Blog posts

Most enterprise AI project don’t fail because of the AI, they fail because they never make it to production. We all been there. The proof of concept works, the demo impresses stakeholders, the model generates the right answers, it’s time to deploy… This is when new challenges appear, such as governance, workflow orchestration, integrations, security, compliance, user adoption, and operational readiness. We can pinpoint it that this is exactly where many AI initiatives stall.

This case study shows how we helped a UK legal-tech startup bridge that gap to go-to-market with a legal workflow automation AI in Azure. Rather than building another AI prototype, we transformed an existing proof of concept into an investor-ready, production-ready solution in just two weeks - reducing manual effort by up to 80% while accelerating the client’s path to market.

The problem with the legal AI PoC

The client had already built a legal workflow automation AI PoC in Azure, which worked, the AI performed well. But when investors and enterprise customers asked to see a complete Data Subject Access Request (DSAR) workflow, there wasn’t one. Instead, requests were handled across emails, manual reviews and disconnected systems.

There was no single process to:

• receive new requests
• classify them
• collect the required information
• generate response packages
• monitor statutory deadlines

The legal AI PoC existed, but the business workflow didn’t. That meant the product wasn’t ready for enterprise AI customers, even though the underlying AI already was.

Why did this matter?

For an early-stage legal-tech company, this wasn’t just an engineering problem, it was a commercial one. Without an end-to-end workflow they couldn’t confidently demonstrate the product to investors or enterprise buyers. Every delayed release meant slower go-to-market, delayed customer acquisition and increasing pressure on the leadership.

The solution - A legal AI with audit trail and human oversight

Rather than rebuilding the client’s platform, Peruzzi Solutions focused on what mattered most: creating the missing business workflow that transformed an AI capability into a product customers could actually use. We designed and delivered a complete DSAR process that automated request intake, AI-assisted classification, workflow management and response generation, while keeping legal professionals in control through a human-in-the-loop approach.

Keeping people at the end of the workflow was critical: we all know that AI can hallucinate, as we already discussed this in a** previous article,** therefore building in human monitoring and review helps with the decision-making. The result was a solution that could be demonstrated with confidence and scaled for real customer deployments.

Business impact of the production AI

Within just two weeks, the client had an investor-ready DSAR workflow, an AI legal inbox triage with human review, that dramatically reduced manual effort while accelerating product readiness.

• 70–80% reduction in manual effort

Routine administrative work across the DSAR process became largely automated, allowing legal teams to focus on review rather than repetitive administration. For employees, the time spent was reduced from 20-40 hours to 2-4 hours, which is a huge improvement in their workflow.

• Faster path to market

Instead of spending months extending an internal prototype, the client was able to demonstrate a complete end-to-end workflow to customers and investors.

• A legal AI that fits regulated industries

The workflow was designed with human oversight built in, ensuring AI supported legal professionals without replacing decision-making.

• Built for future growth: DSAR automation in Azure AI

The solution provided the foundation for a scalable compliance platform rather than another isolated proof of concept.

Key takeaway

Many organisations believe their AI project is “almost finished” because the model works, while in reality, the hardest part is rarely the AI. It’s everything around it: from production workflows, through governance, integration security to human oversight. That’s the difference between a promising demo and a product customers are willing to trust.

Finance reporting, resource planning, and operational tracking are complex processes that involve not only numerous data sources but also a significant amount of manual effort. Connecting systems, reading Excel spreadsheets, and understanding formulas - and sometimes even macros - can be challenging.

Our client required a modern, scalable data foundation in Microsoft Fabric to improve financial reporting, reduce manual data preparation, and establish a robust foundation for future analytics. We successfully delivered this solution, enabling a more efficient and scalable reporting environment.

The problem

The organisation relied on multiple fragmented systems for finance, fundraising, resource planning and operational tracking.​ Reporting was supported by Azure SQL, Azure Data Factory and Power BI, but data remained spread across separate sources and preparation processes.​ This made it difficult to create a consistent, trusted view of performance across the organisation.​

The organisation wanted to evaluate Microsoft Fabric as a strategic platform for future data and analytics initiatives. Microsoft Fabric is Microsoft’s all-in-one data and analytics platform that brings together data engineering, data integration, data warehousing, business intelligence, data science, and AI into a single SaaS solution.

The solution

The challenge of the project was to demonstrate how Fabric could simplify data management, improve reporting capabilities, and provide a scalable foundation for future integrations, while minimising disruption to existing operations. To achieve it, we delivered a Microsoft Fabric solution using a modern, layered data architecture.​

This solution connected multiple source systems, including the accounting system, Salesforce and other operational platforms, into a unified data foundation.​ Power BI reporting was delivered directly from Fabric, giving users access to trusted data through a single reporting layer.​ It automated data ingestion, transformation, and reporting while establishing a roadmap for future integrations and enterprise-wide Fabric adoption.

The impact

Unified platform: Created a single reporting platform across finance, fundraising and operational data.​

Efficiency: Automated key data ingestion and transformation processes.​

Trusted reporting: Improved consistency and reliability of business reporting.​​

Scalability: Established a reusable foundation for future data sources and analytics use cases.​

Governance: Implemented data governance with Microsoft Purview.

Key Takeaway

A successful data platform is about creating a trusted, scalable foundation that supports reporting today and AI initiatives tomorrow.

The electricity market processes an unimaginable number of events: we turn on the lights, dry our hair, or work on our laptop while charging our phone—these are all separate events which need to be handled by the electricity industry. Our client requested one thing: build a platform which can handle it all. Not a small request!

The problem

The electricity market operates on half-hourly settlement periods, requiring precise calculation and reconciliation of energy consumption across multiple market participants.

This creates significant technical challenges:

• Massive data volumes exceeding 1 billion events daily
• Strict accuracy requirements for financial settlement
• Real-time processing needs across the entire market
• Complex coordination between multiple systems and stakeholders

The existing landscape required a platform that could handle scale, speed, and reliability simultaneously.

The solution

We designed and built a cloud-native, microservices-based platform on Microsoft Azure to process and aggregate energy market data at national scale.

The architecture included:

• Azure Databricks → high-throughput data processing
• Azure Data Factory → orchestration of data pipelines
• Azure Event Hubs & Service Bus → real-time data ingestion and messaging
• Azure Kubernetes Service (AKS) → scalable microservices infrastructure
• Cosmos DB & Azure SQL → storage and transactional processing

How does it work?

The platform enables:

1. Continuous ingestion of high-volume energy data
2. Real-time aggregation and transformation
3. Settlement-period calculations across the market

As the platform evolved, machine learning capabilities were introduced to support predictive analytics, advanced forecasting, and automated model pipelines.

The impact

The platform delivers:

• Scalability, which handles 1B+ events daily with room for growth.
• High reliability designed for continuous, uninterrupted operation.
• Accuracy, supporting critical financial settlement processes.

Key takeaways

At this scale, the challenge isn’t just processing data, it’s designing systems that remain reliable under constant load, flexible and maintainable over time. Cloud-native architecture makes this possible, but only when designed correctly from the start.

More and more organisations realise that gathering and validating documents is a real hassle. Scanned PDFs, handwritten reports or spreadsheets in several places is making it almost impossible to keep track or update them. Our client in the construction industry faced the same issue: so we implemented an AI-powered document validation tool.

The problem

A workforce solutions provider in the UK construction industry was managing compliance and documentation for thousands of workers.

This resulted in:

• 100,000+ documents 1,000 different document types
• Files arriving in inconsistent formats, layouts, and quality

Many documents were low-quality scans, poorly structured and demanded hours of manual labor to process. As expected, this resulted in a growing operational burden of manual data extraction, time-consuming validation and high risks of human error. There was one solution: the organisation needed a scalable way to process, validate, and monitor documents efficiently without increasing headcount.

The solution

We designed and implemented an AI-powered document validation pipeline on Microsoft Azure.

The solution combined:

• Azure Document Intelligence: OCR and structured data extraction from diverse document formats
• Azure OpenAI: Normalisation and interpretation of unstructured data
• Azure Functions: Scalable processing pipeline for batch validation and reprocessing

How does it work?

1. Documents are ingested into the system in various formats (PDFs, scans, images)
2. OCR extracts raw text and structural elements
3. AI models identify and normalise key fields
4. Validation logic checks for: Missing data, Incorrect values, Expired documentation
5. The system flags issues and triggers alerts for review

The architecture was designed to handle the high document volumes and ongoing revalidation.

The impact

The results were immediate and measurable:

• 75%+ reduction in administrative workload
• Automated processing of 100,000+ documents
• Proactive alerts for expired or invalid records
• Improved compliance and auditability
• Scalable foundation for future growth

Key takeaways

Most operational bottlenecks are not caused by lack of data. They are caused by data that is unstructured and inconsistent, hence difficult to process. By combining AI with a well-designed pipeline, organisations can turn messy data into reliable, actionable information.

AI is such a pillar in our lives that we don’t even think about when to use it, because the answer is: always. Microsoft, a frontier in AI engineering and the developer of Microsoft 365 Copilot, a chatbot that can code and solve problems, ran its annual survey. The 2026 Work Trend Index survey is based on data from 20,000 people across 10 countries, along with trillions of anonymised Microsoft 365 productivity signals.

With the idea in mind that AI and agents can expand human potential at work while decreasing cognitive effort, the survey tried to answer one question: how do we use AI exactly?

How do we use AI at work?

According to the survey, employees use AI to lift the ceiling of what they can do, meanwhile leaders decide what humans and AI do, hence shifting the workplace entirely. It’s not only about marketers asking “propose 10 new slogans for me”: 49% of workers use AI to analyse, process and understand new information, make decisions and solve problems.

Only 19% answered that they use it mainly to communicate with supervisors or peers or interpret info for others. 17% already produce outputs, such as documents, while 15% use it to gather information. What used to be Google is a chatbot now. Furthermore, 66% of users said that thanks to AI they spend more time on high-value work.

What does this mean in a work setting? It might mean that the “dead internet” theory came alive: more people copy-paste the received email from their colleagues into Copilot and ask it to respond. Then that colleague does the same thing. It also means that 58% say they’re producing work they couldn’t have a year ago.

Juniors are not necessarily juniors anymore, as they have a 24/7 available professional system to help them out. The learning curve is shorter as they can access every piece of information and are capable of doing work they shouldn’t yet be capable of.

What are the four modes of working with AI?

We can establish four modes of working with AI according to the results.

Delegation:

• Turning raw data and notes into structured data
• Pulling, formatting and reporting

Collaboration:

• Refining a proposal or document through multiple rounds of feedback and prompts
• Writing a communication where the tone or framing needs to be adjusted through multiple rounds of feedback

Asking:

• Looking up facts, dates or definitions
• Reformatting a table or figure
• Rewriting sentences, checking grammar

Exploration:

• Probing what agents can do autonomously
• Trying different prompt strategies
• Testing whether Copilot can handle new workflows

These modes shows how we think about AI: a companion that supports our work and allows us to focus our cognitive energy on more important tasks.

But if this is all done by AI, what do humans do?

Agents are now used in every industry, but the pattern of adoption varies widely. Software and technology are frontiers, but banking, manufacturing, healthcare, media and even nonprofit firms have adopted some form of agent.

New studies emerge frequently showing that leaning on these tools might negatively affect creativity, problem-solving skills, attention span, memory and critical thinking specifically. We are prone to believe that everything an AI chatbot says must be true. We are convinced that because it gathers information from all around the globe, more information means more accuracy, hence it doesn’t make mistakes.

The most important question is: are companies ready to support employees in using AI and agents in the best way possible, without allowing them to lose their creativity and problem-solving skills? Or: can we learn how to use AI fairly but not over rely on it?

The opportunity is there for every leader and organisation: building a place where agents increase what people can do, where human judgment stays at the center of the work that matters, and where we all have the agency to decide what comes next.

Sources: https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization#wti2026-modular-nav-shell

https://www.bbc.com/future/article/20260505-how-to-use-ai-without-turning-your-brain-to-mush

https://news.harvard.edu/gazette/story/2025/11/is-ai-dulling-our-minds/

As healthcare organisations grow, operational complexity often increases faster than their internal systems can handle. This was the case for a senior care facility in that needed a more scalable and efficient way to manage daily tasks for its social workers. Microsoft Power Platform is an effective tool, but has its limits.

This is where we stepped in: we partnered with the client to redesign and optimise their existing Power Platform-based system.

The problem

The client operates a long-term elderly care facility, providing essential daily support to senior citizens. Their job is key in society so they needed a platform which is based on Microsoft Power Platform and easy enough to maintain, but capable of managing task assignments and client-specific to-do lists for social workers.

• Task management processes became difficult to oversee and maintain
• The system relied on hardcoded elements, limiting flexibility
• Reporting capabilities were limited and fragmented
• Scaling the solution to support future growth became a challenge

In addition, the client had a strategic goal: to transform their internal system into a product that could be offered to other care facilities. This required a more structured, modular, and easily deployable solution.

The solution

We approached the project with a focus on simplification, scalability, and long-term usability.

First, we conducted a thorough review of the existing system to identify bottlenecks, redundancies, and areas for improvement. Based on these insights, we redesigned the architecture to support a more modular and maintainable structure.

The optimised solution included:

• Power Platform enhancements to streamline task and checklist management
• Azure SQL integration for structured, reliable data storage and improved reporting
• Azure Logic Apps to automate recurring and manual processes
• Template-based architecture to allow easy deployment across different environments
• Removal of hardcoded dependencies, making the system more flexible and easier to maintain

By reorganising the system into a scalable framework, we ensured that it could be customised and extended without increasing complexity.

The impact

The improvements delivered measurable results across operations, adoption, and cost efficiency:

• 40% reduction in task assignment time: Social workers can now manage their daily responsibilities more efficiently, reducing administrative overhead.
• 25% reduction in administrative costs: Automation and improved tracking reduced the need for manual oversight and coordination.

Beyond these metrics, the most important outcome was qualitative: Social workers gained more time to focus on patient care instead of administrative tasks - and this is priceless.

Key Takeaways

This project highlights a common pattern in digital transformation initiatives:

• Systems built for immediate needs often struggle to scale over time
• Complexity increases when flexibility and structure are not designed upfront
• Automation alone is not enough - architecture and usability are equally critical

By focusing on simplification, modularity, and real user needs, organisations can unlock significantly more value from their existing technology stack.