SI Integration in Retail and E-Commerce: Preparing for 2027

In retail and e-commerce, SI is transforming a far wider area than the chat window a customer sees. It affects how a product is discovered, which options it is compared with, how inventory is planned and which decisions can be taken after the sale.

The shopping journey is also diversifying. A consumer may try a product in a store, ask an SI assistant for a comparison, verify the details on the brand’s website and pay in a different interface. Across all of these touchpoints, they expect to find the same price, accurate product information and a consistent service experience.

In the Webtures approach, preparing for 2027 starts with four questions: Is your brand being discovered? Are your products understood correctly? Does your promise still hold at the moment of transaction? Does this experience generate sustainable profit?

Answering these questions requires marketing, commerce technology, operations and data teams to work together. A strong transformation becomes possible when these teams move forward with a shared measurement system and clear responsibilities.

Executive summary: what should come first when preparing for 2027?

What is SI integration in retail and e-commerce?

SI integration means prediction, recommendation, content generation or task execution capabilities working in connection with the company’s real data and business processes. A model producing an impressive answer is not the same capability as accessing accurate stock information in the sales system or creating an order securely.

The scope of integration varies with the purpose of use:

These approaches can work together. A demand model calculates the stock requirement; a generative model explains the reasoning; and an agent, if it has the authority, creates a purchase request. The data needs, failure modes and control mechanisms of each layer should be defined separately.

What does 2026 data tell us?

Understanding the direction of the market requires reading the scope of the research, the measurement period and the type of data together. A survey expectation, platform usage data and an experimental result do not measure the same thing.

Adobe’s August 2026 report is based on an analysis covering more than one trillion visits to US retail sites and more than 100 million products, together with a survey of more than 5,000 US consumers. Using the results as a global average would not be accurate.

Amazon’s May 2026 announcement also states that Rufus has been relaunched as “Alexa for Shopping”. Walmart’s fiscal year 2027 second-quarter earnings call is dated 20 August 2026; the fiscal year and the calendar year should not be confused.

The data on agentic SI expectations comes from a 2026 global retail outlook survey and is based on 330 executives. Of the respondents, 86% work at retailers with annual revenue of at least $1 billion. The results are therefore not a representative picture of small businesses.

From the Webtures perspective, the common message of these data points is that new interfaces are gaining commercial importance. A business’s own investment decision, however, should be made on the basis of its own customer behaviour, margins and controlled experiments.

How should the size of the SI market be read?

Fortune Business Insights’ 14 September 2026 update estimates the global SI in retail market at $12.40 billion for 2025, and projects $16.54 billion for 2026 and $105.88 billion for 2034.

These figures rest on one research firm’s market definition and model. They do not represent the same magnitude as total retail sales influenced by SI. The 2026 figure is also a projection published before the year has ended.

For a management team, the real question is this: which of our business decisions contains an inefficiency that is large enough, measurable and solvable? Market growth provides context for investment; on its own, it does not prove the return of the project to be implemented.

How is the shopping journey changing?

Three different models are developing side by side today.

The brand’s own assistant: The user describes their need on the brand’s website or app. The assistant works with the catalogue, order and support systems. The design of the experience and the customer relationship remain largely under the brand’s control.

Referral from SI to the store: The user researches with an external assistant, chooses the product and moves on to the seller’s checkout. The brand has to be represented accurately at the discovery stage and preserve context during the handover.

Transaction within the interface: On eligible platforms, the customer can complete the steps of finding and buying a product in the same interface. The availability of this model depends on country, seller, payment method and integration.

OpenAI’s 24 March 2026 announcement highlights a focus on product discovery and on sellers being able to use their own checkout experiences. Shopify’s agentic storefronts documentation defines ChatGPT referral separately from the direct checkout flows in some other channels. Eligibility and activation conditions should be assessed channel by channel.

This diversity does not remove the importance of the website. A strong store infrastructure retains its value for product verification, comprehensive information, brand experience and the post-purchase relationship. The new requirement is for this infrastructure to work consistently with different interfaces.

What is agentic commerce and how is it implemented?

Agentic commerce is the participation of SI agents in shopping tasks in line with the goals and permissions set by the user. Finding products, comparing options, preparing a basket or carrying out permitted transactions can all fall within this scope.

For example, a customer might say: “Find me a running shoe I can use in the rain, that stays within a set budget and can be delivered by Friday.” A successful system does not only compare product descriptions. It also checks size, stock, delivery area, total cost and returns conditions.

A workable flow includes the following steps:

• Understanding the need and the constraints, and asking questions where information is missing.
• Finding suitable products in the current catalogue.
• Verifying price, stock and delivery against the transaction system.
• Presenting the options with the reasoning behind them.
• Checking for the required customer approval or a previously granted permission.
• Executing the transaction securely and verifying the outcome.
• Keeping order, cancellation and returns information accessible to the customer.

If an agent resends the same request because of a technical error, two orders should not be created. If the price changes, it should not silently proceed at the old amount. If the payment fails, it should not talk as though the transaction had been completed. These issues relate to the transaction design of the commerce system far more than to the model’s language ability.

In a first implementation, it may be enough for the agent to find products and prepare a basket. Automatic purchasing authority requires more mature data and control structures.

What do ACP, UCP, MCP and AP2 mean?

Knowing the protocol names is useful when preparing for 2027, but distinguishing which problem each one solves matters more.

Google’s UCP announcement aims for this framework to work alongside other protocols. OpenAI’s commerce developer documentation shows the current scope of product data and integration. Google’s April 2026 AP2 announcement addresses payment scenarios based on previously given instructions when the user is not present at the moment of the transaction.

Connecting to a protocol does not automatically make product data accurate; it does not guarantee payment acceptance, security or higher sales. Technical compatibility, business policies and customer experience should be tested together. How the protocols complement one another is covered in detail in our agentic commerce protocol map.

Product data: the shared foundation of SI visibility and transactions

A person can fill the gaps in an incomplete product description by asking customer service. For a shopping agent to compare a product correctly, however, the information needs to be explicit and consistent.

Product data work should cover the following areas:

• Persistent product and variant identifiers; GTINs and manufacturer codes for eligible products.
• Category attributes such as size, material, colour, capacity, compatibility and intended use.
• Price, together with currency, tax and promotion conditions.
• Stock or orderability at variant and location level.
• Delivery area, estimated delivery time and additional charges.
• Returns, warranty, subscription and renewal conditions.
• Images that represent the actual product and descriptions that can be verified.

The product information management system, the web page, the catalogue feed and the order system should not contradict one another. Freshness requirements should also be set field by field. Material information may change rarely; stock and promotional prices may be updated far more often.

Every data field should have an owner. When an incorrect delivery promise is spotted, it should be possible to tell whether the marketing team, the warehouse system or the carrier integration is responsible. “The SI got it wrong” is no substitute for root cause analysis.

The Webtures recommendation is to start with a data quality audit in the product group that generates the most revenue or experiences the most errors. Rolling the same rules out to the entire catalogue afterwards offers a more manageable path.

How is SI visibility measured?

A brand’s appearance in an SI answer should be examined at different levels: is it mentioned, is the product recommended, is it cited as a source, is the information accurate and can the user move on to the purchase process?

Measurement requires more than a fixed list of questions. Product category, intended use, budget, country, language and customer need all produce different queries. “Best headphones” and “Lightweight headphones with long battery life for long-haul flights” are different commercial opportunities.

A sound monitoring system records the query, date, platform, country, language and response. A general visibility score is not derived from a single answer. Variability becomes visible when the measurement is repeated under the same conditions.

Recommended indicators include:

• The brand’s appearance rate across the relevant question set.
• The accuracy rate of product and brand information.
• The distribution of pages cited as sources.
• Visibility compared with competitors.
• Conversion and contribution margin from trackable SI referrals.
• The frequency of answers containing incorrect price, stock or product attribute information.

Google’s 16 September 2026 update states that SI performance insights are available through Merchant Center in some countries. This does not mean there is a complete measurement layer for every country and every platform.

Visits where referral information can be lost should be handled separately. For the measurement side of the topic, see our Dark SI Traffic guide. Attributing all untrackable traffic to SI is not a sound method.

How do personalisation and product recommendations create value?

Personalisation creates value when it helps customers meet their immediate need with less effort. Not every past purchase is a lasting preference. A product bought as a gift may not represent the user’s own interests.

A recommendation system should weigh context, product suitability, stock levels and customer preferences together. For new users facing a cold start, popularity, category information or preferences asked for explicitly can be used.

Measuring the success of recommendations by click-through rate alone can push low-margin products or excessive discounts to the fore. Add-to-basket, completed orders, post-return revenue and repeat purchases should be tracked together.

Experiments with a control group help to establish whether a recommendation genuinely creates incremental value. Crediting the recommendation system with sales from customers who would have bought anyway overstates the return on investment.

Giving customers the ability to understand how recommendations are shaped and to change their preferences is also part of the experience. Personalisation should not turn into an invisible filter that narrows the user’s options.

Visual search, virtual try-on and generative content

Visual search can help customers find similar products from a photo. Virtual try-on makes it easier to show a product in its context of use, especially in fashion, cosmetics and home decoration. How accurately these experiences represent the actual product should be measured separately.

A garment simulation that looks good on screen does not guarantee fit or the physical behaviour of the fabric. In a furniture placement experience, a measurement error can turn into delivery and returns costs. Customer expectations should therefore be managed alongside model performance.

In generative content applications, product descriptions, translations and campaign variations can be produced faster. Information such as materials, warranty, health claims, certifications or technical compatibility, however, must come from a verified source.

A good workflow uses approved product data, flags missing information, sends risky claims to an editor and keeps a record of the published version. Fake customer reviews or synthetic testimonials presented as real user experience should have no place in this system.

An SI approach to advertising and retail media

SI can speed up the production of ad copy and visuals, classify product groups and support budget decisions. Its commercial value depends on the campaign being connected to real product economics.

Advertising heavily for a product that is about to sell out, optimising only for sales revenue in a category with a high return rate or showing conflicting prices across channels all create performance problems. Advertising systems should be fed by stock, margin and operational constraints.

New SI advertising surfaces should also be assessed by country and programme. In Google’s September 2026 announcement, Business Agent for YouTube ads was introduced as a beta in the US. Presenting a product announcement as a standard service available to all advertisers is not accurate.

In measurement, conversions reported by the platform should be separated from genuine incremental sales. Where appropriate, control groups, geographic experiments or campaign pause tests can be used. The aim is to reduce cases in which different platforms each claim the same order as their own success.

Customer service and post-purchase automation

In customer service, the first goal is to resolve the customer’s problem correctly and at a sustainable cost. A conversation closing without being passed to a human agent does not by itself prove that the problem was solved.

Order status, delivery information and standard returns conditions can be automated with more limited risk. High-value refunds, account changes, exceptional warranty decisions or suspicious transactions require stronger verification and authorisation.

When handing over to a human agent, the conversation summary, the checks carried out and the steps already tried should be passed on. Customers having to give the same information again and again is the hidden cost of automation.

To judge success, first-contact resolution, repeat contacts, incorrect transactions, resolution time and customer satisfaction should be tracked together. If complaints and returns rise while support costs fall, the application should be reassessed.

Post-purchase processes matter for agentic commerce too. Cancellation, delivery tracking and returns for an order created through an agent should be presented to the customer clearly. A system that makes shopping easier should not make problem-solving harder.

Demand forecasting, inventory and the supply chain

Forecasting systems can support inventory decisions by using past sales, promotions, seasonality and relevant external variables. Past sales, however, are not always true demand. Zero sales on days when a product was out of stock do not mean that demand was zero.

When evaluating a demand model, testing should respect chronological order. Information that would only become known in the future leaking into the training data creates an appearance of success that cannot be repeated in real life.

Forecast quality should be examined by product group. Fast-moving consumer goods, slow-selling parts and short-lived fashion collections do not behave in the same way. New products may require learning from similar products and expert judgement.

Alongside forecast error, operational success should be measured by in-stock availability, holding cost, wastage, the need for markdowns and service level. Lower forecast error may not deliver the expected financial result if decision rules do not change.

Data latency between ERP, warehouse management and order management systems is especially important. An accurate SI recommendation is not enough if lead times or warehouse capacity are ignored.

How do physical stores and the digital experience come together?

Assistants that give store staff product information, computer vision systems that assess on-shelf availability and in-store pickup processes are examples that can be applied in physical retail.

These applications must fit into employees’ existing workflows. An assistant whose answers are unreliable, or whose use slows down the moment of sale, may not be adopted even if the training time is short. Pilots should be tested under real store conditions.

In projects that use image processing, the purpose should be clearly defined. Counting products on a shelf and analysing customers’ identity or emotional state involve different levels of data and risk. If the same business outcome can be achieved with less data, that option should be preferred.

For many businesses, the consistency of basic processes such as online stock visibility, in-store reservations and returns acceptance is a problem that needs solving before any new SI feature.

Pricing and promotion optimisation

SI can help to understand price elasticity, compare promotion options and forecast the effect of discounts. Limits such as margin floors, brand policy and stock targets should be built into the decision system.

A price that changes for everyone according to demand or time and a price set according to an individual user’s profile are different practices. Data use, consumer disclosure and discrimination risks in the target market should be assessed for each specific use case.

A promotion may lift sales in the short term while teaching customers to expect constant discounts. The measurement window should therefore not be limited to the campaign period. Demand in the following period, sales shifting within the category and returns behaviour should also be examined.

Rather than changing prices automatically because SI recommends it, a recommend-and-approve model can be used in the first phase. Limited automation can be introduced as reliability and commercial results are proven.

Technical architecture: which connections should be in place before the model?

A robust architecture clearly separates which information the model can access and which actions it can take. The source of a product description and the source of its price may not be the same. While the returns policy can be taken from documents, the actual status of an order should be queried from the transaction system.

The recommended architecture consists of six layers:

• Data layer: Sources of catalogue, customer, stock, order and policy data.
• Access layer: Authentication, and permissions based on role and customer context.
• Knowledge and model layer: Search, retrieval of relevant documents, and forecasting or generative models.
• Tool layer: Functions for stock queries, basket creation and permitted transactions.
• Control layer: Business rules, spending limits, approvals and error handling.
• Observability layer: Records of answer quality, tool calls, cost, latency and transaction outcomes.

RAG can be used to generate answers grounded in documents. A retrieval approach, however, does not replace live stock and payment verification. Current transaction information should be obtained by connecting to the relevant system.

Adobe’s April 2026 Commerce announcements show that an MCP approach that simplifies access to commerce data and transactions, together with conversational shopping integrations, is becoming productised. The availability, licensing and implementation scope of each feature should be checked separately.

When choosing models, there is no need to use the largest model for every task. Simple classification, complex product advice and verification of risky transactions may require different solutions. Total cost should include model usage, data preparation, integration, monitoring and human oversight.

How should trust, authority and regulation be handled?

A product description or customer message may contain instructions that try to change the system’s task. Content from external sources should therefore not be treated as a trusted command that grants transaction authority. The actions an agent can take should be limited by tool-level and server-side rules.

Exposing payment details to the model unnecessarily, issuing broad access keys and granting open-ended purchasing authority should be avoided. The scope, duration, budget and revocation method of any authority should be explicit. It should be possible to stop a faulty release and revert to the previous flow.

In global operations, country-level privacy, consumer and SI regulations should be addressed together. Article 50 of the EU AI Act sets out transparency obligations for certain SI interactions and synthetic content, depending on role and conditions.

Under amendment 2026/1744, the 2 December 2026 transition date relates to Article 50(2) for providers of relevant systems placed on the market before 2 August 2026. This date should not be generalised as “mandatory watermarking of all marketing content on the same day”.

A product recommendation system is not automatically classed as high-risk simply because it uses SI. Other purposes, such as recruitment or credit assessment, require separate evaluation. The project inventory should make this distinction visible.

Which metrics should be tracked?

The measurement system should combine technology performance with customer experience and financial outcomes.

Denominators should be clearly defined. Is the conversion rate calculated per session or per user? Are an agent’s retries counted as separate tasks? Over what time window is post-return profit measured?

Until these definitions are fixed, reports from different teams should not be compared. The apparent success of a new system may stem from a change in measurement method.

How is return on investment calculated? An example scenario

Assume that in a store receiving 500,000 eligible visits a month, a controlled test shows the conversion rate rising from 2.0% to 2.2%. The increase is 0.2 percentage points; in relative terms, it is 10%.

This calculation assumes constant volume and benefit; it does not include ramp-up time, seasonality or the cost of capital. A real evaluation should prepare low, base and high impact scenarios.

If it cannot be shown that the conversion uplift comes from the application, attributing the entire difference to the SI investment would be wrong. It is also important to make sure that support cost savings and sales impact do not count the same benefit twice.

The first 90 days: how do you get started?

Rather than a broad tool procurement programme, the first phase should focus on producing a measurable result in a single use case.

When choosing a pilot, data access, integration difficulty and the cost of errors should be weighed as much as revenue potential. An assistant that explains product features and an agent that issues automatic refunds, for example, are not the same kind of starting project.

Stop criteria should be written down at the outset. The application should be restricted when unauthorised transactions, incorrect answers above the set threshold, unacceptable latency or negative contribution margin are observed.

Small businesses can start with off-the-shelf platform features. In large businesses, the variety of data, channels and countries may require more integration work. The same technology roadmap should not be applied at every scale.

2027 readiness roadmap

The 2027 plan should start from a verified picture of the current situation and avoid critical goals that depend on future features announced by platforms.

Q4 2026: strengthening the foundation. Catalogue accuracy, price and stock consistency, current SI use and the measurement infrastructure are audited. Pilot candidates are identified. Data use, exit options and responsibilities in supplier contracts are reviewed.

Q1 2027: generating evidence. The selected applications are tested in a controlled way. Commercial impact, customer experience and the cost of errors are assessed together. An operational owner is assigned to the successful pilot.

Q2 2027: connecting channels. Suitable SI shopping surfaces and store experiences are integrated. The consistency of product data across channels is monitored. The tools and permissions available to agents are reviewed.

Q3 2027: controlled expansion of scope. New categories, countries or processes are added gradually. The post-purchase experience and the human support flow are strengthened. Changing model and platform versions are retested.

Q4 2027: portfolio decision. Profitable applications are scaled; those that do not deliver results are closed down or redesigned. The following year’s budget is tied to verified value rather than to usage volume.

In Google’s September 2026 announcement, some new UCP-related features are being rolled out gradually in the US, while expansion to Australia and Canada is planned for early 2027. This shows the importance of country-by-country preparation. Announced timelines should be checked again before implementation.

Team structure and technology selection

A successful programme should not be owned by the IT or marketing team alone. The owner of the commercial goal, the person responsible for data and technology, an operations representative and those carrying out the risk assessment should work together.

The following questions should be asked when making a purchasing decision: Can we export our data? Which components are affected if a model or platform changes? Are transaction logs accessible? Is human intervention supported? How does cost change as volume grows?

An off-the-shelf solution can provide a faster start. Custom development can create value in differentiating processes and complex integrations. A hybrid approach is also possible: the underlying model and platform can be bought off the shelf, while data, controls and the customer experience are designed specifically for the business.

Employees’ roles change as well. The product specialist strengthens the verified source of information; the support manager defines the rules for exceptions; the analyst measures commercial impact. Training should cover not only writing effective prompts but also the ability to recognise errors and intervene correctly.

What developments should we prepare for in 2027?

In the Webtures assessment, preparation plans can rest on three scenarios. These are not definitive market forecasts.

Discovery moving to more SI interfaces: Brands work harder to ensure their products are represented accurately and to provide comparable data. The reliability of their own web experience supports this process.

Transactions moving inside the interface in selected channels: In some categories, the steps of shopping become shorter. Sellers develop designs that preserve their responsibility for the customer relationship, payment, delivery and returns.

Controlled spread of operational agents: Clearly bounded tasks such as reviewing stock exceptions, preparing campaigns or classifying support requests become more automated. The need for authorisation and verification continues for critical transactions.

Category differences are decisive. Freshness and substitution come to the fore in grocery, size and returns in fashion, compatibility and warranty in electronics, and contract pricing and purchase approval in B2B commerce. Being “agent-ready” means being able to meet these different requirements.

The most durable investment is in accurate data, open interfaces, portable measurement and a controlled transaction structure. These capabilities continue to serve the business even if the popularity of a particular platform changes.

Frequently asked questions

Where should you start with SI in e-commerce?

Choose a measurable problem. Gaps in product data, recurring support questions or a stock problem in a particular category can be suitable starting points. Choosing a solution before establishing data quality and the current cost of the problem increases the risk of investing in the wrong project.

Is setting up a chatbot enough?

A chat interface on its own is not integration. Answers need to be based on reliable information, connected to current systems and handed over to human support when necessary. If it is going to carry out transactions, identity, authorisation and verification layers must also be put in place.

Do you need to rebuild your website for agentic commerce?

A full redesign is not always necessary. Catalogue data, site performance, accessibility, transaction systems and supported integrations should be assessed first. A visual refresh and technical readiness are different needs.

Does SI-driven traffic bring more sales?

Some studies have observed higher conversion, but results can vary by country, category and user intent. A business should measure its own data and investigate the reason for any difference it observes. A rate from external research should not be used directly as a budget assumption.

Can an SI agent make purchases automatically on a customer’s behalf?

It may be possible in technically supported environments. Conditions such as the product, country, payment method and customer authorisation are decisive. The scope of any previously granted authority, its spending limit and the ability to cancel it must be clear.

How much budget does an SI investment require?

There is no single standard amount. An off-the-shelf platform feature and a custom, multi-country system with transaction authority create different costs. Initial investment, ongoing usage, human oversight and integration maintenance should be calculated together.

Will human customer service disappear in 2027?

No definitive conclusion can be drawn on this. Repetitive tasks may become automated; exceptions, sensitive situations and complex customer needs may require human support. Success should be judged by how appropriately tasks are shared.

The Webtures approach: from visibility to commercial outcomes

The approach Webtures recommends in this guide is to assess SI visibility together with product data and the customer journey. A brand being discovered, being understood correctly and being reliably purchasable are parts of the same commercial process.

The first step is to establish the current situation: how are you represented on SI platforms, where does your product information conflict, what obstacles exist in the purchase journey and can your measurement system show these losses?

The second step is to set priorities. Not every business needs to set up autonomous payments straight away. For some, catalogue quality creates more value; for others, personalisation; for others still, inventory and support processes.

The third step is to move forward by measuring results. Technical implementation should be handled together with the business’s technology teams, platform providers and the necessary specialists. Responsibilities for payments and legal compliance should be clearly assigned to the relevant parties.

For a more detailed readiness framework, see the Agentic Commerce Readiness Report.

Build your 2027 roadmap on data-driven priorities. Contact Webtures to assess your current visibility, your product data and your commercial conversion opportunities.

Get in touch →