# Agentic Commerce Is Becoming a Services Market Agentic commerce began with a shopping demo: ask for a product, let software compare the options, and authorize a checkout. That is useful, but it understates the larger change. Services are where agents can remove the most work from a purchase. Buying a service means turning an objective into a scope, finding qualified providers, collecting bids, asking follow-up questions, negotiating price and timing, monitoring delivery, and deciding whether the result is acceptable. Checkout is only the last step. The direction of the market is therefore not simply **agents buying products**. It is **agents running bounded procurement loops**, increasingly against agents representing the seller. Humans and companies remain the accountable principals, especially for expensive, regulated, or subjective work. But agents are beginning to operate both sides of the transaction. ## The existing services market is already large The strongest market signal is not a forecast. It is the amount of service purchasing that already happens through structured digital marketplaces. | Marketplace evidence | 2025 result | What it establishes | | --- | ---: | --- | | Upwork | **$4.0B** in gross services volume | A large market already uses digital discovery, proposals, contracts, milestones, and payment for knowledge work. | | Fiverr | Approximately **$1.06B** in marketplace GMV | Millions of buyers already purchase packaged and custom digital services online. | | Combined observed base | More than **$5B** | There is a meaningful transaction pool that could become more automated without inventing a new category of demand. | [Upwork reported $4.0 billion of GSV in 2025](https://investors.upwork.com/static-files/30673492-032c-4fea-b707-94969a5942f0). Fiverr reported [3.1 million annual active buyers and $342 in annual spend per buyer](https://investors.fiverr.com/news-releases/news-release-details/fiverr-announces-fourth-quarter-and-full-year-2025-results); multiplying the two rounded figures implies roughly $1.06 billion of marketplace GMV. The measures and customer definitions are not perfectly identical, so the combined figure is a directional floor for digitally mediated services, not a precise agentic-commerce market size. The wider opportunity is much larger but less certain. Upwork cites a commissioned estimate of a **$1.3 trillion flexible digital knowledge-work market in 2028**. That is a company-sponsored market study, not observed marketplace volume, so it is better treated as an upper boundary on the adjacent spend pool than as a forecast for agentic transactions.
$5B+ in observed digital service marketplace volume
Upwork's reported 2025 GSV plus directional Fiverr GMV shows that the digital service-buying machinery already exists.
This distinction matters: almost none of the $5 billion observed base should be called autonomous agent spending today. It shows that the supply, demand, contracts, and payment behavior already exist in digital form. Agentic commerce can begin by compressing the work between intent and purchase. ## Buyer agents are moving from search into sourcing The first wave of commerce agents concentrated on product discovery and checkout. OpenAI's [Agentic Commerce Protocol](https://openai.com/index/buy-it-in-chatgpt/) lets a merchant accept or decline an order while keeping payment, fulfillment, returns, and support in its existing systems. That is a clean model for products with known inventory and prices. Service procurement requires a longer state machine. The buyer does not know the final product at the beginning; the specification and price emerge through interaction. Upwork already exposes most of that state machine in its [GraphQL API](https://www.upwork.com/developer/documentation/graphql/api/docs/index.html). A permitted client integration can create a job posting, retrieve proposals, message and shortlist bidders, create a fixed-price or hourly offer, activate and approve milestones, and send payments. Upwork's April 2026 [ChatGPT integration](https://investors.upwork.com/news-releases/news-release-details/upworks-work-marketplace-comes-chatgpt) makes the consumer-facing direction even clearer: a business can describe a project, discover talent, and draft a job post in ChatGPT before continuing into Upwork to scope, contract, and begin the work. Enterprise procurement is advancing along the same path. Oracle's 26A [Autonomous Sourcing Assistant](https://docs.oracle.com/en/cloud/saas/readiness/scm/26a/proc26a/26A-procurement-wn-f41767.htm) can turn eligible requisitions into negotiations, invite suppliers, require a minimum number of bids, apply an award, and generate purchasing documents. Its policy file defines categories, bid duration, bid-count minimums, and maximum line amounts. Approvals can remain in the loop before publication and award. That is already agentic buying in a meaningful sense. The agent is not merely searching a catalog. It is translating policy into a competitive purchasing process. ## Seller agents are already answering the door The seller side is arriving through lead qualification rather than autonomous delivery. Fiverr's [Personal Assistant](https://help.fiverr.com/hc/en-us/articles/32545737221649-AI-Personal-Assistant-for-freelancers) responds to new prospective clients using the freelancer's profile, services, and chat history. It can answer questions, gather requirements, and hand complex conversations back to the freelancer. Fiverr's brief system then lets matched freelancers respond with personalized offers, while [custom offers](https://help.fiverr.com/hc/en-us/articles/360010559198-Creating-and-managing-custom-offers) can specify one-time, subscription, milestone, or hourly work. Upwork's Uma similarly helps freelancers analyze and write proposals and manage project tasks, while clients receive AI-assisted search, shortlisting, interviews, and contract generation. Neither marketplace needs to declare the provider itself autonomous for agents to negotiate on both sides of the front door. The likely near-term service transaction looks like this:
1

Scope

Turn the objective into a bounded request.

2

Contract

Agree on the work, price, and acceptance test.

3

Verify

Check the result before payment.

The agent's job is to carry one bounded service purchase from intent to an accepted result.
This is agents on both sides, but not two unaccountable bots improvising with money. Each agent acts for a principal. The buyer's principal owns the budget and authorization policy; the provider's principal warrants the result. ## Public access still lags the product capability The marketplace workflows exist, but they are not yet open agent infrastructure. Upwork requires an approved API key tied to a verified person with an active payment method. Its [API application requirements](https://support.upwork.com/hc/en-us/articles/115015857647-How-to-request-an-API-key-from-Upwork) say the API is not available for commercial use, and its [automation policy](https://support.upwork.com/hc/en-us/articles/43342677368467-Use-bots-and-other-automation-properly) limits automation to the approved use case. That makes the API plausible for an internal purchasing agent acting for a verified company, but not an unrestricted foundation for a third-party marketplace. Fiverr has the inverse shape. Its product supports briefs, matched offers, negotiation, and purchase, and the company advertises a [Fiverr API through its Platform Solutions partnership program](https://www.fiverr.com/partnerships/platform-solutions). But its public materials do not expose a self-serve buyer API for creating briefs, retrieving competing offers, accepting one, and paying. Unauthorized bots are prohibited by its terms. The current boundary is revealing: platforms are comfortable using their own agents inside controlled workflows. They are more cautious about letting arbitrary external agents become account operators. ## Fully autonomous commerce starts with machine services The autonomous edge of the market looks less like hiring a designer and more like buying a unit of computation, data, communication, or access. Stripe and Tempo launched the [Machine Payments Protocol](https://stripe.com/blog/machine-payments-protocol) in March 2026 so an HTTP service can return a payment challenge and deliver the result after the client pays. MPP supports one-time charges, recurring access, and metered sessions. Stripe's launch examples included browser sessions, physical mail, and even a sandwich order, but the native center of gravity is paid APIs and other machine-consumable services. A July 16, 2026 snapshot of the [public MPP services directory](/mpp) contained **137 services and 1,252 paid or runtime-quoted endpoints** across AI, data, search, compute, storage, blockchain, media, social, and web categories. Of those endpoints, 1,059 advertised fixed prices and 193 used dynamic pricing. These are provider-advertised listings, not verified demand or completed purchases; a live payment challenge remains authoritative. Mastercard's June 2026 [Agent Pay for Machines](https://www.mastercard.com/us/en/news-and-trends/press/2026/june/mastercard-launches-agent-pay-for-machines.html) points in the same direction: permissioned, high-frequency, low-value transactions across cards, accounts, and stablecoins, with examples spanning compute, data, freight, warehouse access, and monitoring. This segment automates first because the output is immediate and easier to verify. An endpoint either returns the contracted data or it does not. A design project, legal opinion, recruiting search, or security assessment has a longer delivery cycle and a more contestable definition of success. ## Two markets are forming, with a hybrid between them Agentic services are forming a continuum between two operating models.

Delegated

People deliver the work.

Hybrid

People and agents share delivery.

Machine

The service returns the result.

The models differ by one variable: how much human work remains inside each transaction.
The hybrid is likely to become the broadest category. It can combine machine speed at the edges with human accountability where judgment matters. A provider does not have to automate the entire job to make the purchase substantially more agentic. Automating intake, qualification, scope, quote, scheduling, status, and evidence can remove most of the coordination cost while leaving the crucial work with a specialist. This extends the economic argument in [What is the actual Point of agentic commerce?](/blog/actual-point-of-agentic-commerce): buyers save time, while sellers lower the cost of acquiring and serving a customer. Services create more room for both gains than a checkout button does. ## The contract becomes the product interface Product commerce can rely on a SKU. Agentic service commerce needs an executable contract. At minimum, that contract has to carry:

Executable contract

Boundsscope · budget · deadline

Acceptancetest · evidence

Recourseowner · escalation

A service contract replaces the SKU by making the purchase boundary explicit.
In practice, those three jobs require the contract to name the accountable buyer and delegated authority; bound the inputs, exclusions, timing, and spend; make offers comparable; define acceptance evidence; preserve funding and receipt records; and provide an escalation path for failure or dispute. The agent can reason about which offer best satisfies the objective. Authorization and settlement should remain deterministic. An April 2026 [IMF note on agentic payments](https://www.imf.org/en/-/media/files/publications/imf-notes/2026/english/insea2026004.pdf) makes this separation explicit: probabilistic orchestration belongs upstream, while policy checks, authorization, and settlement need fixed, auditable rules. This also prevents a dangerous version of agents on both sides: recursive subcontracting with no accountable provider. If one agent accepts a job and hires another agent, the customer still needs to know who warrants the final result, whether further delegation is allowed, and where recourse lands. ## The next evidence will be operational, not promotional Protocol announcements show that the infrastructure is moving. They do not show how much useful commerce agents are completing. The most informative market measures over the next two years will be: 1. **Request-to-contract time:** whether agents materially shorten scoping, bidding, and negotiation. 2. **Agent-originated service volume:** completed service spend, separated from product checkout and speculative payment traffic. 3. **Human exception rate:** how often policy, ambiguity, or risk forces a person back into the transaction. 4. **First-pass acceptance:** whether agent-scoped work is delivered correctly without rework or dispute. 5. **Repeat autonomous purchase rate:** whether principals allow the agent to buy again after observing an outcome. 6. **Provider economics:** whether seller-side agents reduce lead-response time, quoting cost, and customer-acquisition cost without increasing refunds. The base case is not that autonomous agents replace every buyer, salesperson, and service professional. It is that agents absorb the coordination layer around service work. Buyer agents will source and manage more routine purchases; provider agents will qualify and quote more demand; fully autonomous delivery will expand fastest where outputs are immediate and testable. Agentic commerce may have started at checkout, but its more consequential direction is toward contracting for work. The interfaces will increasingly be agent-to-agent. The obligations will still belong to people and companies. The payment stack for that world is mapped in [Agent-to-Agent Commerce, Layer by Layer](/blog/agent-to-agent-commerce). The unresolved question after payment—whether an unfamiliar provider will actually deliver—is examined in [The Barrier to Agent-to-Agent Payments Isn't Money. It's Trust.](/blog/trust-barrier-agent-to-agent-payments).