• turns metrics and deviations into clear recommendations backed by specific figures;

• suggests next steps (Next Best Action / Next Best Offer) based on comparisons with the plan and a peer group, as well as internal rules and reference data;

• generates structured talking points for client conversations — short prompts for discussing assortment, sales, and equipment — without replacing the system of record.

Challenges

• Different data channels and varying data quality. Some information is available via REST API, while some comes through regular data extracts, and updates may be delayed. We needed a robust synchronization approach that could also handle incomplete data.

• Factual accuracy of recommendations. We could not allow discrepancies between the assistant’s recommendations and the data in the corporate system, so conclusions were strictly tied to verifiable input metrics.

• Clear explanations, not just generated text. Users needed more than an abstract suggestion: they needed explanations based on specific figures and comparisons so each recommendation could be checked quickly.

• Operational supportability. We needed transparent integration error diagnostics, automatic retries after failures, and observability so support would not turn into manual incident investigation.

Solutions

• Integration with a fallback path. We implemented the primary data exchange via REST API and a backup path based on CSV/XLSX extracts and ETL. We also added caching to reduce dependency on source availability.

• Standardized input for analysis. We introduced a unified “data packet” assembled before recommendation generation: entity metrics, transaction and order history, peer-group comparisons, reference data, and plan targets when available.

• Transparent output with user confirmation. Key source data is displayed alongside recommendations, while proposed next steps remain subject to user confirmation.

• Reliability and observability. We set up error handling and retries, centralized logs and metrics, and added alerts to monitor system stability and data quality.

What We Implemented

• End-to-end AI workflows within the product. We implemented the full path from selecting an entity to receiving recommendations and prompts, including generated explanations and numerical justification based on calculated metrics.

• RAG for enterprise context. We connected search across the internal knowledge base — procedures, rules, standards, and reference materials — so responses rely on up-to-date internal documents and remain consistent in format and terminology.

• AI orchestration and recommendation generation. We implemented a service that accepts the standardized “data packet,” applies rules and context, including information from the knowledge base, and returns structured output with insights, recommendations, and next steps.

• Integration service and data synchronization. We deployed connectors to the corporate system and configured refresh schedules and degradation rules for cases where some data is temporarily unavailable.

• Comparative analytics. We implemented calculations for key metrics — averages and latest values, peer-group comparisons, and assortment mix analysis — and structured the resulting data for further analysis and recommendation generation.

• Quality checks and diagnostics. We added integration tests and data validation checks, as well as dashboards and logs for analyzing recommendation quality and investigating incidents.