IskayPet's SharePoint assistant cuts answer time from 3–5 minutes to 7 seconds

IskayPet, a pet-care company with more than 330 stores and 4,500 employees across Spain, Portugal and France, has put a retrieval-augmented generation (RAG) assistant over its SharePoint documents and says staff now find operational answers in about 7 seconds instead of 3–5 minutes. The account was published on 22 October 2026 on the AWS Machine Learning blog, co-written by AWS solutions architects and IskayPet's data team, so it is a vendor-published customer story, not independent reporting.

The job

Office staff searched hundreds of SharePoint documents by hand to answer questions about return policies, veterinary protocols, contracts and onboarding, in four languages. The post says answers were often outdated or wrong, and that every new store, country or procedure made the problem larger.

What it runs on

The assistant uses Amazon Bedrock Managed Knowledge Bases, which handles ingestion, parsing, embeddings, vector storage, hybrid (semantic plus keyword) search and reranking. A native SharePoint connector reads the existing documents, so there was no migration. The post says the team enabled the connector's setting for visual content, which makes figures inside PDF, Word and PowerPoint files searchable. A generation model then writes a cited answer and is told to say so when the answer is not in the retrieved text. The team built the frontend and the AWS setup using Kiro, AWS's agentic development tool, running Claude Haiku 4.5.

The design choice worth copying

IskayPet weighed one shared knowledge base with country and area filters against one knowledge base per assistant, and chose the second. Each assistant points at specific SharePoint sites or folders, so adding an area or country means adding an assistant rather than re-architecting. Access follows SharePoint permissions twice: employees see only assistants whose documents they may read (hidden if access cannot be confirmed), and at retrieval time the employee's identity is passed so only permitted documents inform the answer. The post notes that two employees can legitimately get different answers to the same question.

Cost and tuning

The assistant retrieves 25 passages and, after removing duplicates and fragments too short to help, injects up to 15 into the prompt. The authors call that number the main cost lever because it multiplies input tokens on every question, and say input tokens dominate cost in production. Prompt caching and batch inference are being evaluated, and Bedrock Guardrails are being added as a production control. Thumbs-up and thumbs-down feedback from employees feeds the evaluation set.

The results, and what they leave out

IskayPet reports about 7 seconds per answer against 3–5 minutes before, and around 95% less time spent searching for procedures and policies, over roughly two months. It expects about 100,000 queries a month across all areas. The post does not say how answer quality was measured beyond feedback clicks, what share of answers were wrong or refused, what the assistant costs per query, or how the 95% was calculated. Guardrails are described as still being added, so the content filtering is not yet reported as live.

What to do

• If your documents already sit in SharePoint, check whether your RAG tooling can enforce the source system's permissions at query time before you build a separate authorization layer.
• Scope one assistant per business area or country, as IskayPet did, if cross-contamination between policy sets would be costly.
• Tune the number of retrieved passages against an evaluation set, and track input tokens per question, since the post names that as the main cost driver.
• Ask for the figures a vendor story omits: accuracy rate, refusal rate and cost per query.