AI Sales Signal Brief | Retail, E-Commerce & CPG

Unilever is using five Microsoft Foundry agents to read, interpret, check, and enter customer orders that arrive as emails, PDFs, spreadsheets, and even photos of paper forms. Across 11 pilot markets, the company says the first stage of order processing has fallen from 15–20 minutes to about 90 seconds, while exceptions still go to people. The important part is not the speed alone. Unilever adapted the workflow to how retailers and distributors already place orders, rather than asking customers to adapt to its systems. Source

Executive snapshot

• Retail AI is shifting from conversational demos to the less glamorous work of product truth, local availability, order entry, measurement, and exception handling.

• NIQ and Google want to connect AI recommendations to what is actually available and purchased, while BCG argues that most CPG companies remain early in agentic commerce.

• Circana’s new Anthropic-backed program offers a useful operating model: expert-verified test cases, explicit exception routing, and domain operators who own quality.

• For sales teams, the strongest signal is a closed learning loop. Real calls become targeted practice, and the next live call shows whether behavior changed.

• OpenAI’s default-on ad creative enhancements create a new governance question: who owns accuracy when generated variants can run without individual review?

Top developments

Google turns Search and Gemini into an app distribution surface

Google previewed a Connected Apps Platform that lets developers build once through MCP, submit through a common console, and distribute an app in Search, Gemini, or both. The platform includes testing plus traffic, latency, and error telemetry. For sellers and brands, discoverability is beginning to include reliable actions inside answer surfaces, not only content that ranks. Google

Small multimodal models make the edge more practical

Liquid AI released the open-weight LFM3 family, covering text, vision, audio, omni, and decision models in the 2B-parameter range. The company says the models can be fine-tuned locally and run on devices. That expands the design space for store, warehouse, and field workflows where latency, privacy, connectivity, or device cost matter, although the benchmark claims are vendor-reported. Liquid AI

OpenAI gives ad creative more autonomy by default

Starting today, OpenAI says creative enhancements are enabled automatically for new and existing ChatGPT Ads campaigns unless advertisers opt out. The system may adjust copy, translate text, resize images, or optimize destination URLs, and variants can run without individual preview or approval. The operational response should be stronger source-page hygiene, change monitoring, and clear ownership for regulated or promotion-sensitive claims. OpenAI

Circana makes evaluation part of the agent platform

Circana engaged Ode with Anthropic to build governed agents for data processing, trend forecasting, and product classification. Every agent is to be measured against expert-verified cases, with uncertain work routed to specialists; Ode will also train Circana engineers and operators. No outcome data has been published yet, but the design is notable because evaluation, escalation, and internal capability are included from the start. Circana

AI for Sales and field execution implications

Turn the last call into practice for the next one

Outdoo’s new Close CRM integration uses logged calls and deal context to create AI roleplays, then scores practice and live calls on the same rubric. That is a more useful pattern than generic coaching content: identify a specific behavior gap, practice it against a realistic scenario, and check the next customer interaction for change. The company published product details, but no customer outcome data. Outdoo

Field execution depends on exception design

Unilever and Circana point to the same operating principle. Automate the repeatable path, define uncertainty, and route exceptions to people with the right context. For field sales and service, the practical evaluation set should include messy documents, incomplete records, local terminology, low-connectivity conditions, and the handoff itself.

Measure the loop, not the assistant

Useful measures include time to accepted order, correction rate, exception rate, time to resolution, coaching transfer to live calls, and customer effort. Usage and answer quality still matter, but they do not show whether the commercial workflow improved.

Friday focus: Retail, E-Commerce, and CPG

The recommendation is only as good as the shelf truth behind it

NIQ and Google announced a collaboration intended to combine Google’s Shopping Graph with NIQ product intelligence covering more than 250 million products and store intelligence across 90 countries. The plan also brings Google’s Meridian marketing-mix framework into NIQ Cadence so brands can connect recommendations and media activity to real purchase data. Specific milestones and results remain forthcoming, but the direction is clear: product identity, availability, and measurement are becoming core agent infrastructure. NIQ

BCG finds a readiness gap behind the agentic-commerce enthusiasm

BCG reports that 59% of surveyed U.S. CPG companies have no meaningful agentic AI capability. Its research says companies are concentrating near-term investment on generative-engine optimization and structured product data, while fewer are building owned agents or ecosystem integrations. BCG also reports revenue lifts among pilots, but the article does not provide enough methodological detail to treat those ranges as general benchmarks. BCG

Salesforce connects discovery to checkout while keeping the merchant in the transaction

Salesforce’s October B2C Commerce release enables UCP-powered checkout on Google surfaces while the retailer remains merchant of record. It also adds natural-language catalog filtering, automated categorization, local-inventory-aware shopping search, and an assistant that can query commerce APIs with approval required before changes. The pattern is broader than e-commerce: action becomes useful when it is grounded in current operational data and bounded by explicit authority. Salesforce

What to watch

• Unilever scale-up: Whether NOVA’s touchless rate and exception quality hold as it expands from 11 to more than 50 markets.

• Agentic-commerce measurement: Whether NIQ and Google publish concrete milestones for availability accuracy, conversion, and incremental impact.

• Creative governance: How advertisers audit automatically generated ChatGPT Ads variants, especially for price, promotion, health, or performance claims.

• Holiday operations: Whether retailers measure agent-assisted discovery against completed, fulfilled, and retained orders rather than clicks and conversations alone.

Source list

Microsoft: Unilever NOVA · Google Connected Apps · Liquid AI LFM3 · OpenAI creative enhancements · Circana and Ode with Anthropic · Outdoo and Close CRM · NIQ and Google · BCG agentic commerce analysis · Salesforce B2C Commerce release