Plain Questions. Grounded Answers. Straight From the ERP.

The client is a diversified, government-linked enterprise group in the Gulf, with operating companies spanning logistics, industrial services, and trading. Group operations run on a centralized SAP ERP, with an enterprise data warehouse layered on top for consolidated reporting. For years, the path from question to answer ran through the BI team: an executive or operations lead raised a request, an analyst pulled extracts, reconciled them in spreadsheets, and sent back a deck. It worked when the group was smaller and questions moved slower.

It stopped working as the group grew. The request queue grew faster than the analyst team, standard dashboards answered last quarter's questions rather than this morning's, and leaders fell back on asking whoever sat closest to the data. Answers to the same question diverged depending on who pulled the extract and when. Then staff began pasting figures into public chatbots to speed things up, and the tools returned fluent, confident numbers with no lineage. After a fabricated figure surfaced in an operations review, finance banned the practice outright. The appetite for conversational access survived the ban. The trust problem was what remained.

Kentro was engaged to put a conversational intelligence layer over the ERP and warehouse: a system any authorized employee could question in plain language, that answered only from governed data, respected each person's existing access rights, and refused rather than guessed. The engagement covered platform deployment, semantic model design, security integration, and rollout across the group's operating companies, built on Kentro's Enterprise-GPT platform in its ERP-GPT configuration.

The hard part was never the language model. It was making sure that what the model said matched what the ERP knew, for every user, every time.

• Answers lived behind a queue.: Every non-standard question became a report request routed through the BI team. By the time the extract came back, the decision it was meant to inform had often already been made on instinct.
• Same question, different numbers.: Metric definitions lived in analysts' heads and in scattered spreadsheet logic, not in a governed layer. Revenue, margin, and utilization meant subtly different things depending on who calculated them.
• Public chatbots poisoned the well.: Early experiments with consumer AI tools produced confident figures with no source, and a fabricated number reached a management review. Any replacement had to make hallucinated figures structurally impossible, not just unlikely.
• Access control had to survive the interface.: ERP authorizations kept payroll, margins, and supplier pricing scoped to the right roles. A conversational layer that flattened those boundaries would have been a data breach with a chat window.

Kentro deployed its Enterprise-GPT platform, in the ERP-GPT configuration, inside the client's own cloud tenancy and in-region. The architecture separates language from arithmetic. The model interprets the question and composes the answer, but every figure comes from SQL executed against a governed semantic layer, and every policy claim comes from retrieved, cited passages. Nothing numeric is ever generated by the model itself.

• A governed semantic model as the only source of figures.: Kentro built a dimensional semantic layer over the warehouse, with conformed dimensions and metric definitions signed off by group finance. Quantitative questions resolve against this layer, never against raw ERP tables, so the definition argument was settled before the AI ever answered.
• Schema-constrained query-to-SQL.: Natural-language questions are translated into SQL constrained to the semantic model's schema. Generated queries are parsed, validated against permitted tables and columns, and executed read-only against a warehouse replica. The model narrates the returned result set, and if a query fails or returns nothing, the assistant says so.
• Retrieval-augmented generation for unstructured content.: Policies, procedures, and contract documents are chunked, embedded, and indexed in a vector store. Answers drawn from them cite the retrieved passages, so a user can open the source and verify the claim.
• Role-based access inherited from the ERP.: Sign-in runs through Microsoft Entra ID, and each user's ERP roles map to row-level and column-level security in the semantic layer. A plant manager and the group CFO can ask the identical question and each receives an answer scoped to their own entitlements.
• Guardrails and audit as defaults.: Grounding checks reject any numeric claim not traceable to an executed query. Low-confidence retrievals return a refusal and point to the data-owning team. Every question, generated query, and answer is logged for review by internal audit.

The results that mattered were behavioral: what teams stopped doing by hand, and what became possible in the room.

• The report queue stopped being the bottleneck.: Routine questions about stock positions, receivables, and order status are asked and answered in the flow of work. The BI team shifted its time from servicing extract requests to improving models and data quality.
• Shared definitions, group-wide.: Because every figure resolves through the same semantic layer, meetings stopped relitigating whose spreadsheet was right. The argument moved from the data to the decision.
• Leadership asks the system directly.: Management reviews open with live questions instead of pre-cut decks, and follow-ups get answered in the room rather than deferred to the next reporting pack.
• Trust in AI answers held.: Every answer carries lineage back to an executed query or a cited document, and the system refuses rather than invents. Internal audit approved the rollout, and staff gained a sanctioned alternative to public chatbots.