An AI assistant for workplace audit readiness

Jump to

Context

The Goal

Workflow Mapping

The Solution

Role

Experience Designer

Client

Multi-National Food Manufacteurer

Services

UX Strategy, Enterprise Workflow Design, Product Design

Role

Experience Designer

Client

Multi-National Food Manufacteurer

Services

UX Strategy, UX Research, Onsite Workshop Facilitation, ServiceNow

We designed an AI-powered audit readiness workspace to help safety and quality teams find the right documents faster, review incident evidence, and prepare more confidently for audits and inspections.

We designed an AI-powered audit readiness workspace to help safety and quality teams find the right documents faster, review incident evidence, and prepare more confidently for audits and inspections.

Contract workflows at this large financial organisation had become messy and hard to track. We evaluated how teams worked and redesigned the core experience to make contract management simpler and more centralised.

Contract workflows at this large financial organisation had become messy and hard to track. We evaluated how teams worked and redesigned the core experience to make contract management simpler and more centralised.

Jump to

The Story

AI Model

AI Interaction Patterns

System Workflow

The Solution

Reflections

Sol2

Jump to

Context

The Goal

Workflow Mapping

The Solution

The Story

For Safety managers, preparing for an audit or inspection is usually stressful and time-consuming. They have to go searching for documents, verify incident logs, and piece together what happened if things weren't documented perfectly from the beginning.

By leveraging generative AI, we wanted to build a product that could instantly parse through their dense documentation and logged data, pulling up the exact information and specific details needed to be audit-ready on demand.

A familiar interface backed by connected records

A familiar interface backed by connected records

A familiar interface backed by connected records

A familiar interface backed by connected records

AI Model

Using HybridRAG technology for unstructured, multimodal data

The source material would come in different formats, so the proposed model used HybridRAG to combine semantic search, keyword matching, OCR, and relationship mapping.

The system could find the right evidence, understand how it connected, and return information that users could review before using it in an audit response.

Connection model showing how related safety information can be linked to give AI more context.

What is HybridRAG

HybridRAG combines different retrieval methods, like semantic search, keyword matching, OCR, metadata, and relationship mapping, so AI can find evidence that is both relevant and traceable before generating a response.

AI Model

Using HybridRAG technology for unstructured, multimodal data

The source material would come in different formats, so the proposed model used HybridRAG to combine semantic search, keyword matching, OCR, and relationship mapping.

The system could find the right evidence, understand how it connected, and return information that users could review before using it in an audit response.

Connection model showing how related safety information can be linked to give AI more context.

What is HybridRAG

HybridRAG combines different retrieval methods, like semantic search, keyword matching, OCR, metadata, and relationship mapping, so AI can find evidence that is both relevant and traceable before generating a response.

AI Interaction Patterns

Designing with AI patterns users already understand

Initial CTA

We used a Direct Input box as the initial CTA, and the user could keep the prompt simple because the AI already had access to the relevant documents and evidence.

Inline follow-up

Users could highlight part of the response and ask a follow-up question directly from that selection, keeping the next prompt tied to the evidence they were reviewing.

References (via RAG)

Related documents stayed visible beside the response, so users could see which records supported the AI output and open source material when they needed to verify it.

System Workflow

From request to review-ready response

Once the user submits a query, the AI assistant searches across the available records and connects the available evidence. Using the HybridRAG model, the system can combine semantic search to understand meaning with syntactic search to capture exact terms such as names, dates, equipment, and document types.

Connection model showing how related safety information can be linked to give AI more context.

System Workflow

From request to review-ready response

Once the user submits a query, the AI assistant searches across the available records and connects the available evidence. Using the HybridRAG model, the system can combine semantic search to understand meaning with syntactic search to capture exact terms such as names, dates, equipment, and document types.

Connection model showing how related safety information can be linked to give AI more context.

System Workflow

From request to review-ready response

System Workflow

From request to review-ready response

1

Direct input CTA

The landing screen uses a familiar direct input pattern, letting users type what they need for the audit.

2

Left case sidebar

The case history stays visible in the sidebar, so users can access previous audit work and specific sections of the current case.

Initial interaction

A familiar starting point lowers the barrier to entry

The safety manager enters the information they are trying to gather, and the system begins finding the records and evidence that could support it. The direct input CTA lowers the barrier to entry, and shows off the product’s flexibility and capabilities.

Step 1

Tell the AI agent what you need

The safety manager enters the information they are trying to gather, and the system begins finding the records and evidence that could support it.

1

Direct input CTA

The landing screen uses a familiar direct input pattern, letting users type what they need for the audit.

2

Left case sidebar

The case history stays visible in the sidebar, so users can access previous audit work and specific sections of the current case.

Step 1

Tell the AI agent what you need

The safety manager enters the information they are trying to gather, and the system begins finding the records and evidence that could support it.

1

Direct input CTA

The landing screen uses a familiar direct input pattern, letting users type what they need for the audit.

2

Left case sidebar

The case history stays visible in the sidebar, so users can access previous audit work and specific sections of the current case.

Step 2

Capturing case details

The intake form captures the core case details so the system has enough context to retrieve relevant details and build a case.

Step 2

Capturing case details

The intake form captures the core case details so the system has enough context to retrieve relevant details and build a case.

Step 3

Building the case from available records

Once the case details are submitted, the system begins parsing available records and collecting the documents connected to the incident.

Step 3

Building the case from available records

Once the case details are submitted, the system begins parsing available records and collecting the documents connected to the incident.

The Solution

The final experience gave safety managers a single place to access the documents and information needed for an audit. They could share what they were looking for, add any relevant incident details, and let the system pull together the supporting records and evidence.

The Solution

The final experience gave safety managers a single place to access the documents and information needed for an audit. They could share what they were looking for, add any relevant incident details, and let the system pull together the supporting records and evidence.

The Solution

We designed an AI-powered audit readiness workspace to help safety and quality teams find the right documents faster, review incident evidence, and prepare more confidently for audits and inspections.

The Solution

The final experience gave safety managers a single place to access the documents and information needed for an audit. They could share what they were looking for, add any relevant incident details, and let the system pull together the supporting records and evidence.

Step 4

Collected evidence summary

The AI parses through all the available records and presents the case details, collected documents, and an initial summary in one place, with all the sources available in the side panel.

Step 4

Collected evidence summary

The AI parses through all the available records and presents the case details, collected documents, and an initial summary in one place, with all the sources available in the side panel.

Step 4

Collected evidence summary

The AI parses through all the available records and presents the case details, collected documents, and an initial summary in one place, with all the sources available in the side panel.

Step 5

Asking follow-up questions

Any area can be investigated further by asking a more specific question within the same workspace, with follow-up questions building on what has already been found, rather than sending the user back to a new search.

Step 6

Digging deeper into specific evidence

Once a follow-up question is submitted, the AI re-parses the relevant documents and provides a more focused summary, with the available evidence presented upfront.

Reflection

Turning the AI model into a usable review flow

This project reinforced in me that AI is most useful when it has a clear role within the workflow. In this case, it acted as a task-specific assistant, helping safety managers gather, connect, and summarise evidence without taking over the final judgement.

That distinction mattered because audit work depends on trust. The AI could reduce the manual effort of searching through records, but the safety manager still needed to understand where the information came from, whether it was relevant, and what was appropriate to share. Designing it as support, rather than as the decision-maker, made the experience feel more useful and more accountable.

Future Considerations

Source confidence

Show why a record was retrieved and how strongly it supports the answer.

State missing evidence

Make it clear when the system cannot find enough information.