Reading time: about 11 minutes
Most AI spending in eCommerce starts at the wrong end: with a tool demo instead of a process. A store sees something impressive, buys a subscription, and then spends weeks looking for a job to give it. The results show up in the numbers. Nearly every company now uses AI somewhere, yet only a small minority can point to a real effect on profit.
The stores that get value do the opposite. They start with the work their team already repeats by hand every week, ask which of those tasks are safe to hand to a machine, and only then pick a tool.
There is also a quieter shift underneath the tooling question. AI assistants have started to shop on people's behalf, and they read your product data, not your design. The 2025 and 2026 experiments by OpenAI, Google and Shopify showed both how fast this is moving and where it breaks. This article covers both: a practical framework for choosing what to automate, and what agentic commerce means for your catalogue today.
Why most AI projects stall: they start with the tool
Adoption is no longer the problem. Value is. McKinsey's 2025 global survey found that most organisations now use AI in at least one business function, but only 39% report any effect on enterprise-level EBIT, and most of those attribute less than 5% of EBIT to AI. The same research points to one factor more than any other behind real returns: redesigning the workflow, not just adding a tool to it.
Other research tells the same story from different angles:
-
MIT's Project NANDA (July 2025) estimated $30–40 billion of enterprise spend on generative AI and found that most organisations in its sample saw no measurable P&L return. The authors blamed brittle workflows and poor fit with day-to-day operations, not model quality. The headline 95% figure is debated, but the direction is consistent.
-
Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear business value or weak risk controls. It also warns of "agent washing": vendors relabelling ordinary chatbots and automation as agents .
The common thread is the order of operations. "Which AI tool should we get?" produces an answer before anyone knows the problem. Ask a different question first: what does our team repeat every week that eats hours and produces something easy to check? Once you know that, choosing a tool is the easy part. Bending a process to fit a tool you already paid for is the expensive part.
The three-filter test: what makes a good automation candidate
A task is worth automating only when it passes all three filters below. Two out of three usually means either no payback or too much risk.
|
Filter |
The question to ask |
Why it matters |
Fails when |
|
Frequency |
Does it happen daily or weekly? |
Setup, prompts and checks are a fixed cost; repetition pays it back |
It happens quarterly or once a year |
|
Cost |
Does it eat real hours or money? |
Automating a 10-minute monthly task changes nothing on the P&L |
Nobody would notice if it stopped |
|
Checkability |
Can a person verify the output before a customer sees it? |
Errors get caught in review, not in a complaint |
Mistakes surface only after an order, refund or claim |
The third filter is the one teams skip, and it is the one that decides risk. Generative AI is good at producing plausible text quickly. It is not reliable at guaranteeing that a number, a date or a legal condition is correct. So the practical question is never "can AI do this?" but "who reads it before it ships, and how long does that take?"
A useful rule of thumb: if reviewing the output takes almost as long as producing it by hand, the task fails the cost filter in disguise. If nobody can review it at all, it fails the checkability filter and belongs in the human-only column.
Where AI fits, and where humans stay in charge
The dividing line is not how "advanced" a task is. It is what an error costs and who catches it.
|
Rare tasks |
Daily or weekly tasks |
|
|
High cost of error |
Keep manual. Handle case by case. |
Human decides, AI assists. Prices and promotions, stock and delivery promises, legal text. AI may draft or flag; a person approves. |
|
Low cost of error |
Not worth automating. Setup costs more than the task itself. |
Automate with review. Descriptions, FAQs, review and ticket analysis, email drafts, segmentation, catalogue checks. |
Good fits: high volume, reviewable output
-
Product descriptions for large catalogues with many similar items. AI drafts from structured attributes; a content manager edits and approves. The attributes stay the source of truth, so a wrong draft is easy to spot.
-
FAQ blocks on category and product pages, built from real customer questions and support tickets.
-
Customer segmentation from purchase history: frequency, basket size, category affinity, time since last order. The model proposes segments; marketing decides what to do with them.
-
Email and campaign drafts, especially variants for different segments.
-
Review and ticket analysis: what customers praise, what they complain about, which issues repeat. This turns thousands of free-text comments into a short list a product owner can act on.
-
Catalogue hygiene: flagging missing attributes, duplicate products, inconsistent units or titles that break category rules.
What these have in common: a person reads the output before a customer does, or the output informs a decision rather than becoming one.
Human-in-the-loop only, or no AI at all
-
Prices and promotions. A wrong price becomes a wrong order, a cancelled order or a public complaint.
-
Stock availability and delivery promises. Overselling costs refunds and trust.
-
Legal and policy text: return policies, terms, privacy notices, warranty conditions.
-
Anything a customer can hold you to. In Moffatt v. Air Canada (2024), a Canadian tribunal ordered the airline to honour a refund rule its website chatbot had invented. Air Canada argued the bot was responsible for its own statements; the tribunal rejected that and held the company liable for not taking reasonable care that its chatbot was accurate.
Customer service shows the same lesson from the cost side. Klarna said in 2024 that its AI assistant was doing the work of roughly 700 agents. By May 2025 its CEO acknowledged that cost had been too dominant a factor and quality had dropped, and the company began hiring human agents again, keeping AI for routine queries. The lesson is that a hybrid model, AI for the repeatable and humans for the consequential, outperforms either extreme.
Agentic commerce: what actually happened in year one
Agentic commerce means an AI assistant does more than recommend a product: it searches, compares and, eventually, completes the purchase for the shopper. The first year produced fast infrastructure and a sharp reality check. The durable pattern so far is discover in AI, buy on the merchant's site.
|
Date |
Event |
What it means for merchants |
|
Mar 2026 |
OpenAI deprioritises standalone Instant Checkout and refocuses on product discovery; purchases move to merchant sites and apps |
The assistant sends the shopper; your checkout closes the sale |
|
Mar 2026 |
Shopify switches on Agentic Storefronts by default for eligible US merchants: catalogues appear in ChatGPT, with Copilot and Google AI Mode available |
Distribution to AI channels becomes a platform feature, not a project |
|
Jan 2026 |
Google and Shopify announce the Universal Commerce Protocol (UCP), co-developed with Etsy, Wayfair, Target and Walmart and endorsed by 20+ companies |
A second open standard covering the whole journey, from discovery to order management |
|
Sep 29, 2025 |
OpenAI and Stripe launch Instant Checkout and the open Agentic Commerce Protocol (ACP); Etsy live, "over a million" Shopify merchants announced as coming |
First mass-market attempt at in-chat purchase |
|
Sep 2025 |
Google announces the Agent Payments Protocol (AP2) with 60+ payments partners; Visa and Mastercard build agent payment frameworks on their networks |
The payment-authorisation layer forms underneath the shopping protocols |
Why in-chat checkout stalled
Three reasons recur in the coverage. First, uptake was thin: reports put the number of live Shopify merchants at roughly a dozen to 30 against the million announced. Second, CNBC reported that product selection stayed limited and the feature did not always show up-to-date item information. Third, OpenAI itself said the first version lacked the flexibility merchants and shoppers needed. Standardising product data across millions of catalogues turned out to be part of the problem.
Discovery is real, and it converts
The retreat on checkout did not slow AI-driven shopping. Adobe's analysis of more than one trillion visits to US retail sites found that traffic from generative AI tools rose 693% year over year in the 2025 holiday season, and those visitors converted 31% better than other traffic. Shopify reported AI-referred sessions to its merchants up more than 8x and AI-referred orders up nearly 13x in Q1 2026.
Treat forecasts as ranges, not targets
McKinsey estimates agentic commerce could orchestrate up to $1 trillion of US B2C retail revenue by 2030 and $3–5 trillion globally. Narrower definitions give much smaller numbers: Morgan Stanley's base case for US agentic e-commerce in 2030 is around $190 billion. The spread comes from what each firm counts as "agentic", from purchases an agent merely influenced to ones it completed end to end.
The takeaway for a store owner is not "buy an AI feature". It is this: a machine is already reading your catalogue on the shopper's behalf, and it repeats whatever your data says. If the price, stock or return terms it reads are stale, the shopper arrives with the wrong expectation and your team handles the fallout.
Why clean product data is the real AI priority
For most stores, the highest-return "AI investment" in 2026 is not a new feature. It is getting the catalogue into a state that machines can read and trust. The same work pays off three times: in classic SEO, in AI answers inside search, and with the agents that shop on a buyer's behalf.
The platforms are explicit about this. OpenAI's product feed specification asks merchants to push structured data rather than relying on crawling, with required fields such as a stable item ID, title, description, URL, brand, image, price with currency, availability and seller details, plus flags controlling whether an item can appear in search. Google now accepts store-wide return policies (since June 2024) and shipping policies (since November 2025) as Organization-level structured data, so the rules no longer have to be repeated on every product.
|
Data layer |
What "clean" looks like |
Typical failure |
|
Attributes |
Complete, consistent attributes per category: same names, same units, real values instead of "see description" |
Half the catalogue has size in the title, half in a field, some in neither |
|
Identifiers |
Stable product and variant IDs; GTIN or MPN where they exist |
IDs change on re-import, so channels lose history and reviews |
|
Price and stock |
Synced automatically from the source system to the site and every feed |
A weekly manual export; promotions end on the site but live on in feeds |
|
Structured markup |
Product, Offer, AggregateRating on product pages; MerchantReturnPolicy and shipping policy under Organization |
Markup missing, or contradicting the visible price |
|
Policies as text |
Shipping times, costs and return terms written in plain, crawlable sentences |
Terms shown only as a banner image or inside a PDF |
|
Media |
Several images per product, descriptive alt text, variant-specific photos |
One image shared across all colours |
No tool fixes a messy catalogue. An AI writer fed inconsistent attributes produces confident, inconsistent descriptions. An assistant reading a stale feed quotes a stale price. Data quality is the precondition for every other item in this article, which is why it deserves its own budget line rather than being treated as cleanup someone does on Fridays.
The 30-minute automation audit
You can run this with your team in one meeting, with a shared spreadsheet and no vendor on the call.
-
List every task the team repeats by hand each week. Be literal: "write descriptions for new arrivals", "answer where-is-my-order emails", "update prices from the supplier file".
-
Score frequency: daily, weekly, monthly. Drop anything rarer than weekly.
-
Score cost: estimate hours per week, multiplied by the hourly cost of whoever does it.
-
Score checkability: can someone verify the output in a fraction of the time it takes to produce it? Yes, partly or no.
-
Flag the red zone: anything that sets a price, promises stock or delivery, or states a legal condition stays with a human, whatever its score.
-
Pick one or two tasks with the highest cost that pass all three filters. These are your pilots.
-
Audit the data separately: take 20 products from your top category and check attribute completeness, feed price against live price, markup validity and whether shipping and returns are readable as text.
The output is a short list of pilots and a data backlog. Both matter; the second usually matters more.
A realistic 90-day path
-
Days 1–30: data foundation. Fix attributes in the top categories, automate price and stock sync, add Product and Offer markup, publish shipping and return policies as text with Organisation-level markup.
-
Days 31–60: first pilots. Launch one or two reviewed automations, for example, description drafts for new arrivals and weekly review analysis. Measure hours saved and the edit rate on AI drafts.
-
Days 61–90: scale or stop. Keep what saved real time at an acceptable edit rate; drop what did not. Then connect the catalogue to AI shopping channels your platform supports, and track AI-referred sessions and orders in analytics.
Decide the success metric before each pilot starts: hours saved per week, edit rate, conversion on pages with new content. A pilot without a metric becomes a subscription nobody cancels.
Next step
List your team's five most repetitive weekly tasks and send them to us. We will help you spot which ones are safe, high-ROI automation candidates, and check whether your catalogue is ready to be read by AI shopping assistants.