, providing context and managing exceptions. That is highly relevant to procurement. The human-in-the-loop role is not merely clicking approve. Someone has to know when the system is missing a commercial fact, when an assumption is stale and when the apparent answer does not fit the actual deal.

In procurement, explanation changes reliance

A September 2026 experiment in the International Journal of Industrial Ergonomics gives the issue a procurement-specific mechanism. Yining Ye and colleagues tested five explanation types in a simulated procurement decision task with 40 participants.

Participants preferred explanations that were informative, technically specific and logically coherent. Explanation type affected both subjective trust and whether participants accepted the algorithmic recommendation.

The study is exploratory and small. Participants were aged 20 to 34, with an average age of 24.5, and the task used a predefined, finite set of relevant information. Real procurement decisions are often messier, with missing context distributed across engineering, finance, operations, contracts and supplier history.

That limitation is almost the point.

If explanation quality changes reliance even in a bounded task, a professional using AI in a live commercial environment needs more than a recommendation. They need to see enough of the reasoning, assumptions and information boundary to judge whether the recommendation deserves authority.

This is also why a live negotiation transcript is not automatically enough context for coaching. The transcript can show the move that happened. It may not contain the mandate, switching cost, qualification risk or package trade that determines whether the move was commercially sound.

AI can become an internal anchor

Procurement already understands anchoring in supplier negotiations. The first credible number or frame can shape everything that follows, and an AI-generated recommendation can create the same effect inside the organization.

Imagine a category review that concludes an eight percent supplier increase should probably be negotiated down to three percent. The number may rest on incomplete market data and assumptions about leverage, but once it appears in a polished briefing it starts organising the discussion. Finance asks why procurement cannot reach three percent, the category manager adjusts around three percent and the negotiation plan begins treating a model estimate as an organizational reference point.

The human is still formally in the loop. In practice, however, disagreement has become more expensive than agreement because challenging the machine now requires explanation. That is a governance problem that cannot be solved merely by leaving the final approval button with a person.

Human oversight only matters when disagreement is genuinely possible

A 2026 conceptual paper in AI & Society describes a related question through “epistemic well-being at work”: whether professionals retain meaningful capacity to judge, contest and take responsibility when algorithmic recommendations carry organizational authority. The paper is conceptual, not evidence of a measured procurement effect, but the question is useful.

If a sourcing recommendation defaults to an AI score and overriding it requires an exception note or senior approval, the buyer may remain accountable while practical authority has shifted elsewhere. A human can sit in the loop and still become a rubber stamp.

Meaningful oversight therefore requires more than presence. It requires enough understanding, evidence access and organizational permission to say that a plausible recommendation is not good enough for this decision.

Procurement may be overtraining AI use and undertraining challenge

SAP's five-year Procurement Imperative analysis, published on 17 September 2026, gives this judgment problem a procurement-specific skills signal.

The latest study surveyed 2,648 C-suite executives across 23 countries. In the 2025 research within the series, 68.2 percent identified AI proficiency and ethics as the leading skills priority, while traditional procurement competencies were cited by 34.1 percent. Curiosity was prioritised by only 10 percent.

The same research series reports that by 2026, AI-driven and predictive insight had become the dominant category-management priority, cited by 66.6 percent of respondents.

These figures do not prove that procurement teams are losing curiosity or judgment. They do show where formal attention is going.

That creates a potential skills paradox. Procurement can become much better at operating AI while remaining underdeveloped in the behaviour required to challenge AI.

The practical skill is not generic curiosity as a personality trait. It is observable commercial behaviour: asking what evidence supports the recommendation, testing assumptions, looking for missing supplier context, challenging a convenient answer and knowing when a model output should not become the internal anchor.

A buyer who accepts a plausible AI-generated cost argument without probing it has essentially the same capability problem as a buyer who accepts a plausible supplier argument without probing it.

The better the system becomes, the more calibration matters

A poor tool keeps users alert because every answer feels questionable. A very useful tool creates the opposite risk: it is right often enough that checking starts to feel inefficient. The capability challenge is therefore not teaching people to resist AI; it is teaching them to know when confidence is warranted and when a high-consequence decision deserves another layer of scrutiny.

A mature user can rely on an output when sources are current, assumptions fit the case and the cost of error is limited. The same user can stop when a recommendation depends on an unverified switching-cost assumption and would shape a critical supplier decision.

Both are competent uses of AI. Maximum skepticism wastes the technology; maximum trust outsources judgment.

Negotiation exposes the boundary

Negotiation makes the issue visible because the prepared answer meets new information in real time. The supplier rejects the cost model, engineering reveals a longer qualification path, a price concession arrives tied to an unexpected volume commitment or the counterpart escalates. The buyer has to decide immediately what the new information changes and what it does not.

Voice2Evolve works at that execution boundary. A buyer can arrive with an AI-assisted plan, meet resistance from a simulated supplier and then review why the plan was followed, modified or abandoned. The useful measure is not obedience to the preparation; it is whether the decision to rely on or depart from it had a commercially defensible reason.

AI is making procurement much better at producing information. That is precisely why procurement needs to become better at judging which information deserves authority.

Sources

• IBM Institute for Business Value (21 September 2026). New IBM CHRO Study: AI Puts Critical Thinking at the Center of Workforce Priorities.

• Ye, Y., Zheng, J., & Ma, L. (September 2026). Decoding user preferences: A study of explanations for algorithmic decision support in procurement. International Journal of Industrial Ergonomics, 115, 104029.

• OpenAI Economic Research (16 September 2026). How workers are unlocking new ways of working, Work at the Frontier.

• Igbodor, E. et al. (2026). Trust calibration in human-AI collaborative decision-making: a cross-domain investigation of climate intelligence and organisational governance systems. Human-Intelligent Systems Integration.

• Salazar-Altamirano, M. A., & Ravina-Ripoll, R. (2026). Beyond efficiency: epistemic well-being at work under algorithmic governance. AI & Society.

• Borisov, I. G. (2026). From Outputs to Warranted Claims: Reframing AI Literacy as Culturally Mediated Quality Judgement for HRD. Human Resource Development Review.