Logiciel Solutions Contact Us
Success Stories Tech News Investors Contact Us
whitepaper
8 min read

AI in Retail and Inventory Intelligence.

AI in retail has moved past pilots in two places, inventory decisions and the shopping front end, and both depend on the same stock data. IHL Group puts the cost of out-of-stocks and overstocks at $1.73 trillion a year worldwide, 6.5% of retail sales, while fewer than one in four retailers have rolled out AI or machine learning where that cost hits hardest.

At the same time, traffic from generative AI tools to US retail sites rose 693.4% in the 2025 holiday season. This paper maps eight use cases to published evidence and goes deep on inventory intelligence and agentic commerce.

The evidence is public benchmarks, industry surveys and vendor datasets. Survey and vendor figures show correlation rather than cause, and none of it measures any single retailer's results.

In depth

The Margin Sits In Inventory Decisions, And Agents Now Read The Same Data.

01

Inventory distortion is the largest measurable cost AI in retail can address.

IHL Group puts out-of-stocks and overstocks at $1.73 trillion a year worldwide, or 6.5% of retail sales, even after $172 billion in improvements over the prior year. North America carries $415 billion of it. Supply chain disruption is the largest single cause at $301 billion, and personnel issues add $248 billion. Fewer than one in four retailers have rolled out AI or machine learning in the areas it hits hardest. Retailers that use AI and machine learning for inventory report sales growth 2.3 times and profit growth 2.5 times that of competitors, though those are survey figures that show correlation rather than cause. The size of the loss is well documented, and the share of retailers applying AI to it is still small.

In shortthe share of retailers applying AI to it is still small
02

Machine learning beats statistical forecasting on the best public retail evidence.

In the M5 competition, run on 42,840 Walmart series, the winners used LightGBM ensembles trained across many series at once and beat statistical benchmarks by up to roughly 20%. Time-series foundation models are the newer option. Amazon's Chronos reports strong zero-shot results across 42 datasets, Google's TimesFM is pretrained on about 100 billion real-world time points, and Chronos-2, released in October 2025, adds covariates such as price and promotions. Some pretrained models saw public benchmark data in training, and they are less proven on promotion-heavy retail demand. A point forecast also cannot weigh the cost of a lost sale against the cost of excess stock, which needs the full demand distribution. The forecasting evidence favors cross-learning models, and the newest models still need testing on your own holdout data.

In shortThe forecasting evidence favors cross-learning model…
03

Inventory accuracy limits every model built on top of it.

When book stock is wrong, replenishment orders against phantom units, and an agent quoting availability sells items that are not there. Item-level RFID raised inventory accuracy from 63% to 95% in a 2018 Auburn University study, and IHL expects RFID deployment to grow 291% over two years as retailers close the gap. Shrink adds to the same problem. NRF put average shrink at 1.6% of sales in fiscal 2022, or $112.1 billion, and its 2024 study found reported shoplifting incidents up 93% in 2023 versus 2019. Computer vision can help, but the FTC banned Rite Aid from facial recognition for five years after its system wrongly flagged shoppers. Clean stock data is the input that forecasting, replenishment and agent channels all share.

In shortreplenishment and agent channels all share
04

AI shopping agents are a real discovery channel, and control of checkout is still contested.

OpenAI and Stripe released the Agentic Commerce Protocol in September 2025, and Google launched the Universal Commerce Protocol in January 2026, with the Agent Payments Protocol as a payment layer. Salesforce found shoppers referred from AI search converted nine times more often than social referrals, and brands running their own agents grew holiday sales 6.2% versus 3.9%. Both are vendor datasets. In March 2026 OpenAI moved purchases into retailer apps inside ChatGPT, after trade reporting said in-chat checkout did not drive meaningful sales and some merchants did not want OpenAI running the transaction. The protocols are published, but who owns the order is not settled.

In shortwho owns the order is not settled
The detail

Three Questions That Show Whether Retail AI Will Change Outcomes.

None of this requires choosing a model first. It requires knowing whether the data and the decision path can carry one, and that is visible before any vendor is shortlisted. Three questions do it, each answered with one named artifact your teams can produce or admit they lack.

Zone · 01

Is zero sales flagged

Ask whether sales history marks the days an item was out of stock. Without that flag a forecast reads zero sales as zero demand, learns to order less of exactly the items that sold out, and repeats the stockout. Every forecasting method in the paper, from exponential smoothing to foundation models, depends on this input. The artifact is two or more years of SKU-location-day sales with stockout flags.

Zone · 02

Can an agent read it

Ask whether the catalog, live price and live availability can be read by a machine without a person in between. An agent that sells an out-of-stock item creates a cancellation and a lost customer, and incomplete attributes keep products out of agent results entirely. The artifact is a catalog audit against the Merchant Center and ACP feed specs, starting with top sellers, plus an inventory API with a stated freshness target.

Zone · 03

Which prices use personal data

Ask which pricing decisions draw on personal data and which move only with supply, demand or time of day. New York requires a prescribed notice on prices set by algorithms using personal data, with penalties up to $1,000 per violation, while market-driven pricing sits outside that law. The artifact is a written map of every pricing input, with personal-data pricing separated, logged and able to be switched off.

By the numbers

The figures that make it a board-level conversation.

$1.73 trillion
yearly global cost of inventory distortion, equal to 6.5% of retail sales, with $415 billion of it in North America
Up to about 20%
better accuracy than statistical benchmarks for the LightGBM models that won the M5 competition on 42,840 Walmart series
693.4%
rise in traffic from generative AI tools to US retail sites in the 2025 holiday season, from a modest base
Inside the report

What you'll take away.

01

Step 1 - Rank use cases by decision frequency and data readiness

Start where decisions repeat daily and history is clean, such as forecasting, replenishment and markdowns. They produce the most auditable gains and build the data foundation agent channels need later.

02

Step 2 - Backtest forecasting methods on your own data before choosing

Run a tuned gradient-boosting model against a foundation-model baseline per segment. Score both on weighted error and bias, since public benchmark results may not transfer to promotion-heavy demand.

03

Step 3 - Make inventory and catalog data readable by agents

Expose location-aware stock and price through an API with a freshness target, complete product attributes against feed specs, and decide whether to offer native protocol checkout, app checkout or discovery only.

04

Step 4 - Measure every AI decision against a control group

Track forecast error, in-stock rate, turns, markdown rate, shrink and conversion by channel against a holdout or pre-period, not a vendor's benchmark, and watch the planner override rate as a trust signal.

Questions

Frequently asked.

How is AI used in retail?

Retailers use AI in eight main areas: demand forecasting, replenishment and allocation, pricing and markdowns, inventory visibility and shrink, personalization and search, conversational commerce and shopping agents, store labor planning, and supply chain planning.

Inventory and pricing decisions show the most auditable gains. BCG reports retailers adopting AI pricing raised gross profit 5% to 10%, and McKinsey finds personalization most often drives a 10% to 15% revenue lift.

How accurate is AI demand forecasting compared with traditional methods?

On public retail data, machine learning is more accurate. In the M5 competition on 42,840 Walmart series, LightGBM ensembles trained across many series beat statistical benchmarks by up to roughly 20%. Results vary by category, and some foundation models saw public benchmark data in training.

Backtest candidate methods on your own data and measure both weighted error and bias before choosing one.

How much do out-of-stocks and overstocks cost retailers?

About $1.73 trillion a year worldwide, according to IHL Group. That inventory distortion equals 6.5% of global retail sales, and North America accounts for $415 billion of it. Supply chain disruption is the largest single cause at $301 billion, followed by personnel issues at $248 billion.

Fewer than one in four retailers have rolled out AI or machine learning in the areas where these losses are largest.

What are the Agentic Commerce Protocol and the Universal Commerce Protocol?

They are open standards that let AI agents buy from merchants on a shopper's behalf. OpenAI and Stripe released the Agentic Commerce Protocol in September 2025, covering product feeds, checkout and delegated payment.

Google launched the Universal Commerce Protocol in January 2026, covering discovery, checkout and post-purchase, and it can use Google's Agent Payments Protocol for signed payment authorization.

Do AI shopping assistants improve e-commerce conversion rates?

Early data says yes, though most of it comes from vendors. Salesforce found shoppers referred from AI search converted nine times more often than social referrals in the 2025 holidays, and brands with their own agents grew sales 6.2% versus 3.9%.

Adobe recorded a 693.4% rise in generative AI traffic to US retail sites from a modest base. Measure conversion on your own traffic with holdout groups.

What privacy laws apply to AI in retail in the US?

No single federal law covers it. Retail AI falls under the FTC Act, state comprehensive privacy laws in 23 states as of April 2026 per the IAPP, biometric laws such as Illinois BIPA, and New York's Algorithmic Pricing Disclosure Act, in effect since November 2025.

Facial recognition for loss prevention carries high risk after the FTC's 2023 order banning Rite Aid from using it for five years.

What does this whitepaper cover, and how far does its evidence reach?

It covers a use-case map, the evidence on impact, inventory intelligence, agentic commerce protocols, architecture, privacy, KPIs, a phased roadmap and a readiness scorecard, current as of November 2026. The evidence is public benchmarks, government data, industry surveys and vendor datasets.

Survey and vendor figures show correlation rather than cause, unaudited vendor case studies are left out, and no figure measures any single retailer's results.

Get the whitepaper

Have it emailed to you.

Drop your details and we'll send AI in Retail and Inventory Intelligence straight to your inbox - no spam, unsubscribe anytime.

Download whitepaper
Next step

Find out which retail AI decision your data can support first.

Bring your sales history, your inventory accuracy picture and the use case your team most wants to fund. We will work through which decisions your data can carry today and what a first pilot would need to be measured against. SECTION 7 - FAQ - 5 to 8 questions

Book a retail AI session