← workbypk.comWaveMiner · Research note 02 · Wave: agentic-ai-consumers
Agents as Consumers
What happens as AI agents start researching, buying and negotiating for people and companies, mapped through five stages. The wave is in stage 2 today, and stage 3's problems are already showing. Stage 3 is where a small operator with a controls background can get there first.
Patrick · workbypk.comPublished 24 Sep 2026 · evidence as of that datePrevious note: AI Data Center Spend MapEvidence tiers A–D under Sources
Where the wave stands
Stage 2 now, stage 3 next
1EnableBuiltPayment credentials, checkout protocols and agent identity
2AdoptNowAgents research; people mostly keep the final click
3ScaleNext · targetAgents buy within limits; errors, disputes and fraud surface
4MatureLaterAgents switch vendors and negotiate; loyalty pricing erodes
5ReconfigureLaterBusinesses built on human attention shrink; new intermediaries form
Why stage 2
The rails are built. Stripe and OpenAI's checkout protocol launched September 2025; Google, Shopify, Walmart and Target launched a shared commerce standard in January 2026; Coinbase's payment protocol moved to a Linux Foundation body with 52 members including Visa, Mastercard and Amex.1
Platforms are enrolling merchants for them. Shopify switched on agent checkout in ChatGPT for 2M+ eligible US stores on 11 March 2026.3 Amazon's "Buy for Me" agent completes purchases across 100M products from 400,000+ outside merchants.4
Shoppers arrive through AI, and they buy. AI-referred traffic to US retail sites grew 138% year over year in May 2026 and converted 54% better than other traffic.5
But people keep the final click. Only 11% of US consumers would let AI make the purchase decision, even for household supplies.7 A third want to approve every agent transaction, though they expect agents to make 15% of their purchases within five years, up from 9% a year earlier.8
Stage 3 is already visible
Nobody knows who pays for mistakes. US payment rules assume a person on the other side.10 Under Reg E a transfer is presumed authorized once the consumer gave the agent credentials; under Reg Z a "not as agreed" dispute pushes the loss onto the merchant.9 No card network had published binding agent-dispute rules as of August 2026.2
Networks now demand controls and records. Visa and Mastercard require agent providers to register, prove an authorization chain, enforce limits matching what the user delegated, and keep detailed records.9 Amex covers erroneous purchases by registered agents, though its announcement did not address fraud.11
Fraud has moved in. 47.9% of AI bot traffic on Akamai's network in the second half of 2025 hit commerce sites, with agent hijacking and synthetic identities named as the new attacks.12
Insurance is pricing the risk. AIUC ties premiums to independent audits of an agent's controls; Munich Re rebuilt its aiSure cover for agents in 2026. AIG, Chubb, WR Berkley and Great American added AI exclusions to standard policies from 1 January 2026.13
Stage 3 signals already seen: protocols handed to neutral standards bodies, first network rules, first insurance products. Still missing: regulator guidance, binding dispute rules, lawsuits at scale. Those are the arrival signals to watch (see forecasts).
The evolution map
Five stages, what breaks, who pays
Each stage creates different work. Opportunities are sorted by path: build something, supply something, service something, or own something that gains value.
1
Enable
Built
What happens
Card networks, payment processors and big tech build agent credentials, checkout protocols and agent identity: Visa's Trusted Agent Protocol (October 2025), Mastercard Agent Pay (April 2025), Google's payment protocol (now with the FIDO Alliance).1,2
Opportunity
None for a small operator. This layer belongs to the networks and platforms.
Signal it's done
Protocols handed to neutral bodies. Seen: FIDO (April 2026), Linux Foundation (July 2026).1
2
Adopt
Now
What happens
Agents research and compare; people approve. Platforms plug their merchants in by default. Business buyers get purchasing assistants.
What breaks
Product data. The average US retail product page is only 66% readable by AI models.6 Businesses off the big platforms are invisible to agents. Distributor sales reps lose their opening as agents compare suppliers on data and price.4
Who pays
Sellers who need to be found, especially B2B distributors and suppliers.
Opportunities
Service Catalog and data readiness for businesses off the big platforms: regional distributors, industrial suppliers, service firms that quote and schedule. Own Distributors with clean data and APIs gain share.4
Crowding
Consumer brands are crowded: Shopify does readiness for free, and "generative engine optimization" agencies are multiplying.3,16 Off-platform B2B looks thinner (inference).
3
Scale
Next · target
What happens
Agents complete purchases within limits people and companies set. Volume exposes errors, disputes, fraud, and spending nobody approved.
What breaks
Proof of what was actually authorized. Dispute evidence and reason codes. Merchant chargeback losses. Agent hijacking. Corporate spend control once employees' agents can buy.9,12
Who pays
Merchants (they carry chargebacks), corporate finance teams (they answer to auditors), and anyone seeking insurance, since insurers price on audited controls.2,13
Opportunities
Service → product Controls for agent spending inside companies: delegation limits, approval rules, audit trails, reconciliation, insurer-audit readiness. Service Dispute evidence and chargeback handling for agent orders. Service Independent agent-control reviews for underwriting.
Arrival signals
Regulator guidance on agent authorization; network dispute rules for agent transactions; first class action over agent purchases; audit-profession guidance on agent-initiated transactions.
Timing
Most signals likely within 12–24 months. Confidence: medium
4
Mature
Later
What happens
Agents shop continuously, switch vendors and renegotiate recurring bills. The "loyalty penalty" (higher prices for customers who never switch) stops working when a robot does the talking. Bill-negotiation agents already exist, and AI is pushing their fees from 35–60% of first-year savings toward flat subscriptions.14
What breaks
Renewal pricing in telecom, insurance and subscriptions. Brand loyalty in commodity categories. The distributor sales relationship.
Opportunities
Service Helping regional suppliers compete in agent-run procurement: pricing rules, data, ordering APIs. Own Businesses agents can't route around: physical fulfillment, local service, same-day availability.
Timing
2–5 years. Confidence: low to medium
5
Reconfigure
Later
What happens
Businesses built on capturing human attention shrink, and new intermediaries form around agent reputation and agent-to-agent deals. An early sign: Google search referrals to Chartbeat's publisher network fell 40.2% year over year to July 2026, after a 21.9% drop the year before.15
Who loses
Affiliate and comparison sites, ad-funded publishers, lead-generation brokers, retention call centers, some distributor sales teams (inference).
Opportunities
Own Buying displaced businesses that still have physical or relationship assets, at distressed prices. Service Helping them transition.
Timing
3–7 years. Confidence: low
Be one stage early
Three opportunities to test
Ranked on timing, evidence, and fit with the author's operations, risk-and-controls and financial-services background.
Stage 3 · Service, then product · Top candidate
Agent spending controls for mid-size companies
Thesis
As employees and systems let agents buy, finance teams need delegation limits, approval rules, audit trails and reconciliation that will satisfy auditors, card networks and insurers. Networks already require "controls matching delegation scope" and records;9 insurers price on audited controls.13
Buyer
Controllers and CFOs at mid-size firms; audit and advisory firms as a channel.
Founder seat
The practitioner. This is controls design work, and financial-services relationships are the channel.
Red team
Procurement suites (Amazon Business, SAP Ariba, Coupa) will build controls in; Amazon Business already flags out-of-policy spend.4 Big advisory firms will sell this to large enterprises. The mid-market may not feel pain until losses land.
Cheapest test
Ask five controllers or CFOs: is anyone at the company buying through an AI agent today, and what policy covers it? Would you pay $2,500–5,000 for an agent-spending controls review before your next audit? Kill it if fewer than two expect agent purchasing within 12 months.
Stage 2 → 4 · Service
Agent-ready catalogs for regional distributors
Thesis
Shopify and Amazon made consumer sellers agent-readable by default. Regional electrical, industrial and supply distributors with messy catalogs were not, and agent-run procurement routes orders to whoever has clean data and APIs.4
Buyer
Owner or COO of a regional distributor, reached through the contractors who buy from them.
Red team
Distributors' ERP and e-commerce vendors may ship this. Data cleanup is commodity labor unless it becomes an ongoing feed you maintain.
Cheapest test
Take three regional distributors' public catalogs, run 20 real contractor purchase questions through an AI assistant, and count the failures. Show the owner the result.
Stage 3 · Service · Watch
Dispute evidence for agent orders
Thesis
Merchants carry agent chargebacks, and no binding network rules exist yet.2 Logging what the agent was authorized to do becomes the evidence that wins disputes.
Red team
Chargeback specialists are already publishing agent-dispute playbooks, so this will crowd fast.10 Better as a partner or referral line than a business.
Cheapest test
Ask three merchants whether they can tell agent orders from human ones today.
Seen this from the inside?
Three questions for finance and procurement people
If you run finance, purchasing or a distribution business, your answer is worth more than any source on this page. Reply through patrick@workbypk.com. Short answers are fine, and anything you share is kept anonymous.
Today. Is anyone at your company buying anything through an AI agent or assistant yet, even informally?
Policy. If an agent made a purchase nobody meant to approve, which policy or control would catch it?
Suppliers. Have any of your suppliers or customers asked about making catalogs, pricing or ordering readable by AI agents?
Forecast log · made 24 Sep 2026
Predictions this map will be scored on
Each is dated, tied to an observable signal, and reviewed quarterly. If they miss, the stage timing above gets revised.
#
Prediction
By
Confidence
If it happens
F1
Visa or Mastercard publishes binding dispute rules or reason codes specific to agent-initiated transactions.
30 Jun 2027
Medium
Stage 3 confirmed; dispute-evidence work gets a rulebook.
F2
The CFPB or FTC issues guidance, a proposed rule, or an enforcement action specifically about agent-initiated consumer payments.
31 Dec 2027
Low to medium
Authorization proof becomes a compliance requirement, not a best practice.
F3
A major procurement suite (Amazon Business, SAP Ariba or Coupa) ships delegation controls built specifically for AI-agent purchasing.
30 Jun 2027
Medium to high
The controls opportunity shifts from product to mid-market implementation and review.
F4
In Global Payments' next annual survey, consumers' expected share of purchases made by AI agents within five years exceeds 20% (15% in May 2026).
2027 survey
Low to medium
Consumer comfort is rising faster than the timing above assumes.
Where waves meet
Intersections with other waves
Programmable moneyCoinbase's machine-payments protocol now sits with a Linux Foundation body, and Mastercard launched Agent Pay for Machines in June 2026: agents paying agents across cards and stablecoins.1,2
Data-center buildoutThe electrical and supply distributors on the spend map are exactly the B2B sellers whose catalogs purchasing agents will query (inference).
Business successionOwner-run distributors that don't adapt lose agent-routed orders, and some will be retiring owners open to a sale (inference).
Aging populationBill-management and negotiation agents for aging parents, run by their adult children (inference).
Method
How this was made
First manual run of WaveMiner's evolution mapper: identify the current stage with evidence, map what each stage requires and breaks, flag next-stage opportunities, and log dated predictions. Every stage claim names a signal that can be observed. Directed and reviewed by Patrick; research compiled with Claude (Anthropic).
Provenance
Sources
Tier A primary · B high-quality analysis · C industry reporting · D vendor marketing or weak signal. A D-tier source can suggest a lead but cannot validate it on its own.
Data file: sources.csv (every source on this page with its tier).