The shopper who is not a person
The 2026 story of agentic commerce is a reality check, not a retreat. On March 4, 2026, OpenAI pivoted its checkout effort to merchant-controlled checkout: the agent-side Instant Checkout experiment had reached only about 30 Shopify merchants, sales tax handling was unresolved, and multi-item carts never shipped. Meanwhile Google AI Mode crossed 75 million daily active users with checkout live at select merchants, and two protocols are racing to standardize the rail: ACP, the Agentic Commerce Protocol backed by OpenAI and Stripe, and UCP, backed by Google and Shopify.
Read those facts together and the direction is unambiguous. Assistants are moving from recommending products to carrying purchase intent, but the transaction itself is landing back on the merchant's own checkout. The agent shops; you close.
For telehealth operators that should sound familiar, because a funnel where a third party delivers a briefed, high-intent visitor to a gate only you can open is already your business model. The question is whether an agent can read your storefront well enough to deliver anyone at all.
The state of the rails
| Rail | Backed by | Status, August 2026 | What it means for telehealth |
|---|---|---|---|
| ACP (Agentic Commerce Protocol) | OpenAI, Stripe | Pivoted to merchant-controlled checkout on March 4, 2026, after agent-side Instant Checkout reached only ~30 Shopify merchants, with sales tax unresolved and no multi-item carts | The agent hands intent to your checkout instead of completing it, which fits Rx flows perfectly |
| UCP | Google, Shopify | Rival protocol; Google AI Mode at 75M+ daily active users, with checkout live at select merchants | Mainstream reach already exists; structured product data decides which programs agents surface |
| Your own storefront | You | The destination both roads currently end at | Machine-readable programs and a clean handoff win regardless of which protocol wins |
The strategic takeaway is that you do not need to pick a protocol in 2026. You need a storefront both can read, because the losing move is being illegible to all of them while a competitor publishes structured pricing.
Why the agent can never finish the job
In retail, the endgame is an agent that completes the purchase. In telehealth, that endgame is legally impossible, and it always will be for prescription programs.
A prescription is not a purchase. It is the output of a care process: a state-licensed provider reviewing a real intake, identity verification, informed consent given by the actual patient, and pharmacy dispensing rules. No commerce protocol changes any of that, and no responsible operator should want it to.
What an agent can legitimately do is everything before and around that gate:
- Research and compare programs on price, inclusions, and medication options
- Check published eligibility criteria against what its user has told it
- Prefill demographic and history context the patient confirms later
- Deliver the patient to intake with the decision already made
That is most of the funnel we mapped in Anatomy of the Modern DTC Telehealth Funnel, executed by software in thirty seconds. The clinical gate is not a bug in your agentic story. It is the part of your story that deserves the most design attention, because handled well it reads as trust, not friction.
Building the agent-legible storefront
Agents choose among programs they can parse. Four investments make yours parseable.
Machine-readable program and pricing data. Publish plans, monthly prices, what is included, refill cadence, and medication options in structured form: clean HTML tables, consistent naming, schema markup on program pages. If your pricing is "book a call," the agent recommends whoever published a number. This is the same discipline that wins AI citations in Generative Engine Optimization for Telehealth, now doing double duty as commerce infrastructure.
Structured eligibility. State availability, age ranges, exclusion criteria, and lab or BMI thresholds, stated plainly on the page. This lets an agent pre-qualify its user honestly, which protects your intake from unqualified handoffs and protects patients from dead ends.
Stable URLs and consistent entities. Agents build maps of your catalog over time. Renaming programs quarterly and rotating URL structures breaks the map. Keep program names, URLs, and brand entity data boringly consistent.
A machine surface, not just a web page. This is where we have placed our own bets: embeddable widgets that capture the handoff on any page, and a GraphQL API that exposes program, pricing, and intake structure directly, so the storefront humans see and the data agents read come from one source of truth.
Designing the human handoff
The patient an agent delivers behaves like the cohort we described in When Your Next Patient Arrives via ChatGPT: briefed, compared, and impatient with repetition. The handoff design goals follow from that.
Confirm the context. The receiving page should acknowledge what the agent established: the program, the price, the medication option under discussion. A generic homepage that makes the patient re-find their choice loses a meaningful share of them.
Do not restart the interview. Intake should accept what is already known: prefilled demographics the patient verifies, a free-text on-ramp for history, branching that skips answered questions. The patient confirms; they should not re-type.
Make the human-only moments explicit. Identity verification, consent, and provider review should be framed, in one sentence, as the reason an agent could not finish this: prescription care requires a licensed clinician reviewing your real information. Patients trust a gate that explains itself.
Instrument the channel. Tag assistant referrers and agent traffic distinctly, then track handoff-to-intake and intake-to-approval rates against your other channels. This cohort should convert well; if it does not, your receiving experience is dropping what the agent carried. The third-party trust layer that feeds agent recommendations, Reddit very much included, is doing its job. The last hundred meters are yours.
What to build this quarter
- Publish pricing and program structure on crawlable pages with schema markup
- Rewrite eligibility as structured, scannable criteria on every program page
- Freeze program names and URLs; fix redirects for anything already renamed
- Build the agent-referral receiving state and prefill-aware intake
- Add assistant and agent referrers to your analytics segmentation
None of this is speculative work. Every item also improves SEO, AI citations, and human conversion today, which is exactly why it is worth doing before the protocol race resolves.
FAQ
Can an AI agent buy prescription medication for a user? No. Prescription therapy requires a licensed provider's review, identity verification, and the patient's own informed consent, and no commerce protocol changes that. An agent can research programs, compare pricing, check published eligibility, and deliver the patient to intake, but the clinical steps are human by law.
What are ACP and UCP? ACP is the Agentic Commerce Protocol backed by OpenAI and Stripe; UCP is the rival protocol backed by Google and Shopify. Both standardize how AI assistants carry shopping intent to merchants. After OpenAI's March 2026 pivot to merchant-controlled checkout, both roads currently end at the merchant's own checkout rather than inside the assistant.
How should a telehealth brand prepare for agentic commerce? Publish machine-readable program, pricing, and eligibility data, keep URLs and program names stable, add schema markup, and build a receiving experience that continues where the agent stopped instead of restarting the patient. The same work improves AI citations and human conversion immediately.
Does agentic commerce matter if agents cannot complete the purchase? Yes, because the agent controls the comparison step. It decides which two or three programs its user ever sees, based on what it could parse and verify. Losing that step means losing the patient before your funnel begins.