Skip to main content

Agentic Commerce

Your in-app assistant already sees your store — sales, catalogue, the traffic funnel, business health. Agentic commerce lets it act on what it sees. It can prepare the things it would otherwise only suggest — a gift card for a lapsed customer, a promotion on slow-moving stock, a dynamic-pricing rule that targets a customer segment — and it handles the reversible catalogue housekeeping outright. Agentic commerce has a shopper side as well: the Agent API lets any AI shopping assistant search a store, price a basket and hand a checkout to the shopper, described at the end of this page. When a request names something the assistant cannot make exactly, it says so first and tells you which mechanism it chose, or asks. A request for a sale or a discount that does not say how gets one short question — a markdown, which shoppers see as the lower price everywhere, or a promotion applied at checkout? An explicit request such as “20% off my graphic tees” is acted on directly. Anything that moves money, goes live to shoppers, or changes something already live arrives as a draft you approve — a gift card, a promotion, a pricing rule — and is inert until you do. The organisational actions, which are reversible, happen directly: creating a subcategory, enabling a category, tagging a product, grouping products into a collection.

What it can do

Base product prices are off limits — the assistant only ever touches dynamic-pricing rules, never a product’s own price. Anything a plan doesn’t include, or your role isn’t permitted to do, it won’t offer. Written content needs your storefront URL set first (CMS → Settings). The product links in a post are built against it, so without one they would only resolve on your own site and break as soon as the post is shared — the assistant says so and stops rather than writing a post that falls apart. A lookbook hotspot claims a product is at that point in that picture, so it is only ever placed from what is actually visible in the scene. A product the assistant cannot find in the frame is left off and reported by name, rather than given a hotspot at an approximate position. If a scene cannot be produced at all, the lookbook is still created, with your products confirmed against the catalogue and a plain note that they are not placed yet.

Approving a draft

Every draft is decided on Suggestions, near the top of the sidebar beside Dashboard. It appears there only when something is waiting, with a count beside it, and it is available on every plan. Three tabs — Waiting for you, Approved and Rejected — each filterable by type: gift cards, promotions, pricing rules, campaigns, search rules, wholesale quotes and credit limits. Every entry shows what is being proposed, the reasoning behind it, when it was prepared and when it expires. Approving a gift card is the moment it issues and the customer is notified; approving a promotion or a rule is the moment it goes live. Rejecting keeps the record — it moves to the Rejected tab rather than disappearing, so you can see what was suggested and declined, and you can always ask for it again. A suggestion nobody decides eventually expires on its own, which is counted separately: expiring is not the same as saying no. You can also approve a suggestion right in the chat when the assistant proposes it, and its launcher shows a count when any are waiting. Opening Gift Cards, Marketing or Pricing while something of that kind is waiting shows a short notice offering to take you there — once a visit, and the lists themselves stay free of drafts. An advertising campaign the assistant drafts is not here: campaigns are reviewed under Integrations → Meta Ads or Integrations → Google Ads.
Approving an AI suggestion is its own permission, separate from the pages themselves. A store owner can grant a teammate — an admin, a marketer — the authority to approve suggestions without changing anything else about what they can do. The check is enforced on the server, not just hidden in the interface.

What it aims at

A suggestion is only as good as what it is pointed at, so the assistant reads what shoppers wanted and did not buy — not only what sold.
  • Carts — which products sit in baskets that have gone quiet, and what they are worth. A cart whose owner has since bought that product is not counted; they did not abandon anything.
  • Abandoned checkouts — who reached the last step and stopped. The strongest signal in the group, because the shopper had already decided.
  • Wishlists — what has been saved for later, the longest-lived signal of the three.
  • Search outcomes — what shoppers searched for, which product they clicked, and whether that led to a basket, a checkout or a purchase.
The last one matters because it names where a search failed, and the three answers need different fixes: clicked but never added to a basket points at the product page or the price; added but never reached checkout points at delivery cost or stock; reached checkout but never paid points at payment. A promotion aimed at what shoppers are actually trying to buy beats one aimed at what merely sells slowly. And a product that is wanted but out of stock is a restocking decision — the assistant will not propose a discount on something nobody can buy.

Margin protection

Every value the assistant proposes is in your store’s currency, and every money action protects your margin. It reads each product’s cost to make sure a discount never sells below cost, and it prefers higher-margin or slow-moving stock. When too many products have no cost price set, the margin figure it would be checking against is itself unreliable, so it declines to prepare any gift card, promotion, or pricing rule and points you to Products → Bulk edit to fill in the missing costs. The reversible catalogue actions stay available in the meantime.

Dynamic workflows

A request in conversation runs once and ends there. A dynamic workflow is the same intent kept running: ask for something ongoing — “keep my featured collection showing what’s actually selling”, “every week, check for slow stock” — and it becomes a rule that re-evaluates on a schedule, and also as soon as something it depends on changes in the store: a product selling out, new products arriving, a burst of orders. They are dynamic rather than fixed: each evaluation re-reads the store’s current figures and decides whether anything warrants doing, so the outcome follows the business rather than a predetermined sequence — the same workflow features different products as the sales change. This is what distinguishes them from event-driven automations, where a defined action follows a specific event. An event can wake a workflow early, but what the workflow then does is still decided from the store’s figures at that moment. A burst of related changes wakes a workflow once rather than once per change, a workflow’s own changes never wake it again, and the scheduled evaluation still runs, which is how conditions that no single change announces — stock that has stopped selling, a collection member that has gone stale — are noticed. A workflow observes the same approval boundary as a conversation. Reversible merchandising runs unattended: rotating a collection to current best sellers, retiring members that stopped selling, tagging what sells. Anything touching money prepares a draft and waits — promotions and clearances, gift cards for customers who pass a spending threshold in a window, markdown pricing rules over stock that stopped moving, and a marketing campaign when trade drops off against the period before — announced only on the channels the store has connected. Search merchandising works the same way on a schedule. A workflow keeps watching which searches convert badly and drafts a rule when one is worth fixing, so it follows whichever search is failing this season rather than the one that was failing when the rule was written. It proposes only when a search is both busy and converting well below your baseline, and only ever features a product that search already matches. A busy search that returns nothing at all is reported as a stocking decision instead, since a reordering rule has nothing to reorder. Where there is no cost data to choose by margin it says so rather than guessing, and it will not draft a second rule while one is still waiting for you. The money templates carry per-run limits. A promotion template drafts one promotion. A gift-card template drafts one card per qualifying customer, each of which is a liability the merchant owes, so a single run is capped on both the number of recipients and their combined face value; a rule matching more customers than the cap allows stops and reports that rather than drafting an arbitrary subset, and a cooldown keeps the same customer from being rewarded again on the following run. A markdown rule is scoped to the specific products it identified rather than the whole catalogue, is drafted switched off, and leaves only one draft outstanding at a time, so a weekly cadence cannot stack duplicates while the merchant decides. Every one of these limits is enforced in the database, not in the prompt.

Content on a schedule

Writing is a standing intent too: ask for a post or a lookbook on a recurring basis and each run produces a new one — a real piece, built around products you actually sell, with a cover image, saved as a draft under CMS. Nothing is published for you; it is your public voice, so you publish it. Give the rule either a fixed topic or a category. With a category it chooses its subject afresh every run from whatever shoppers are most drawn to there, so the piece follows real demand instead of repeating something you decided months ago. If nothing in that category drew interest that week, it writes nothing and tells you why. These run weekly at most, and a rule pauses itself once three of its drafts are waiting unread — publishing or discarding some starts it again.

Sign-up forms where they fit

A lead capture workflow prepares a sign-up form where one has a clear job: a waitlist for a collection or offer you scheduled for later, a back-in-stock list for popular products that sold out, and a trade enquiry form for a store that sells wholesale. Each form is prepared once per purpose and reused afterwards, and it never asks for a phone number. Every form waits for your approval, except that a store in automatic mode publishes the waitlist for a drop you scheduled directly. Once published, your storefront shows it.

Seasons and days

Any workflow can be scoped to part of the year and to particular weekdays — promotions, collections, tagging, gift cards, pricing and content alike, not only the writing ones. A season recurs. Set up a Black Friday rule for November or a Valentine’s rule for February once, and it wakes every year on its own; out of season it is skipped rather than switched off, so it comes back without you re-enabling it, keeping its name, settings and history instead of being deleted and rebuilt each year. A window can wrap the year end, for a mid-December to early-January run. Weekdays pin a rule to the days you mean. On its own a weekly rule is an interval, so it drifts by a day or so each time; naming Friday makes it land on Friday and stay there. They combine: Dynamic workflows are managed at Settings → Workflows, which is also where each one can be paused or removed. Alongside the configuration is the record of every run — what it did, and the figures behind the decision, so a change always has an explanation attached rather than appearing unannounced. A workflow that fails repeatedly switches itself off and says so, rather than retrying indefinitely.

Autonomous workflows

A workflow still begins as a request. Autonomous mode removes that step: the assistant watches how the store is performing and keeps the set of standing rules current on its own — setting up rules worth running, retuning ones whose thresholds have drifted from the business, and retiring ones that have stopped producing anything. It is opt-in per store, at Settings → Workflows, and it changes who maintains the rules rather than what a rule may do. The approval boundary is unchanged. Discounts, gift cards, pricing rules and search rules are still prepared as drafts and still wait for a person, under the same permission and on the same pages. What becomes automatic is the upkeep, not the spending. Full autonomous mode, a separate level above this one, is the only setting that changes the approval boundary. Three constraints hold it in place, and each is enforced in Galactic Core’s own code rather than left to the assistant’s judgement — they sit alongside the wider set of platform guardrails:
  • A rule a person paused stays paused. Pausing is an instruction, not an untidy state to be corrected, so it is never restarted automatically.
  • A rule that has ever acted is never retired. Retirement is reserved for rules that have produced nothing across a sustained run of evaluations, counted from the run log rather than asserted.
  • Nothing happens without a record. Every adjustment writes a run entry carrying the figures it was based on, beside the entries for the actions themselves — and the two are distinguishable, so “what changed in my shop” and “what changed about my rules” remain separate questions.
Running in the background consumes AI credits, which is why the same setting establishes how they are replenished: an amount, the balance that triggers it, and a ceiling for the month. A payment that does not go through is reported by email, and a second failure stops the attempts rather than re-presenting the card. Without replenishment the assistant simply stops maintaining the rules; the rules themselves carry on running as they were. Where a store has a marketing goal recorded, the assistant reads it — what the merchant said they are optimising for, and until when — so its decisions follow that objective. Where there is none, which is the ordinary case, nothing changes.

Learning from results

Standing rules learn from what their own suggestions achieved. When a promotion’s window closes, the sales it ran against are compared with the same length of time before it, and the outcome is recorded against the rule that proposed it. A rejected suggestion counts too, and a merchant may say why — the discount was too deep, too shallow, the wrong products, the wrong time — so that a reason reaches the setting it concerns rather than being averaged away. What a rule learns is a small set of named settings: how deep its discount runs, how often it suggests, the return an advertising rule treats as losing money, the hour a social network’s posts do best. Each moves only inside a range, by a bounded step at a time, and only once enough results exist; until then the standing default applies. Every value carries a sentence saying why it is what it is, and Settings → Workflows → What it has learned shows each one with its evidence, with controls to pin it, reset it to the standard value, narrow its range or switch learning off. A preference learned from rejections — “turned down three slow-stock offers as too big a discount” — is remembered for the assistant’s own answers too, and can be forgotten from the same page. Learning tunes how a rule proposes; it never changes what a rule may do. The bounds and the approval step are described under guardrails.

Product pages

One workflow arrives with every store rather than being asked for. It looks for products whose search title and description are empty and writes them, so those pages can be found by search engines. A product added or edited without them is picked up shortly afterwards, and a pass once a month works through anything still outstanding. It works from the product itself and from the terms shoppers on that store have searched for, it only ever fills a gap rather than replacing wording a merchant wrote, and it starts with the pages that have nothing at all. A catalogue larger than a month’s budget is covered over several months, emptiest pages first.

Seasons

A store’s seasons depend on where its customers actually are, so they are read from real storefront traffic rather than assumed. A store selling into the United States meets Thanksgiving and Black Friday; one selling into Britain meets Boxing Day instead; a store selling into both meets all three. Where there is too little traffic to say, nothing is proposed rather than something guessed. Each season carries the lead time an offer needs, so a rule is set up when there is still time to approve and stock it — not on the day. Seasons that suit particular catalogues are proposed only to the stores that sell those things. Dates that move from year to year, like Thanksgiving, are resolved to the exact day. What arrives is a standing rule rather than a finished promotion: it recurs each year, and what it prepares still waits for approval like any other offer.

Advertising and merchandising

The two sides read each other. A clearance rule leaves out products that advertising is driving sales to — otherwise the same merchant pays for the click and then discounts the sale it produced — and records how many it held back. In the other direction, the advertising side sees which promotions are live or scheduled, which markdown rules are active, and that changes are waiting on the merchant’s approval.

Rules that arrive with a store

A store set up by an agency can start with standing rules already running, chosen by that agency for the merchants it serves. They are ordinary workflows from the moment the store exists: the merchant pauses, retunes or removes them like any other, and what they change stays changed. Defining that set is described under agencies.

Limits

The number of standing rules a store may have, how long the run record is kept, and how many consecutive failures switch a rule off are all set per store, under Settings → Workflows → Settings. Each starts from a default appropriate to the plan and can be adjusted: a catalogue spanning dozens of categories and a focused one have different needs, and the ceiling follows the plan for any store that has not chosen its own.

Full autonomous mode

Full autonomous mode is a level above autonomous mode for a store whose owner wants its standing rules to act rather than suggest. With it on, what a workflow would otherwise prepare for approval is carried out: promotions start and end, price rules and search rules switch on, campaigns start and ad budgets move within the monthly advertising limit, gift cards are issued within their limits, posts go to the connected social accounts, and blog posts, lookbooks and forms are published. It never, on its own, changes a base price, publishes or edits a returns, shipping, privacy or terms page, sends a wholesale quote, changes a credit limit, issues a refund or payout, or changes settings, staff or roles. Only the store’s owner can turn it on, from a session signed in with two-factor authentication, after confirming each of the statements the Terms of Service list and entering a one-time code sent by email. The panel at Settings → Workflows → Advanced shows each precondition with its current state — cost prices on nearly all of the catalogue, a month of autonomous mode with decided suggestions, funded credits — and the button stays disabled until every one is met. Every action is checked against limits the owner sets before it runs: a minimum margin over cost price, a monthly discount budget, how much of the catalogue may be on offer at once, how many products one action may touch, and how often and how far a price may change. An action that would go past a limit is not carried out; it is recorded as refused, with the reason in a sentence, and the suggestion waits for a person as it would without the mode. Each action that runs can be undone for 72 hours, and an hourly check switches the store back to autonomous mode — and tells the owner by email — when margin falls, refusals pile up, the owner’s two-factor setup is removed or cost prices go missing. A daily summary and a weekly results email account for what it did. How each limit is enforced is described under guardrails.

Replies to shopper messages

A store can let the assistant answer shopper messages in its inbox. It is off by default and, once on, works in one of three ways: answering only questions the store’s own records settle — whether a product is in stock, what it costs, where an order is — suggesting a reply for the merchant to send, or also answering general questions. A reply is signed as the store’s automatic reply and quotes only what the records hold. Returns, refunds, cancellations, complaints and payment problems are always left for the merchant, as is a conversation a person has taken over, and the merchant’s new-message email is sent only for conversations that still need them. The checks are listed under guardrails.

Where the decision sits

Unless the owner has turned on Full autonomous mode, the assistant never issues a gift card, activates a promotion, or turns on a pricing rule on its own. The margin check and the approval step are enforced in Galactic Core’s own code rather than left to the assistant’s judgement, so a prompt cannot talk its way past either; they are part of a wider set of platform guardrails. Every action it takes is recorded in your activity log under its own name, so an AI-assisted change is as traceable as a manual one. What it can see is bounded the same way. The assistant reads your store through a business schema — a fixed, named layer of business concepts — rather than querying your data directly, with the store bound before any concept is read, and margin arriving as a computed figure rather than as the cost behind it. That is how it can refuse a discount for selling below cost without the cost itself ever entering the conversation. The arrangement is described in full under the business schema the assistant reads through.

Shopping assistants: the Agent API

The in-app assistant works for the merchant. The Agent API serves the other side of the counter: an AI shopping assistant helping a shopper buy from a store — the store’s own chat widget, a voice interface, a generated storefront, or a third-party assistant. Galactic Core supplies the tools rather than the assistant. The developer brings the model and the conversation design; the tools supply what a model cannot know or must not guess — live stock, the price a shopper is actually charged, the promotion that applies, shipping and tax for an address, the return window — and a checkout the shopper confirms themselves. The endpoints live under /v1/agent/* and are available in the SDK as client.agent; the Agent reference documents every method with real responses. Every response shares one envelope: the result in data, and an evidence array naming the operation and field each figure was read from, so an assistant can cite a number and a reviewer can trace it. Most tools need only a publishable key.

Finding a store

A store publishes agent.json at /v1/feeds/{store}/agent.json: the API’s base address, a production publishable key dedicated to assistants, links to the tool manifest and the OpenAPI specification, the store’s feeds, and checkout.human_confirmation_required: true. No key is needed to read it. A storefront serves the same document at /.well-known/agentic-commerce.json by forwarding that path, and the store’s llms.txt links to it, so an assistant that reads either can start calling the Agent API without being configured for the store in advance. The descriptor is published by default, except by a wholesale store, which publishes it once the merchant turns it on. The merchant controls it from Settings → Agentic commerce, and turning it off also deactivates its key. A store on a trial does not publish one. The Feeds reference shows the document.

The tools

A specification the merchant has not published is reported as absent, never inferred, and a basket line that cannot be bought is flagged rather than dropped. No tool exposes cost, margin, or what other shoppers are doing.

Two ways to pay for the model

A helper call runs entirely on the store’s side — the model, the tool calls it makes, and the pricing and stock checks behind its answer — and returns a finished result, so the calling application holds no model provider key and runs no model of its own for that step. The conversation itself, and when to call each tool, remain the application’s. Every tool other than the two helpers is deterministic. The helpers are off until the merchant turns them on, which requires automatic recharge of the credit balance with a monthly limit. Each call is debited at its true cost and reports the amount; a failed call keeps nothing. When the helpers cannot run — not enabled, paused, or out of credits — they say so, and every deterministic tool keeps working. A server holding the store’s secret key reads the balance, the recharge state and each month’s helper spend with getUsage.

Checkout intents

An assistant never places an order or takes payment. It creates a checkout intent: the basket is priced for the shipping address, the price is frozen, the stock is held for 30 minutes, and a single-use confirmation token is returned. The shopper’s approval is collected in one of two ways. By default the shopper reviews the frozen total in the developer’s interface and approves it; the application then confirms the intent with the token, alongside the shopper’s credential or, for a guest, their contact details. Alternatively the assistant asks for hosted confirmation and receives a link to the store’s own confirmation page, which it gives to the shopper. The page shows the frozen quote, takes the shopper’s contact details, confirms the intent, and passes payment to the store’s payment provider — a Stripe Checkout redirect, a Paystack popup, PayPal buttons, or a custom provider’s redirect — so an assistant without a checkout interface of its own can still complete a purchase. Confirmation prices the basket again. If anything the shopper would pay has changed — a line price, the discount, shipping, tax or the total — nothing is placed, and the response carries the new quote so the shopper can decide again. Otherwise the order is placed with payment pending, using the held stock, and payment proceeds through the store’s normal payment flow. The token is stored only as a hash, works once, and a wrong one reveals nothing about the intent. An intent left unconfirmed expires and releases its stock; one the shopper declines can be cancelled at once.

Limits on a publishable key

A checkout intent holds stock, and a publishable key can be read from any storefront, so intents created with one are bounded: 60 new intents an hour from one address, 5 waiting for confirmation at once per address (per shopper when a shopper credential is sent), 10 units of one item per intent for a guest and 25 for a signed-in shopper, and guests together may hold at most half of an item’s stock. Past a limit the intent is refused with 409 hold_limit_reached and the limit in details.limit, or 429 rate_limited with Retry-After. A secret key, used from the store’s own server, is bound only by its plan. A recognised assistant platform, identified by a Web Bot Auth signature or its published IP ranges, has its own allowance per store of 600 new intents an hour and 100 waiting; the details are on the Agent API reference.

Returns

getOrderStatus lists request_return among an order’s allowed actions while the order is inside the store’s return window, and requestReturn lodges it for the shopper — the whole order, or chosen lines and quantities, with a reason. The return arrives in the store’s returns queue as pending, exactly as one lodged from the storefront would; the store reviews it and decides the refund, and the shopper is emailed that it was received.

Standing instructions

A shopper can leave an instruction for later: reorder an item every so many days, or buy it when it is back in stock or reaches a price. Every instruction carries a maximum unit price, a quantity of at most 10 and an expiry of at most a year; a reorder runs every 7 to 365 days, at most 12 times, and the other two fire once. A shopper can keep 10 with one store. When an instruction’s condition holds, Galactic Core prepares a checkout intent at the price it found and sends the shopper a link to the store’s confirmation page, by email under the store’s own name and in the agent.mandate.triggered webhook event. The link stays open for 48 hours. The shopper confirms and pays on that page; nothing is bought without them. Instructions need the shopper’s credential and personalisation consent, and withdrawing consent pauses them.

Events and attribution

Checkout intents emit three webhook events: agent.checkout_intent.created when an assistant prepares one, agent.checkout_intent.confirmed when the shopper confirms and the order exists, and agent.checkout_intent.expired when one lapses unconfirmed. A standing instruction that fires emits agent.mandate.triggered, carrying the confirmation_url the shopper confirms at. An order placed from a confirmed intent is recorded as an agent order: its order.created event carries source_channel: "agent" (an order placed directly through POST /v1/orders carries "storefront") and agent_client_ref, which identifies the API key or the connected application that created the intent. The store also keeps a record of assistant activity for its own merchant, separate from its storefront analytics: the words of each search and how many products matched, the products an assistant opened or compared, cart drafts, and how each checkout intent ended. These records carry no shopper identity. The merchant reads them through the in-app assistant, which can answer what shopping assistants asked for, which searches found nothing in the catalogue, and which products they looked at most.