A mid-sized cleanliness and facilities compliance company, operating across healthcare, education, building services, and hospitality in Canada and the US, had outgrown its own success. The platform processed more than 130K+ assessments in the last 3 years and recorded over a million individual assessment items across key customers, all while maintaining 100% service uptime. As its digital inspection and field-operations platforms scaled into more regulated environments, audit volumes and documentation demands began growing faster than its inspectors, custodians, and sales teams could keep up with by hand.
Auditors were buried in manual post-inspection work, reviewing dozens of criteria, typing summaries from scratch, stitching together report narratives, interpreting photo evidence on their own judgment, and hunting down fragmented SOPs mid-task. The platforms had already digitized inspections and centralized field data, but the intelligence layer sitting on top of that data was still entirely human.
Net Solutions partnered with the client to build a RAG layer into its digital ecosystem, one that grounds every AI-generated summary and recommendation in the company’s approved documentation rather than in a language model’s general knowledge. The result is a knowledge layer that summarizes, reads images, and answers field questions in real time, with every output traceable back to a source document and reviewed by a person before it counts as official.
Assessments completed in 2024 through 2026
Service uptime maintained
Individual assessment items recorded across key customers in 2024 through 2026
Consecutive SOC 2 Type 2 audits completed with no exceptions
The client’s platforms had already digitized inspections and centralized field data collection, but the intelligence layer sitting on top of that data was still entirely human. A handful of recurring problems made it hard to scale.
Inspectors were completing assessments with roughly 400 scored criteria, each carrying comments and photo evidence. Turning that into a coherent summary meant re-reading every item and typing a narrative by hand for every single assessment, a process that could eat up 60 to 90 minutes per audit. Back-to-back inspections wore inspectors down, and summary quality tended to slip as the day went on.
Producing one location-level report meant manually pulling together findings from several assessments completed over a date range and writing a “super summary” from scratch each time, which slowed down how quickly customers got their reports.
Two inspectors could review the same photo differently and arrive at different conclusions, because each interprets cleanliness standards differently. Since scoring relied on personal judgment, comments and scores varied by inspector, making results inconsistent and hard to defend during audits and customer review calls.
SOPs, product guides, and compliance references were spread across emails, shared drives, and binders. Inspectors and custodians routinely had to step outside the platform or call a supervisor just to get an answer, which broke their workflow and slowed decisions in the field.
Without one source of truth, new or temporary staff had no reliable way to match the judgment of experienced inspectors, so both cleaning outcomes and audit scoring varied more than they should have.
What the client needed wasn’t a generic chatbot bolted onto its platforms. It needed something that could reason over its own operational and compliance data accurately and consistently, in a way inspectors and custodians could actually trust and defend.
We treated AI as an embedded capability from the start, not a bolt-on feature. Every output needed to be explainable, sourced, and reviewable, and that requirement shaped the whole technical approach.
We inventoried the client’s SOPs, scoring rubrics, product documentation, and compliance references to understand what ground truth actually looked like across different service lines.
We built a retrieval pipeline that could ingest varied document types, including SOPs, manuals, scoring guides, and product recommendations, and serve grounded, role-specific answers directly inside the existing mobile and web workflows.
Using agile sprints, we shipped one capability at a time straight into the inspection and field-support workflows people already used, so no one had to learn a separate tool.
Every AI output, whether a summary or a recommendation, was built to be reviewed and approved by a person before it became part of an official record.
Before writing any production code, our engineering team ran structured evaluations to decide on the technology solution and stack.
We paired the platform with strict security controls, including encrypted data in transit, secure key management through AWS Secrets Manager, and firm agreements that client data is never used to train the underlying models.
This combination was chosen specifically to meet the compliance standards that matter for this client’s footprint, including HIPAA for its US operations and PIPEDA for Canada.
We implemented three RAG-grounded capabilities, each mapped directly to a specific point of friction in the inspection and field-support workflow.
Upon audit completion, the system automatically extracts all 400 criteria per assessment, scores, and inspector comments. The RAG engine retrieves matching guidance from the client’s approved SOPs and scoring standards and generates a structured summary broken into:
What was previously a fully manual, criterion-by-criterion write-up became a review-and-approve task.
Uploaded photo evidence is now evaluated against defined cleanliness criteria using the retrieval-grounded approach. If the system detects an issue such as visible residue, it retrieves the specific cleaning product and method documented for that deficiency in the knowledge base, rather than generating a generic response. Because findings are retrieved from documented standards instead of being freely generated, scoring stays consistent across inspectors and locations and is more defensible on review.
For high-volume, rapid audits, inspectors can upload up to 20 surface images in a single batch. Each image is matched against the same knowledge base individually. For every image, the system:
An embedded assistant retrieves relevant content from ingested SOPs, manuals, and policies to answer field questions in real time, supporting text, image, and voice-to-text input. A custodian can ask something like “what’s the approved procedure and product for this surface?” and get an answer on the spot, grounded in the client’s own documentation, without leaving the app. Inspectors can go further, asking something like “what’s the approved dilution ratio for biohazard removal on sealed terrazzo floors?” and get an immediate, context-aware answer pulled directly from the ingested SOPs.
From centralized knowledge to defensible scoring, how the RAG layer changed day-to-day work for inspectors, custodians, and the teams reviewing their reports.
Scattered SOPs, manuals, and emailed guidance were consolidated into one retrieval-backed source of truth.
Assessment and report summaries moved from fully manual drafting to a review-and-approve step, cutting documentation time per audit.
Because image analysis and written findings pull from the same approved documentation every time, scores are more consistent across inspectors and easier to justify on review.
Consolidated, AI-drafted location reports shortened the gap between when an assessment wraps and when the customer sees results.
Inspectors and custodians get role-specific guidance inside the platform instead of leaving it to search for a procedure or call a supervisor.
New and temporary inspectors can follow the same grounded guidance as experienced ones, narrowing the gap in judgment and execution.
The platform has maintained 100% service uptime, and the client’s business continuity and disaster recovery processes have consistently completed within expected timelines during testing.
Evaluation-led, not tool-led: We don’t default to a single model or vendor. Every engagement starts with a structured evaluation of LLMs, response APIs, and architectural tradeoffs against the client’s actual requirements, including accuracy, compliance, latency, and cost, before a single line of production code gets written.
Grounded AI, by design: In regulated, evidence-driven industries, an ungrounded AI answer isn’t just unhelpful, it’s a liability. We build retrieval-augmented systems specifically so every AI output can be traced back to an approved source document.
Embedded, not bolted on: We design AI capabilities to live inside the tools people already use, right at the workflow steps that already cause friction, not as a separate chatbot people have to remember to open.
Human review at the core: Every capability we build keeps a person in the loop for final approval, so AI speeds up decisions without taking accountability off the table.
If your organization is sitting on years of SOPs, manuals, and institutional knowledge that your team still has to dig up manually, there’s a faster way. Let’s talk about how a retrieval-augmented AI layer can turn your existing documentation into real-time, trustworthy guidance for your field and frontline teams.
RAG is an AI architecture that retrieves relevant content from an organization’s own approved documents at the moment of a query, then generates a response grounded in that retrieved text. Unlike a standard chatbot that answers from general training data, a RAG system’s answers are traceable back to a specific source document, which makes it a better fit for regulated or evidence-driven industries.
Fine-tuning bakes knowledge into a model’s weights, so any update to a procedure requires retraining. RAG instead retrieves from live documents, so updating a source file updates every future answer instantly, with no retraining cycle and none of the ongoing cost of fine-tuning.
Yes. In this engagement, uploaded photo evidence is evaluated against defined cleanliness criteria using the same retrieval-grounded approach used for text. If the system detects an issue, it retrieves the specific product and remediation method documented for that deficiency, and inspectors can batch-process up to 20 surface images at once.
Every AI output, whether a summary, a score, or a recommendation, is built to be reviewed and approved by a person before it counts as an official record. The AI accelerates the draft; a person still signs off.
Industries with heavy documentation and compliance requirements benefit most, including facilities and cleanliness compliance, healthcare, education, and any field-operations business where inspectors or technicians need instant, defensible answers grounded in approved SOPs.
Yes, when built with the right controls. This engagement paired the retrieval layer with encrypted data in transit, secure key management through AWS Secrets Manager, and contractual guarantees that client data is never used to train underlying models, meeting HIPAA requirements for US operations and PIPEDA requirements for Canada.
Timelines vary by scope, but Net Solutions delivers in iterative, embedded sprints, shipping one capability at a time directly into existing workflows rather than launching everything at once. This lets teams start using grounded AI features early instead of waiting for a single, large release.
Tell us about your documentation, your compliance requirements, and where your team loses the most time searching for answers. We'll show you what a retrieval-augmented AI layer could look like inside your existing platform.
Results matter. Discover how our customized solutions have delivered measurable business outcomes for clients across various industries. Each case study highlights the strategic impact of our work, turning challenges into opportunities for growth.
Ask Sol anything about digital products, AI, engineering, or growth, and get answers drawn from years of Net Solutions thinking and experience.