The model is not your problem.
The 200 PDFs nobody owns are.
Azure OpenAI consultant for industrial operations. RAG on SOPs, safety MSDS, quality manuals and incident corpora. Document intelligence for AP, quality and compliance. Production-grade with citation, confidence thresholds and human evaluation sets.
The Problem
Patterns we see in every engagement
Most Azure OpenAI pilots die because the buyer asked 'where is the value' before the technical team asked 'where is the data'. We fix the data first.
Vector search is not search.
Pure semantic search misses exact-match queries like part numbers, SKU codes and standards references. We use hybrid (vector + keyword + BM25) every time. Vendors who tell you vector search is enough have not tested it on industrial document corpora.
Answers without citation are unusable.
Operations teams will not act on an LLM answer they cannot verify. Every RAG response we ship includes citation to the source paragraph. No citation, no deploy.
The model is not the bottleneck — chunking is.
Bad chunking → bad retrieval → bad answer regardless of which model you choose. We A/B chunking strategies during prototype. Most pilots that 'don't work' are chunking failures, not model failures.
Cost compounds at scale and ChatGPT Enterprise is not the answer.
ChatGPT Enterprise does not read your industrial documents, does not write back to your ERP, and does not sit inside Power BI Copilot or Copilot Studio. For operations workflows, Azure OpenAI plus a real RAG pipeline is the right shape — not consumer chat.
What we build
What we build
Three production patterns. Each one shipped end-to-end with citation, confidence threshold and human evaluation set.
RAG on industrial documents
Replaces
The 200 SOPs / MSDS / work instructions sitting in SharePoint that nobody opens because finding the right page takes 12 minutes.
- Document ingestion via Azure Document Intelligence — text, tables, headings preserved
- Semantic chunking — pages chunked by meaning, not by token count; tables preserved
- Embedding via Azure OpenAI text-embedding-3-large
- Retrieval via hybrid search (vector + keyword) in Azure AI Search
- Generation via GPT-4o (or GPT-4 Turbo for cost) with citation to source PDF page
- Surface via Copilot Studio agent or custom Power App
Plant managers and operators get sub-30-second answers to SOP questions with cited source. Compliance teams accept the workflow.
Document intelligence — AP, quality, compliance
Replaces
The AP clerk who keys 200 supplier invoice line items into SAP every morning.
- Azure Document Intelligence custom models per document type
- Confidence scoring — anything below 90% routes to human review automatically
- Power Automate flow writes the extracted record to ERP after match-check
- Audit trail per extraction — every write traceable to source document and confidence score
AP processing time drops 60–70%. Quality COA extraction becomes automated. Compliance teams get audit trail by design.
Predictive text classification
Replaces
The customer service ticket inbox where every ticket gets manually triaged and 30% misroute.
- Few-shot or fine-tuned classifier for tickets, NCRs, work orders
- Confidence threshold — high-confidence auto-routes, low-confidence to human
- Validated against held-out set before deployment, refreshed quarterly
- Integrated into existing ticketing tool — no new interface for the team
Triage time drops. Misroute rate falls. Ticket SLA compliance rises without adding headcount.
Production RAG operations — eval, monitoring, refresh
Replaces
The pilot that worked in demo and degraded in production because nobody monitored it.
- Human evaluation set of 100–300 ground-truth Q&A pairs
- Monthly scoring against the eval set — drift surfaces immediately
- User feedback loop — every thumbs-down logged and reviewed
- Periodic re-embedding when source corpus updates significantly
RAG quality is measured, not asserted. Production agents stay accurate as the corpus evolves.
Cost engineering for production GPT workloads
Replaces
The pilot bill that was USD 200 and the production bill that became USD 18,000 — without anyone noticing.
- Per-query cost telemetry — track input tokens, output tokens, model used
- Model routing — GPT-4o for hard queries, GPT-4 Turbo for medium, GPT-3.5 for long tail
- Prompt caching where Azure OpenAI supports it
- Cost dashboard surfaces top-cost users and top-cost query patterns
Production AI cost stays modeled and controlled. Surprises do not land on the CFO's desk.
Tenant residency, identity, sensitivity
Replaces
The 'we can't put confidential data into ChatGPT' wall that blocks every regulated workload.
- Azure OpenAI deployed in your Azure tenant — prompts stay in your tenant
- Entra ID for agent access — same SSO as Power BI and Fabric
- Sensitivity labels respected — labelled content does not leak across teams
- Regional deployment (UAE, Saudi, India, Singapore) where sovereignty matters
Compliance teams sign off. Regulated workloads — pharma, finance, government-adjacent — become tractable.
How we work
From document audit to first production agent in 7–8 weeks
We start with the documents and the ground-truth Q&A set. Without those two, the build is theatre.
01
Discover — score the corpus, define the queries
Two weeks. Score the document corpus. Define the 3–5 highest-value queries the agent must answer correctly. Establish the human evaluation set — 100–300 ground-truth Q&A pairs.
02
Prototype — full RAG pipeline on 200–500 docs
Three to four weeks. Build the RAG pipeline end-to-end. A/B chunking, embedding and retrieval strategies. Evaluate against the human set. Target ≥85% acceptable answers before declaring readiness.
03
Deploy — full corpus, monitoring, train users
Four to eight weeks. Scale to full corpus. Surface in Copilot Studio or custom Power App. Train users on what the agent is good at and not. Establish monthly eval refresh.
Technology stack
AI Foundation
Retrieval
Document Ingest
Surfaces
Eval & Ops
Governance
Common questions
What buyers ask us
Why Azure OpenAI vs OpenAI direct?
Tenant data residency, identity (Entra ID), sensitivity-label awareness, and regional deployment (UAE, Saudi, India, Singapore — where it matters for sovereignty). Same model. Different governance surface.
What about open-source models — Llama, Mistral?
Valid for some workloads, primarily where cost or fine-tuning control matters. We have shipped Mistral and Llama 3 on Azure ML for specific clients. Default for industrial-document RAG is GPT-4o because accuracy floor matters and user trust margin is thin.
Can we just give ChatGPT Enterprise to our team?
For generic productivity, yes — and many of our clients do. ChatGPT Enterprise does not read your industrial documents, does not write back to your ERP, and does not sit inside Power BI Copilot or Copilot Studio. For operations workflows, Azure OpenAI plus the build above is the right shape.
How do we know the answers are correct?
Three layers: citation to source for every answer; human evaluation set of 100–300 ground-truth Q&A pairs scored monthly; user feedback loop where every thumbs-down logs and we tune. We do not promise 100% — we promise the floor and we measure to it.
How much does it cost?
Discover USD 12,000–18,000. Prototype USD 30,000–55,000. Deploy USD 60,000–180,000 depending on corpus size, query patterns and integration scope. Azure consumption USD 800–18,000/month at production scale.
Ready to move
Book a 30-minute Azure OpenAI diagnostic
30 minutes with Amit. No slides. No pitch deck. No obligation to proceed. We walk through your document corpus and the queries the agent would answer first.