02 November, 2026 / 15 min read

AI Procurement: Why Companies Disappear From the Shortlist Unseen

Share:LinkedInXFacebook

White robotic hand selecting a single red wooden figure from a row of identical light-wood figures, illustrating AI-driven procurement shortlists.

More and more large companies let AI agents pre-sort suppliers before a human ever sees a shortlist. A provider that machines read as contradictory or wrong can drop out of the process without ever finding out. This article describes the risk, separates what is documented from what is merely expected, shows it using publicly verifiable data of a consulting firm, and lists concrete countermeasures.

The Illusion of Human Control

When consulting services are procured, one principle still holds: in the end, a human decides. That is true, but it falls short, because what matters is which providers that human gets to see in the first place.

For decades, the award of strategy, transformation and audit mandates followed a familiar pattern: board conversations, personal recommendations, long-standing relationships and a formal tender process, the Request for Proposal (RFP). The longlist was compiled by hand by analysts or the purchasing department, mainly based on market presence, references and industry knowledge.

Once companies deploy AI agents that collect, compare and score provider data, the bottleneck moves forward. The human decision still takes place, but over a selection that a system has prepared. A provider that does not appear there is not rejected, it is simply never considered. There is no rejection letter and no stated reason that could be challenged.

A clarification up front: this article describes a development and a risk. It does not claim that this kind of pre-filtering is how most corporations work today. What is documented is that large companies are introducing AI agents in purchasing. How far they go when selecting consulting firms is hardly documented publicly. That is exactly why checking now is worthwhile: while the systems are still being built, corrections to your own data are cheap.

BMW: What Is Documented and What Is Not

BMW provides a current example. On 13 October 2026, Handelsblatt reported on the company's plans to save billions with artificial intelligence (GEP). According to the reports that followed, BMW already uses four AI agents in purchasing, with an annual purchasing volume of around 80 billion euros. Head of purchasing Nicolai Martin reportedly expects initial savings in the mid hundreds of millions of euros and is thinking in billions over the long term. By 2028, BMW wants to be an “AI-orchestrated company” (all-ai.de).

Two clarifications matter.

  1. The billions are a target, not a result. Several reports explicitly stress that these are long-term ambitions and not savings already achieved (Finanztrends). One source also points out that BMW has not officially quantified a specific amount saved by AI alone (MS Aktuell).
  2. The known applications concern supply chain and capacity planning. In one reported example, an agent merged planning data from three divisions, flagged bottlenecks and proposed countermeasures, a process that previously could take weeks. That BMW uses agents to pre-filter consulting firms is not documented.

What the example does show: corporations are building infrastructure in which agents merge data from many sources, evaluate it and make proposals. It is plausible that such systems will play a larger role in the selection of suppliers and service providers in the future. That is a reasoned expectation, not a fact. This article therefore treats it as a scenario that companies can prepare for.

Shortlist 0: A Model for Silent Pre-Selection

Classic tenders know the longlist, a broad market overview, and the shortlist, the bidders invited to submit a proposal. We use Shortlist 0 to describe the set of providers that an automated system takes into the longlist in the first place. The term is our own working model and not an established market standard. It is meant to describe where part of the selection could take place in the future.

What is special about this level is that dropping out remains invisible:

Stage 1: Market scan

  • Classic process: Market presence, reputation, personal contacts.
  • Scenario with AI pre-filtering: A system evaluates public sources.

Stage 2: Pre-selection

  • Classic process: Longlist compiled by analysts or purchasing.
  • Scenario with AI pre-filtering: Shortlist 0, proposed by the system.

Stage 3: Tender

  • Classic process: RFP sent to the known set of bidders.
  • Scenario with AI pre-filtering: RFP sent to the bidders who passed the pre-selection.

Stage 4: Decision

  • Classic process: Presentations, negotiation, award by humans.
  • Scenario with AI pre-filtering: As before, but from a smaller selection pool.

Stage 5: Risk for the provider

  • Classic process: A rejection is visible and can be queried.
  • Scenario with AI pre-filtering: Not being considered goes unnoticed.

Which data such a system uses depends on the individual case. Plausible sources include commercial registers, company websites, press releases, knowledge bases such as Wikipedia and Wikidata, news archives and the answers of large language models. Where these sources contradict each other or are out of date, uncertainty arises for the system. Whether it treats that uncertainty as a risk depends on its configuration. That it can be read as a warning sign is reason enough to check.

Why Classic SEO and PR Are Not Enough

Search engine optimization and public relations target people who search, click and read. An automated system works differently: it cross-checks statements from several sources. Three differences matter in practice.

Rankings are only one part of the data basis. Many AI systems use web search and therefore also search indexes, but they weight sources differently from a human. In addition, training data, knowledge bases and registers feed in. A good ranking for a keyword does not guarantee that an AI answer describes the company correctly.

Structured data helps but is no guarantee. Schema.org markup, for example as JSON-LD, makes facts such as legal form, locations or service areas machine-readable. That all systems treat this information as authoritative truth is not proven. It does, however, increase the chance of an unambiguous match, especially when external sources say the same thing.

Contradictions between sources are the actual problem. Press releases, archive articles, register extracts and third-party profiles state different figures, legal forms and ownership details, often for good reasons: different reference dates, different corporate levels, different definitions. Humans put this into context. An automated system may read it as uncertainty about the provider's identity. We call this the Identity Ambiguity Risk.

The practical consequence: visibility in classic search and visibility in AI answers are two different tasks. Anyone who only does the first does not know what the second shows.

Case Example Roland Berger: Where Public Data Leaves Room for Interpretation

We chose Roland Berger as an example because the company is well documented publicly and is one of the best-known strategy consultancies in Europe. Roland Berger did not commission this analysis and was not consulted for this article. We do not claim that the company has lost contracts or that any particular AI system describes it incorrectly. We show which publicly verifiable facts an automated system could read in different ways. All information is current as of 2 November 2026; the sources are listed at the end of the article.

A. Independence and Merger History

In November 2010, Deloitte confirmed “advanced discussions” about a merger with Roland Berger; the combined business was to be called “Roland Berger Deloitte Strategy Consultants” (Accountancy Age). A few days later, the partners rejected the merger (Accountancy Age). Further talks followed in 2013: Deloitte had returned, and PwC and EY had also made contact. According to reports, Deloitte and PwC withdrew in September, while the EY offer remained as an option (Consultor.fr). Roland Berger remained independent and describes itself as a partnership exclusively owned by its partners (press release 2019).

Possible reading by a system: The archives contain many reports from a time when the outcome was still open. If a system treats older reports as equal to current self-descriptions, it might consider the ownership status unclear. Whether this actually happens can be tested with targeted questions (see the “Self-Test” section).

B. Figures at Different Levels

Several different figures on size and structure circulate publicly, most of which can be explained by reference date and definition:

Employees: approximately 2,700

Source: Börsen-Zeitung, 2013.

Employees: approximately 2,400

Source: Press release, 2019.

Employees: approximately 3,000

Source: Press release, more recent issue.

Employees: approximately 3,500

Source: Wikipedia), infobox.

Employees: 4,645

Source: Built In, third-party profile, June 2026.

Partners: approximately 250

Source: Consultor.fr, report from 2013.

Partners: approximately 230

Source: Press release, 2019.

Shareholders of the holding company: 196

Source: Wikipedia), as of December 2017.

Legal form of the holding company: GmbH, later GmbH & Co. KGaA

Sources: Wikipedia); Austrian company register (shareholder of the Austrian subsidiary, entry of 24 October 2025).

Each individual figure may be correct for its reference date and level: group or single entity, partners or shareholders, salaried consultants or total workforce. A human recognizes this. A system that merges figures without reference dates sees five different headcount figures and at least three labels for the shareholder level.

Possible reading by a system: Deviations between website, press text and register data could be read as an inconsistency in self-presentation. Here too, this is a hypothesis that can be tested, not a finding.

C. Topic Profile: Restructuring and AI

In May 2026, Roland Berger reported revenues of 1.01 billion euros for 2025, the highest in the firm's history, and names profitability and data-driven, AI-supported solutions as its focus for 2026 (press release). In August 2026, the company reported a new recognition in the area of IT, technology and telecommunications in the Forbes ranking of the best management consulting firms (Roland Berger).

The self-image is therefore clearly oriented toward technology and data. Whether language models nevertheless associate the company mainly with older topics such as restructuring, because far more was written about them over decades, remains open. This is precisely the core of the Semantic Disconnect: self-image and machine image do not have to match, and you only find out by testing.

D. What the Example Shows and What It Does Not

It shows that even well-run companies leave public data traces that differ over time and in content. It does not show that a procurement system would therefore screen out Roland Berger. Such a statement would not be credible without documented tests and without knowledge of the specific systems. The example serves as a pattern for the question every company should ask itself: what would a system find about us, and does it contradict itself?

The AIPM Risk Fields™: Three Recurring Patterns

The example yields three risk fields, which we at EYECONIFAY™ call the AIPM Risk Fields™ (AI Perception Management™). They apply across industries and are intended as a checking framework. The “Possible consequence” column describes a scenario in each case, not a proven sequence of events.

1. Identity Ambiguity Risk

  • Typical trigger: Contradictory or outdated information on owners, legal forms, merger or acquisition history.
  • Possible consequence in an AI-assisted pre-selection: The system cannot determine the provider's identity unambiguously and rates it as hard to verify.
  • Check question: Do all sources state the same ownership structure and the same legal name?

2. False Risk Trigger

  • Typical trigger: Deviating figures between website, press texts and registers, often due to different reference dates or definitions.
  • Possible consequence in an AI-assisted pre-selection: A deviation is read as a lack of transparency or data quality, even though each figure is explainable on its own.
  • Check question: Does every figure state its reference date and level (group, single entity)?

3. Semantic Disconnect

  • Typical trigger: The company's self-image (current focus areas) is weaker in the sources than its historical image.
  • Possible consequence in an AI-assisted pre-selection: The system assigns the provider to the old topic area and does not consider it for new topics.
  • Check question: Which service areas does an AI answer name first, and do they match today's strategy?

The fields do not depend on one another. A company can be unremarkable in one field and have a problem in the next. What matters is that all three can be influenced by working on your own data sources, before anyone else exploits or misreads them.

Recommendations: Consistent Data in 90 Days

The following steps apply regardless of industry and company size. They do not address causes inside any particular AI system, but the one cause you can influence yourself: contradictory public information about your company.

Days 1 to 30: Inventory

  • Create a fact sheet: An internal “single source of truth” with company name, legal form, registered office, ownership structure, subsidiaries, headcount and service areas, each with reference date and level.
  • Review public sources: Website, commercial register, press releases, Wikipedia and Wikidata, industry directories, review portals, careers pages, social media profiles.
  • List deviations: Every statement that differs from the fact sheet, with its location and an assessment (outdated, different level, wrong).
  • Run a baseline test: The prompts from the next section in at least three AI systems, fully documented.

Days 31 to 60: Cleanup

  • Align your own sources: Website, press kit, boilerplate texts and profiles state the same facts in the same wording, with a reference date.
  • Have third-party sources corrected: Update directories and profiles you can access yourself; on Wikipedia and Wikidata, add only verifiable information with sources, disclose conflicts of interest and follow the platform's rules.
  • Put historical topics in context: If older events (merger talks, restructurings) are often misread, publish a clear, factual account on your own website that states the current position.
  • Add structured data: Schema.org Organization markup (name, legal form, address, founding date, parent and subsidiary companies, sameAs links to registers and profiles).

Days 61 to 90: Anchor and Monitor

  • Substantiate service areas: Back up current focus areas with references, publications and contact persons on your own pages so that external sources can pick them up.
  • Introduce monitoring: Repeat the baseline test every quarter and compare the deviation rate.
  • Assign responsibility: One person owns the fact sheet and passes every change (renaming, acquisition, new figure) on to all channels.
  • Define an escalation path: What happens if an AI answer stays wrong for good? Secure the documentation, check the sources and, where applicable, use the providers' correction forms.

We at EYECONIFAY™ call this triad the AI Perception Management™ (AIPM) remediation path: a clear factual foundation (Knowledge Graph Optimization), machine-readable markup (schema architecture) and ongoing control (Continuous AI Auditing). The effect it has on a particular system can only be measured, not promised. That is why the test belongs at the start and the monitoring at the end of every project.

Self-Test: Check Your Company in 30 Minutes

You do not need special software. What matters is that you document carefully, because AI answers fluctuate and can differ from run to run.

Procedure:

  1. Choose at least three systems (for example ChatGPT, Claude, Gemini or Perplexity) and note the date and model version.
  2. Enter each prompt with identical wording, five times each in a new session.
  3. Save the complete answers and screenshots.
  4. Compare every statement with your fact sheet and mark deviations as “wrong”, “outdated” or “different level”.
  5. Calculate for each prompt in how many runs the deviation occurred.

The six check prompts (replace [Company] with your name):

Prompt 1

“Who is [Company]? Describe its ownership structure and independence.”

What it checks: Identity Ambiguity Risk.

Prompt 2

“How many employees does [Company] have? Give the source and the date.”

What it checks: False Risk Trigger.

Prompt 3

“What is the legal form of the parent company of [Company]?”

What it checks: Identity Ambiguity Risk.

Prompt 4

“Name the five most important service areas of [Company].”

What it checks: Semantic Disconnect.

Prompt 5

“Which providers are suitable for [your future topic, e.g. AI transformation in the automotive industry]? Explain why.”

What it checks: Semantic Disconnect, findability.

Prompt 6

“What risks speak against hiring [Company]?”

What it checks: False Risk Trigger.

Evaluation: A wrong answer in one of five runs is an indication; a wrong answer in most runs and across several systems is a call to action. The counterpart matters too: if your company is not named at all for prompt 5 even though it works on that topic, that is a visibility problem even if no answer is “wrong”.

Write down what you do not know: your test shows how a system responds to a question, not how your customers' procurement systems make decisions internally.

Frequently Asked Questions

What is “Shortlist 0”?

A working term for the set of providers that an automated system takes into the longlist of a tender before humans make the selection. It describes a scenario and is not an established technical term.

Do corporations really use AI agents in purchasing?

Yes. BMW, for example, reports four AI agents in purchasing. Whether and how they pre-filter consulting firms is not publicly documented.

Is a good website and classic search engine optimization enough?

They are necessary but not sufficient. AI answers draw on several sources, which should be consistent with one another.

What does it cost to get started?

The self-test and the fact sheet cost working time but no budget. Cleaning up third-party sources and ongoing monitoring are more demanding.

Can you make sure an AI always describes the company correctly?

No. But you can raise the probability by removing contradictions in the sources, and you can measure deviations.

Conclusion

When consulting mandates are awarded, humans still decide. But how many providers those humans see is increasingly determined by systems that evaluate public data. The risk for companies is less that someone says something false about them than that their data contradicts itself and a machine derives uncertainty from it. The Roland Berger example shows that even well-known firms leave data traces that differ depending on reference date and level.

The good news: these contradictions are within companies' sphere of influence. Anyone who maintains a fact sheet, aligns their sources and regularly tests what AI systems say about them improves their starting position, regardless of how quickly AI-assisted procurement spreads. In this phase, the winner will not be whoever communicates the loudest, but whoever presents the same, verifiable and current picture of themselves everywhere.

Notes on This Article

Methodology and date. All information on Roland Berger and BMW is based on publicly available sources (see below) and is current as of 2 November 2026. We make no statements about the behavior of individual AI systems toward Roland Berger. Terms such as “Shortlist 0” and “AIPM Risk Fields™” are working models of EYECONIFAY™.

Independence. There is no business relationship between EYECONIFAY™ and Roland Berger. Roland Berger was not consulted for this article and neither commissioned nor reviewed it. The presentation serves to explain a principle using a publicly documented example.

Corrections. If any information is outdated or incorrect, please let EYECONIFAY™ know. Substantiated corrections will be incorporated, stating the date of the change.

No legal or investment advice. This article is a general professional assessment and does not replace legal or financial advice in an individual case.

Sources

BMW and AI in purchasing

Roland Berger: history and structure

_________________________

Zeljko Perencevic

Founder & CEO of EYECONIFAY™ | Originator of AI Perception Management™

EYECONIFAY logo

Request an Executive Risk Briefing

Tell us who you are. Our team will contact you within 2-3 business days to arrange a no-obligation introductory call.

Please use your company address. Private addresses (Gmail, Outlook, etc.) are not accepted.

EYECONIFAY logo

Reserve your copy of the European Executive AI Perception Risk™ Report 2027

Enter your details and we will send you the report when it is published. The reservation is free and non-binding.

Please use your company address. Private addresses (Gmail, Outlook, etc.) are not accepted.

EYECONIFAY logo

Contact Us

Send us a message and our team will get back to you.