RESEARCH / EVIDENCE BRIEF · GAEO.ai

Applied AI and agent commerce

An AI agent can compare offers, prepare an order or take an authorised action for a shopper. Each step calls for a different level of evidence. A useful product recommendation can be assessed against the shopper's criteria. An order also requires current availability, permission to spend and a way to resolve errors after purchase.

How this brief uses evidence

This is a synthesis of published research by GAEO.ai. It brings together surveys on AI adoption, a controlled experiment on task performance, a shopping benchmark, retail analytics and a market scenario. Each measures a different outcome. Where we suggest tests for operators, those are evaluation designs for local use.

What does the research show?

Value gap (2026). In McKinsey's August 2026 global survey of 1,719 respondents in 97 countries, 80% reported that AI makes them personally more productive, 37% reported seeing any enterprise EBIT impact, and 6% qualified as AI high performers—distinguished by redesigning workflows, visible senior commitment, and active risk management of agents. McKinsey, The state of AI in 2026: On the road to ROI.
Shopping-agent consistency (2026). In Visa's 5 November 2026 evaluation of 22 AI models from 12 labs across 15 shopping challenges (three attempts each), models failed 42.8% of runs, 54% of failed runs still ended in a purchase, single-model pass rates ranged from 82.2% down to 28.9%, and 34.2% of scenarios passed on all three attempts. Visa press release, 5 Nov 2026.
Adoption (2025 baseline). In McKinsey's 2025 survey, 23% of respondents reported scaling an agentic AI system in at least one function, and a further 39% reported experimenting with agents. McKinsey, The State of AI in 2025.
Value at scale. BCG classified 5% of firms in its 2025 global study of 1,250 companies as achieving substantial AI value at scale. BCG, The Widening AI Value Gap.
Task performance. In a randomized experiment involving 758 BCG consultants, GPT-4 users completed 12.2% more subtasks and worked about 25% faster on tasks within the model's capabilities, while performance deteriorated outside that boundary. Harvard Business School AI Institute, Back to the Beginnings of AI at Work.

How is AI changing shopping behavior?

Three measures of AI in the shopping journeyThe measures use different populations, time periods and units. Read each on its own.

$3–5TGlobal B2C goods revenue agents could help orchestrate in 2030Scenario under adoption and merchant readiness assumptions. US range: $900B–$1T.McKinsey · 17 Oct 2025
+693.4%Growth in visits from generative AI tools to US retail sitesNovember–December 2025 versus 2024. Referrals grew from a small base.Adobe Analytics · 7 Jan 2026
20%Share of global Cyber Week orders influenced by AI and agents25 Nov–1 Dec 2025. Salesforce counts recommendations and conversational customer service.Salesforce · 5 Dec 2025

Market scenario. McKinsey estimates that by 2030 agents could orchestrate $900 billion to $1 trillion in US B2C retail revenue and $3 trillion to $5 trillion globally. The scenario depends on shopper adoption and merchant readiness, and covers goods only, excluding services and B2B. McKinsey, The Agentic Commerce Opportunity.
Visits from AI tools. Adobe Analytics recorded a 693.4% year-over-year increase in visits from generative AI tools to US retail sites during November 1 to December 31, 2025, from a small base. The measure tracks shoppers following links from AI tools. Adobe Analytics, 2025 Holiday Shopping Season.
Influence on orders. Salesforce attributed AI or agent influence to 20% of global orders in its Cyber Week 2025 analysis, covering November 25 to December 1. Its definition includes personalised recommendations and conversational customer service, so a shopper can still place the order themselves. Salesforce, 2025 Cyber Week results.

Track AI-referred visits, orders with AI assistance, and orders placed by an authorised agent as separate metrics. For each, record the measurement period, geography and denominator.

How can an operator test readiness?

Discovery. Give the agent a defined product or itinerary request. Check whether the recommendation satisfies the stated requirements, uses the correct variant and cites the current offer.
Transaction. Test in a simulated checkout with explicit spending limits and confirmation. Include price changes, out-of-stock offers, retries and duplicate requests. Record whether each attempted action stays within the customer's permission.
Servicing. Test order lookup, returns, refunds or itinerary changes. Measure whether the agent resolves the request correctly or gives a human enough context to take over.

Compare each stage with the existing workflow on the same cases. Record correct completions across repeated runs, human intervention, time per resolved case and material errors. Review edge cases separately, especially where an incorrect action costs money or leaves the customer without a usable order.

For retail, a useful test is whether an agent respects size, delivery deadline and return constraints after a price change. For airlines, check the total price with bags and seat selection, then test whether the passenger can change the resulting order. These are proposed test cases from GAEO.ai.

What should an operator measure first?

Start with a single order journey and log three events independently: a referral from an AI tool, AI assistance before a customer places an order, and an order executed under delegated authority. A referral tells you how a shopper arrived. An assisted order describes a touchpoint. A delegated order requires evidence of permission, an action and an outcome. Report each against its own denominator and retain an audit trail for retries and reversals.