GK Software has launched GK Agentic, an AI framework aimed at enterprise retail processes that already own price, inventory, promotions and transactions rather than treating a language model as the system of record.
The offering is split across two layers that share one agent library. Retail assistants use retrieval-augmented generation against approved material such as operating procedures, compliance documents and product information. Embedded agents sit inside GK CLOUD4RETAIL workflows and are meant to support work such as promotion creation, campaign management, pricing and analytics.
That split matters more than the “agentic” label. POS still has to complete sales, inventory still needs deterministic records, and pricing still has to honor effective dates and rules. The useful AI layer is one that can act through those systems instead of replacing them as the source of truth.
GK is pitching model choice as well: different large language models, and cloud, on-premise or hybrid deployment. For a large retailer that can be about cost, data residency, latency and procurement, not only benchmarks, including the option to swap a model later without rebuilding the process.
Model Context Protocol is part of the integration story. GK describes the framework as MCP-ready, with connectors that can expose retail backends as structured tools. MCP does not clean up a legacy estate or replace permissions and business rules; it can standardize how an agent requests an allowed action instead of a custom integration for every workflow.
Employee assistants sit at the lower-risk end: a new store colleague can get an answer grounded in approved documentation rather than the open web. Embedded execution raises the bar. Creating a promotion, changing a campaign or touching pricing still needs role-based access, approvals, logging, rollback and a record of what changed. GK says human-in-the-loop governance is built in.
According to streetinsider.com, the company is presenting GK Agentic at NRF Europe in Paris and product material also links Fressnapf to the framework. Public information still does not offer a broad set of comparable production metrics on labor or cycle-time savings across retailers, so commercial evidence lags the architectural claim.
The launch still marks a shift: general-purpose AI is moving closer to systems that already hold operational authority. Retailers are likely to judge vendors less on a single model and more on whether AI runs inside existing controls, whether models can change without changing the process, and whether behavior stays predictable when stores are busy and the network or model service is not.
GK’s argument is that the AI layer should be replaceable while the retail process stays durable—a more useful promise, for long-lived store systems, than another named assistant.