How AI Agents Choose Vendors in Enterprise Procurement: What Gartner's 2028 Forecast Actually Says

Gartner's forecasts make agent-mediated purchasing a present planning problem, but they do not publish one universal rule for how an AI procurement system will shortlist vendors. The defensible response is to make a company's claims verifiable across the sources an agent may be authorized to consult, then measure whether those claims are actually retrieved, cited, and used in evaluation.

Two Gartner forecasts are relevant and distinct. In April 2026, Gartner said that by 2028 more than half of enterprises would stop paying for assistive intelligence such as copilots and favor platforms that commit to workflow results. Gartner described delegated authority across identity, permissions, policy enforcement, systems of record, and auditability—not an AI-search vendor-ranking formula. (IT-Online, April 2, 2026)

Separately, a report on Gartner's 2025 strategic predictions said AI agents could handle 90% of B2B purchases and intermediate more than $15 trillion in spending by 2028 through automated exchanges using verifiable data feeds and standardized trust frameworks. The report does not identify a single retrieval stack, source hierarchy, shortlist algorithm, or required media channel. (Digital Commerce 360, November 28, 2025)

What the Gartner forecasts establish

The April forecast is about workflow architecture. Gartner's quoted distinction is whether AI has delegated authority to trigger actions within enterprise policy and identity constraints. It also forecasts margin pressure for software vendors that merely layer AI over legacy applications rather than redesigning around agentic execution.

The purchasing forecast is about the possible scale of agent mediation. It describes machine-to-machine negotiation, data verification, contracting, and execution with minimal human intervention.

Together, those forecasts support three operating questions for a vendor:

1. Can a machine identify the company, product, category, and evidence behind its claims?
2. Can the evidence survive the buyer's policy, security, legal, and procurement constraints?
3. Can the team observe which sources and claims appeared during an actual agent-assisted evaluation?

They do not establish that ChatGPT, Perplexity, Google AI Overviews, or any public answer engine supplies the citation graph for enterprise procurement. They do not establish that earned media is the primary input, that one publication placement causes shortlist inclusion, or that a missing public citation means automatic exclusion.

Human validation has not disappeared in 2026

Forrester's January 2026 business-buying research points in the opposite direction from an immediate no-human-review story. Forrester says generative AI is reshaping discovery and evaluation, but buyers increasingly rely on internal and external networks to validate and de-risk decisions. Its typical buying decision includes 13 internal stakeholders and nine external influencers; procurement is a decision-maker in 53% of buying cycles. (Forrester, January 21, 2026)

Forrester's 94% figure refers to buyers in groups of six or more who report benefits from the larger group. It is not the share of all buyers using AI. The report also says human interactions remain essential because AI answers can be incomplete or unreliable. Boundary: Forrester's 94% figure is buyer-survey behavior evidence; it does not establish that AI answers cite, recommend, or shortlist a brand, or that citation presence produces pipeline or revenue.

The useful planning tension is therefore bounded: Gartner forecasts more delegated execution by 2028, while Forrester observes large human buying groups and active validation in 2026. Neither source establishes the precise date when a specific category, company, or transaction will become autonomous.

A four-layer vendor evidence audit

A vendor cannot infer the hidden internals of every procurement agent. It can inspect and improve the evidence layers those systems may encounter.

1. Identity

Check whether public and first-party sources agree on the company name, product names, category, ownership, location, leadership, and canonical URLs. Record contradictions rather than hiding them inside a single visibility score.

2. Claim support

For every claim likely to matter in evaluation—security certification, integration, price, implementation time, customer outcome, geographic coverage—record the source, measured unit, date, population, and owner. Separate first-party documentation from independent validation.

3. Machine access

Verify that the relevant evidence is reachable in rendered HTML and machine-readable representations, with stable canonical identity and no contradictory structured data. Machine access is a prerequisite for evaluation, not proof of retrieval or selection.

4. Observed use

Run representative procurement prompts and workflows across the systems buyers actually use. Capture whether the brand was retrieved, which source was cited, how the claim was described, whether it entered a shortlist, and what happened next. Retrieval, citation, recommendation language, shortlist inclusion, referral, pipeline, and revenue are separate outcomes.

What distribution research can—and cannot—add

Stacker and Scrunch's December 2025 pilot compared eight articles across 944 prompt-platform combinations on five AI platforms. Citation coverage rose from about 8% for brand-only citations to about 34% when syndicated-only and co-citations were included. (Stacker, December 2025)

The expanded 2026 study measured 87 distributed stories from 30 brands across eight AI platforms. It reported a 239% median lift and higher cross-platform coverage, while explicitly labeling the research observational rather than definitive causal proof. (Stacker, 2026)

The measured unit in both studies is citation coverage for selected distributed stories and generated prompt sets on the tested public AI platforms. The studies do not establish that distribution causes a given enterprise procurement agent to retrieve a vendor, that editorial coverage outranks product documentation, that a specific outlet guarantees shortlist inclusion, or that citation produces a purchasing outcome.

That boundary still leaves a useful test: when independent coverage contains precise, source-linked facts, compare how often those facts surface against equivalent claims available only on the vendor's domain. Treat the result as a measured channel contribution, not a universal engine rule.

The Machine Relations frame for agentic procurement

Machine Relations treats machine-facing reputation as an observable system rather than a promise that publicity controls AI recommendations. In procurement, the discipline is to make important claims identifiable, attributable, current, independently checkable where possible, and measurable across machine-mediated evaluation.

The relevant asset is not an assumed “citation authority” score. It is an evidence map connecting:

• the claim a buyer or agent needs to verify;
• the first-party system of record;
• any independent source that tests or corroborates it;
• the machine-readable representations that carry it;
• the observed retrieval, citation, description, and shortlist outcomes;
• the commercial result, measured separately.

This is citation architecture used as evidence infrastructure. It does not require pretending that all agents use the same sources or that external coverage is always superior to current product, security, legal, or pricing documentation.

What to do before 2028

Do not build a countdown around an invented training-data deadline. Gartner's cited forecasts do not say that enterprise agents in 2028 will be trained primarily on content published in 2025 or 2026, and public retrieval systems can change their indexes, providers, and source behavior much faster than a model-training cycle.

Use a shorter operating loop:

1. Choose one procurement decision. Define the category, buyer constraints, and claims that determine eligibility.
2. Build the evidence map. Link each claim to its current first-party record and any independent corroboration.
3. Repair contradictions and machine access. Fix entity conflicts, stale facts, blocked pages, and mismatched structured data.
4. Run dated tests. Record prompts, providers, sources, citations, descriptions, and shortlist outcomes.
5. Change the weakest evidence layer. Publish a missing primary source, correct an unsupported claim, or pursue independent validation where it would reduce decision risk.
6. Remeasure. Do not call visibility, recommendation, pipeline, or revenue improved unless the corresponding observation changed.

Start with what current systems can find and verify about the brand today: AuthorityTech's visibility audit.

FAQ

Does Gartner say AI agents will make 90% of B2B purchases by 2028?

Digital Commerce 360 reports that Gartner forecast AI agents would handle 90% of B2B purchases and intermediate more than $15 trillion by 2028. That is a market forecast, not a guarantee for every industry or transaction, and the report does not specify one vendor-shortlisting algorithm.

Does Gartner say earned media determines which vendors procurement agents select?

No. The cited Gartner reporting discusses delegated execution, verifiable data feeds, trust frameworks, policy, identity, permissions, and auditability. It does not establish earned-media primacy, a publication hierarchy, or a universal source-selection mechanism.

Are enterprise purchases already fully autonomous?

Not as a general rule established by the sources on this page. Forrester's 2026 research describes large human buying groups, strong procurement involvement, trials, and validation through internal and external networks. Gartner's 2028 statements are forecasts about increasing agent mediation and delegated execution.

What did the Stacker distribution studies measure?

They measured citation coverage for selected stories, generated prompts, and tested public AI platforms. The larger study describes its evidence as observational. Neither study establishes that a placement guarantees retrieval, recommendation, shortlist inclusion, pipeline, or revenue for a particular vendor.

How should a B2B company prepare for agentic procurement?

Make procurement-relevant claims current, attributable, machine-accessible, and independently checkable where appropriate. Then test actual workflows and keep retrieval, citation, recommendation language, shortlist inclusion, referral, pipeline, and revenue as separate measurements.

Related Reading

• How AI Security Companies Build Earned Media and AI Search Citations in 2026
• PR for AI Search: How Earned Media Drives AI Citation Authority