Budgeting for Gen AI in 2027: Insights from Walmart, Amazon, and gloozd

October is the budget planning season for 2027. According to a Deloitte survey, 82% of retail and consumer goods executives plan to increase their AI investments over the next year, but only 16.5% can measure the return.

At the same time, generative AI is changing how people shop. Instead of browsing categories, using filters, or typing exact product names, customers can simply describe what they need: “recommend a coffee for a cezve” or “I want to make borscht.”

For retailers, this creates three opportunities: grow sales, better understand demand, and build a new shopping experience. But how do you measure that value? We look at Walmart, Amazon, and two gloozd projects.

1. A higher basket value among assistant users does not prove the assistant generated extra sales. To measure the impact on sales, use a control group: some shoppers do not see the new feature for a limited time, and you compare revenue per visitor.

2. Conversations with an assistant reveal demand that sales data misses. In our project for Kulturrra, around 30% of customer requests were for products not available in the range. You would not learn this without listening to what customers ask for.

3. Some values are hard to measure precisely. Being first to spot what customers are starting to ask for – and being present where they ask, including ChatGPT and Gemini – is a strategic bet. Treat it as one, with its own limited budget.

4. Most generative AI costs usually come after the pilot. Every customer request costs money, so the more popular an assistant becomes, the higher the bill. Track the cost per session or order from day one.

5. Start with product data. Both your own assistant and external AI agents can only recommend products they can read and compare.

On September 15, at a Goldman Sachs investor conference, Walmart U.S. President and CEO David Guggina shared three figures about its Sparky assistant (transcript, Retail Dive coverage):

• customers who use Sparky have a 40% higher average order value;

• weekly engagement with Sparky grew by 60% over the quarter;

• product searches using photos convert 57% better than text searches.

Look again at the first number. It compares shoppers who use the assistant with those who do not. But who tends to try a new website feature first? Often, loyal customers with larger baskets who might have spent more anyway. So the 40% difference could mean the assistant increases spending. Or it could simply mean that people who already buy more are more likely to use it. The number alone cannot tell us which explanation is right, and Walmart has not shared its methodology.

We see the same challenge in our own VARUS.UA case. We built a search feature that understands requests such as “I want to make borscht.” The Gemini generative model interprets the request and passes a list of products to the existing search engine. Within two months of launch, the new search feature's share of all search sessions rose from 15% to 35%. The average price of a product found through it was UAH 130, compared with UAH 106 through regular search. Orders containing products found through the new search had an 18% higher average order value.

That looks like a 23% higher product price and an 18% higher order value. There are two plausible explanations. The first favors AI: with regular search, shoppers look for “beetroot,” “cabbage,” and “carrots” one by one. With “I want to make borscht,” the new search suggests the whole set, so items that might otherwise be forgotten or bought elsewhere end up in the basket. The second explanation: people searching for ingredients for a complete meal may have planned a larger purchase from the start, such as lunch for the whole family. The figures alone cannot tell us which effect matters more – just as with Walmart's 40%. They suggest customers value the new feature, but they do not show how much additional revenue it generated.

A control group can answer that question. The process is simple:

1. Randomly select 10–20% of visitors who will not see the new feature for a few weeks.

2. Compare revenue per visitor between the groups, not the share of search sessions or the price of a product found.

3. The difference between the groups is the measured effect. You can apply it to your traffic and compare it with costs.

You do not need a separate platform: visitors can be assigned to groups on the website, and revenue per visitor is already available in e-commerce analytics. The main constraint is traffic. To detect a difference of just a few percent, you need tens of thousands of visitors in each group, often more. With less traffic, the test simply needs more time.

For specialty coffee multistore Kulturrra, we built an AI barista. The assistant uses a generative model and a coffee knowledge graph to recommend products. Customers ask questions just as they would in a café: “recommend something for a cezve” or “I want sweet coffee, not something acidic.”

The assistant did not deliver remarkable sales results on its own. The real value was in the conversations. When we analyzed them, we found that around 30% of requests were for products not available in the range.

We discovered this by analyzing customers' conversations with the assistant. These conversations reveal both what people want and why: for a cezve, as a gift, or for everyday drinking. By understanding the context and intent, we can see which customer needs our range does not meet – something a sales report alone cannot show.

You can measure this value through the decisions it enables:

1. Each month, label conversations by topic and whether a matching product is available.

2. Track how the share of unmet requests changes across product categories.

3. For products added to the range because of these requests, track sales separately.

These sales help estimate the business impact: analyzing conversations reveals demand that previously went unnoticed, and expanding the product range can turn that demand into actual purchases.

The third benefit is less visible. When customers talk to you in their own words, you can spot emerging needs earlier than others. The first business to notice a new request may also be the first to meet that demand – and sometimes even set a trend.

There is another benefit. A retailer that introduces a new way to shop – one that fits how people actually live – can build loyalty and help shape new shopping habits. Neither effect will necessarily show up in average order value or in a control group after just a few weeks. But “at least we're keeping up with the trend” should never become an excuse for a project that has not paid off.

It is more honest to call this a bet: set a limited budget, a time frame, and a question you want to answer by the end of it. For example, after six months, has the share of repeat purchases grown among customers who talk to the assistant? Has your market share changed in that category?

This new interaction does not happen only on your website. Customers also ask ChatGPT or Gemini for advice, and those tools send them to stores. Walmart allows this traffic. According to Guggina, ChatGPT and Gemini usually take shoppers straight to a Walmart page. Many of these shoppers are new or returning after a break, and they show a “very high” repeat purchase rate. Guggina called open access for agents a clear positive.

Amazon takes the opposite approach. It restricts access by third-party AI agents, including those from Meta, Perplexity, OpenAI, Anthropic, and Microsoft, using technical and legal measures.

The reason is its business model. In 2025, Amazon's advertising business generated more than $68 billion, and third-party AI agents could change how customers interact with ads on its platform. For Walmart, these agents are a potential source of new shoppers. For Amazon, they are a potential risk to advertising revenue.

This leads to a simple rule. If you make your money from product sales, AI agents are likely to help you. If a significant share of your revenue comes from advertising and brand promotion on your website, first measure how much of that revenue agent traffic may bypass.

Before deciding, look at the traffic you already have. ChatGPT adds the tag utm_source=chatgpt.com to links, while traffic from Gemini and Perplexity can be identified by source. In GA4, you can see how many visitors already come from these channels, their conversion rate, and their average order value – in about an hour. For many retailers, this will be their first data point on the AI agent channel.

Walmart shared Sparky's results, but not what it costs to run. This is common: generative AI costs are easy to underestimate because a large share often comes after launch.

According to Gartner estimates reported by Campus Technology, at least 70% of the lifetime cost of a generative AI model comes from processing requests – every time a customer uses it. Gartner also says a production system can cost orders of magnitude more than a pilot, and that at least half of generative AI projects will exceed their budgets by 2028. In August, Gartner predicted that the cost of a single agentic workflow could increase more than fivefold by 2028. Tokens are getting cheaper, but agents take more steps and use far more of them than a standard chatbot.

What does this mean for your budget?

• Track unit costs: per search session, conversation, or order. Then you can see growth in both revenue and costs as the assistant becomes more popular.

• Budget for monthly operating costs separately from launch costs. After the pilot, the bill for requests depends on traffic, so plan for it just as you would plan an advertising budget.

• Bring all costs together: cloud services and generative models, your team, contractors, and data preparation. These costs are often spread across marketing, IT, and e-commerce budgets.

That last point matters more than it may seem. Cloud costs usually sit in IT, assistant subscriptions in marketing, and contractors in e-commerce. Nobody sees the full amount. When no one knows what AI really costs, it is impossible to assess its return, and next year's budget is planned in the dark. The BCG survey shows a similar picture. Among 50 retail leaders in Asia-Pacific, almost nine out of ten invest less than 1% of revenue in AI overall or do not track those costs at all. Meanwhile, more than two-thirds expect AI to improve operating profitability (EBITDA) within three years or see it as a strategic necessity. In other words, companies expect AI to improve profits, yet some do not even track how much they have invested.

Both your website assistant and external AI agents depend on the same thing: product data. An assistant cannot recommend a product properly if its listing lacks key details such as size or ingredients. ChatGPT cannot recommend a product if it cannot read its price and availability. Before starting a new generative AI project, check three things:

• What share of product listings have the key attributes filled in for their category?

• How up to date are prices and stock levels in your product feed?

• Can search engines and AI agents automatically read price and availability from product pages? Developers use standard schema.org markup for this. Shoppers do not see it, but software can read it.

The first project doesn't have to be large-scale. In our project with VARUS.UA, we did not rebuild the search system: the generative model interprets the request, while the existing engine finds the products. This approach gives you data quickly, and it is easy to stop if the results are poor.

And save conversations from day one. Without them, you lose the second benefit we discussed: understanding demand.

Before launch, agree on exactly what counts as a result and use the same definition across teams. Otherwise, after three months, marketing, e-commerce, and finance may bring three different figures. Marketing might count any order where the customer opened the chat at least once. E-commerce may count only orders where a product was added to the basket directly from an assistant response. Finance may also subtract returns. Each figure can be valid by its own definition, but they cannot be compared. And the decision to scale will once again come down to assumptions.

Adding an assistant to the side of an existing website will not change much on its own. Customers may keep browsing the catalog while the chat remains just another button on the page. Real change starts when you redesign the full shopping experience around the conversation: search, product selection, recommendations, and messages. That is when all three benefits discussed here can come together: stronger loyalty and, in turn, more sales; deeper insight into demand; and a new habit of shopping with your brand.

That is why your 2027 budget should support a step-by-step shift toward this new way of shopping. The first step can be small, like the new search in VARUS, as long as you know what wider experience it will become part of. From day one, it also needs a separate cost line, a control group, monthly conversation analysis, and a GA4 report on AI agent traffic. Without these, a year later you may still be unable to answer the key question: how much did we spend on generative AI, and what did we get back?