It might be an email thread buried in a departed employee’s inbox. A contract saved to a shared drive with no retention tag. A customer record sitting in a system no one has audited in years.

The challenge isn’t that these documents exist. It’s that no one knows what they are, how long they should be kept, or when they should be destroyed.

Consider the stakes. GDPR enforcement actions continued their upward trajectory. And beyond the fines, organizations faced litigation holds that froze operations, discovery processes that consumed thousands of staff hours, and reputational damage that eroded customer trust.

At the root of many of these failures sits a deceptively simple problem: organizations cannot comply with retention requirements for documents they haven’t properly classified.

This isn’t a technology problem in the traditional sense. It’s a business risk problem masquerading as an administrative task. And for decision-makers navigating an increasingly complex regulatory landscape, understanding why document classification matters, and how artificial intelligence is transforming the solution, has become essential knowledge.

In this post we will look at how Knowledge Agent, currently in public preview, help mitigate this risk by deploying AI to classify documents and using Microsoft Purview’s auto apply policies that assigns a retention label based on the classification.

The Foundation: Why Classification Drives Everything

Before exploring how artificial intelligence transforms document management, it’s worth pausing on a fundamental question: why does classification matter so much in the first place?

The answer lies in understanding what retention policies require.

Every regulatory framework that governs document retention operates on the same basic principle: different types of information carry different obligations.

· A routine internal memo might require retention for one year.

· An employment record might require seven.

· A contract might require preservation for the duration of a relationship plus a statutory period afterward.

· Medical records, financial statements, intellectual property documentation, customer data, each category carries specific requirements that vary by jurisdiction, industry, and context.

Retention policies translate these requirements into actionable rules. They specify how long each document type must be kept, under what conditions it must be stored, who may access it, and when it may be destroyed.

Well-designed retention policies protect organizations by ensuring that required records remain available while unnecessary data doesn’t accumulate into liability.

But here’s the critical point: retention policies can be applied manually by a user or can be automated only if the documents have been classified.

An unclassified document exists in a compliance vacuum. No retention rule governs it because

· No one has determined which rule should apply.

· It cannot be defensibly destroyed because no one can confirm its retention period has expired.

· It cannot be reliably produced in discovery because no one can locate it through policy-based searches.

· It sits in storage, accumulating cost and risk, contributing nothing to compliance objectives.

Classification is the bridge between regulatory requirements and operational reality. It transforms abstract policy into concrete action. Without accurate, consistent, comprehensive classification, retention policies become theoretical documents that describe what should happen rather than operational controls that ensure it does.

· Organizations with strong classification practices can demonstrate defensible retention, respond efficiently to regulatory inquiries, and manage data as a strategic asset.

· Organizations with weak classification practices face the opposite: indefensible practices, scrambled responses, and data that functions primarily as accumulated liability.

The question, then, is not whether classification matters. The question is how to achieve it on a scale.

The Role of AI in Document Classification

Just as Artificial intelligence is transforming every aspect of business, so it is transforming classification of document. What required months of implementation now deploys in weeks.

Microsoft Purview offers several automated classification methods:

1. Classification based on Sensitive Information Types

2. Trainable Classifiers

3. Exact Data Match Classification

4. Metadata-based classification

In this post we will explore metadata-based classification approach (#4).

This approach leverages a document metadata to apply a retention policy. For instance, applying a 7 years retention policy to documents that have a metadata Contract as document type.

Using Knowledge Agent in M365 Copilot For Document Classification

The new Knowledge Agent, part of Microsoft 365 Copilot, helps automatically tag documents based on the content of the document. Because the agent uses LLM, we have ability to influence the application of the tag using prompting.

In this scenario, we will use Knowledge Agent to automatically classify various marketing related documents for Zava Corporation.

1. After you have enabled Knowledge Agent in your M365 tenant, navigate to the document library, and click on the icon on the bottom right of the screen. Next click on the Organize this library menu.

1. Classify documents

2. Review and update classification definitions

3. Documents are not classified. New documents uploaded will automatically get classified based on the definition. We can include a human in the loop workflow, to ensure that the AI assigned classification is validated and updated if necessary.

Now we have a document library equipped with an agent that automatically classifies documents as they are uploaded and created.

This classification not only helps with compliance, as we will explore more in the post, but also with other aspects of content management, such as findability and process automation.

Configuring managed property in SharePoint admin center

In the previous section, we created a new column named Document Type in the document library. This column contains the classification for the document. In this section, we will create a managed property, which is needed to automatically apply the retention policy.

Detailed steps for configuring a managed property is described in this article.

We will configure the managed property named RefinableString00, mapped to the crawled property ows_Document_Type

Auto application of retention policy based on documentation classification

In this section, we will explore how the auto-apply label policy in Microsoft Purview can be leveraged to apply a retention label to documents that are classified as Marketing Material

We will assume that we have a record retention label named Marketing Collateral Retention. This label has a 7 years retention policy. In the steps below we will configure the metadata-based auto apply policy to apply Marketing Collateral Retention label to documents classified as Marketing Material

1. Navigate to Microsoft Purview admin center -> Policies -> Label Policies

2. Create a new Auto-Apply policy

After you provide name and description for this policy, click Next. In this page, select Apply label to content that contains specific words, or phrases or properties

1. Next, we will specify the conditions in the policy that scans documents that are tagged with Marketing Material

2. Next, we will select the policy to be applied to content in SharePoint sites.

3. Finally, we will select the retention label that needs to be applied to the documents and activate the policy

For step-by-step detailed guidance, refer to Microsoft documentation.

Brining it all together

Task of classifying documents manually, which is time consuming and error prone is now automated by AI with Knowledge Agent in M365 Copilot. As the policy runs in the background, any new and existing documents will automatically be assigned to a retention label based on the document classification.

Aligning AI Classification with Business Imperatives

Technology capabilities matter only insofar as they address real business needs. For AI-powered document classification, the alignment between capability and requirement is remarkably direct.

The speed is imperative. Business moves faster than manual classification can follow. By the time a human review team works through a backlog, new documents have accumulated that dwarf their progress. AI classification operates at the speed of document creation.

The scale requirement. Enterprise data volumes have exceeded any realistic manual management capacity. Organizations that once measured document holdings in millions now count in billions. AI classification scales horizontally without proportional cost increases.

The consistency mandate. Regulators and litigators don’t accept “we classified most documents correctly most of the time” as a defense. Compliance requires demonstrable, consistent application of retention policies across the entire document population. AI delivers this consistency inherently.

The resource reallocation opportunity. Perhaps most importantly for business leaders, AI classification frees human expertise for work that requires human judgment. Compliance professionals who once spent their days manually tagging documents can instead focus on policy development, risk assessment, regulatory interpretation, and strategic planning.

The business case isn’t subtle. Organizations that adopt AI classification achieve better compliance outcomes at lower cost while redeploying skilled staff to higher-value activities. Organizations that don’t continue fighting a battle that mathematics guarantees they will lose.

Things To Consider

The following are few points to keep in mind as you think about implementation:

1. The Knowledge Agent is in preview at the time of this writing. Product feature may change by the time it is GA.

As demonstrated in the post, the agent today works on an individual document library. I understand that enterprises may have a need to apply the classification across multiple SharePoint sites and libraries.

If this analysis resonated with your compliance challenges, subscribe for more insights on how technology is reshaping enterprise risk and regulatory strategy. Share with colleagues who are navigating these same challenges.

These are the resources that helped me ramp up on Generative AI in 2025, and that I’ll keep coming back to in 2026.

One of the most common questions I get is: “What courses, books, or resources would you recommend to learn AI?” To kick off 2026, I’m sharing the books and online training that helped Generative AI really click for me.

In this post, I’ll focus on books and online courses. I’ll follow up with a separate post on blogs and podcasts.

My go-to AI resources

The list is in no particular order:

1. “Generative AI in Action” – Amit Bahree (Book)

   • The first book I picked up to understand Generative AI end to end.

   • Written by my colleague Amit Bahree, Principal Group Product Manager in the Azure AI engineering team at Microsoft.

   • Starts from the basics and goes deep quickly, with real-world scenarios that make the concepts concrete.

   • If you enjoy this book, I also recommend following Amit’s blog, where he writes about building LLMs and practical GenAI patterns.

2. “Co-Intelligence: Living and Working with AI” – Ethan Mollick (Book)

   • A book for everyone, not just technologists.

   • Ethan Mollick, a Wharton professor, focuses on how to partner with AI rather than just treat it as a tool.

   • Great for understanding how to think about AI in day-to-day work and life, and an excellent gift for friends and family who are AI-curious.

3. Udemy courses by Ed Donner (Online training)

   • Perfect if you learn best by watching and building.

   • Ed blends business and technology beautifully and currently has three AI courses on Udemy:

     • LLM Engineering, RAG, QLoRA, Agents

     • Complete Agent & MCP

     • Gen AI & Agentic AI for Business Leaders

Ed starts from the basics, builds to advanced concepts, and walks through real projects so you finish with something working, not just theory.

4. “AI for the Rest of Us” – Sairam Sundaresan (Book)

   • A very approachable introduction to AI concepts.

   • Explains how AI works and how to get the most out of it without drowning you in math or jargon.

   • The best part is how accessible it makes some of the more technical ideas.

5. “Reshuffle: Who Wins When AI Restacks the Knowledge Economy” – Sangeet Paul Choudary (Book)

   • One word: wow.

   • This is a business strategy masterclass on how AI is restructuring industries, value chains, and power dynamics.linkedin+2​

   • Read this not to get ready-made answers, but to sharpen the questions you ask about your business, operating model, and career.

Final thoughts & a question for you

I’m grateful to these authors and instructors for creating such foundational knowledge. Many of these resources focus on principles and mental models, so they’re unlikely to become obsolete any time soon.

If you’ve read any of these books or taken these courses, what did you think?
And what are you reading or watching right now to level up on AI in 2026?

Across any organization, teams run countless initiatives; some succeed, some fail, but every initiative creates stories. These stories explain why a strategic deal was won, why a product experiment took off, or why a promising idea stalled, and they hold the patterns needed to scale success and learn from failure.

Sales teams, for example, accumulate hard-won insight about what differentiates a winning pursuit: the programs that resonated, how the deal was negotiated, which stakeholders mattered, and what objections really changed the outcome. Engineering teams see similar patterns in experiments that drive a step-change in product adoption, revealing not just what worked, but how it was discovered, tested, and shipped.

Every department, team, and individual carries this kind of tacit, experience-based knowledge. The organizations that learn to systematically identify, extract, socialize, and replicate these stories can dramatically accelerate innovation and execution.

This is where AI now plays a transformational role. AI agents can reason across the messy, unstructured exhaust of everyday work; emails, chats, files, and meeting transcripts; to surface the hidden narratives inside an enterprise.

With Microsoft 365 Copilot’s Researcher agent, for example, knowledge workers can ask questions that span their work graph and have AI synthesize the relevant context into clear, reusable insights, turning tacit knowledge into explicit assets the entire organization can learn from and scale

Understanding Tacit v/s Explicit knowledge

Tacit knowledge is the lived, experience-based know‑how people develop over time; insights, intuitions, mental models, and skills that are hard to fully write down or formalize. Explicit knowledge, by contrast, is the codified, structured content that can be easily documented, stored, and shared in artifacts like manuals, wikis, reports, and databases. Both forms are essential, but most organizations vastly under-invest in systematically converting high‑value tacit knowledge into explicit, reusable assets.​

Traditional knowledge management systems were designed primarily for explicit knowledge: they index documents, capture process descriptions, and centralize reference material. These systems struggle with the nuance and context that make human expertise powerful things like judgment in ambiguous situations, pattern recognition across edge cases, and the “feel” for how to navigate complex stakeholders. As a result, the richest insights often remain trapped in inboxes, chats, and people’s heads, even when the organization appears to have a mature KM platform.​

This gap between what systems capture and what experts know creates a structural barrier to agility, innovation, and scale. When tacit knowledge is not surfaced and shared, teams repeatedly solve the same problems, repeat avoidable mistakes, and depend on a few “indispensable” individuals, slowing down decision-making and execution. Organizations that learn to systematically externalize tacit knowledge, making it discoverable, searchable, and recompilable are far better positioned to adapt quickly, innovate continuously, and scale their best practices across markets and functions

Scenario: Driving sales excellence using the Researcher agent in M365 Copilot

A real-life scenario you could include is how Contoso leveraged generative AI, specifically the Researcher agent in Microsoft 365 Copilot, to boost sales excellence by turning tacit knowledge into actionable insights for its sales teams.

Contoso’s own sales enablement and operations teams use generative AI to distill scattered, experiential insights from sales interactions, emails, meeting transcripts, and CRM systems. Rather than simply summarizing documents, Researcher agent synthesizes this tacit knowledge, producing stories that not only describe what helped win an opportunity, but how it was won. Insights such as executive sponsorships, sales plays, working across organizations boundaries, partnerships, technical depth, roadmaps and deal architectures.

Previously, collecting and piecing together this context would take hours; now, it’s available in moments.

Subscribe now

Enter Researcher agent in M365 Copilot

The Researcher agent in Microsoft 365 Copilot is an advanced reasoning assistant designed for complex, multi‑step research tasks rather than quick Q&A. It plans a research workflow, iteratively gathers information from both your work graph (emails, files, meetings, chats) and the web, then synthesizes the results into a structured, source‑cited report you can review, refine, and share.

The recent enhancement to the Researcher agent in Microsoft 365 Copilot allows it to deeply reason over your day‑to‑day work signals, specifically emails, chats, files, and meeting notes/transcripts, rather than just web or static documents. It can now pull together context across Outlook, Teams, OneDrive and SharePoint and synthesize it into a single, structured report that surfaces decisions, risks, follow‑ups, customer context, and open questions that would otherwise stay buried in fragmented conversations.

This new capability is essentially a tacit‑to‑explicit knowledge engine sitting on top of your real work. By reasoning across emails, chats, meeting transcripts, and files, it can surface the “story behind the work” that normally lives only in people’s heads and scattered threads.

In practice, a seller or product lead can ask Researcher to explain how a big deal was won, why a feature direction changed, or what patterns show up across a series of customer conversations, and it will assemble a coherent narrative, with sources, from all those interactions. That narrative turns ad‑hoc judgments, trade‑offs, and negotiations; classic tacit knowledge into explicit assets like playbooks, win‑loss analyses, and decision rationales that can be shared, searched, and reused by the broader organization

The following is a prompt you can use today in the Researcher agent. Please make sure to select Emails, Meetings and Chats as sources for the agent.

Prompt:

Please draft how we expanded usage of M365 Archive, Autofill and OCR within Contoso. The goal is to create a story that can be shared within the organization to help others learn from the process. Story should focus on technical implementation details as well as business outcomes. The narrative should be qualitative in nature. The story should highlight lessons learned and best practices for others to replicate. Audience includes mixed group across organization. Keep the length short within 5 pages

Amplifying the tribal knowledge into strategic differentiator

The output from the Researcher agent can be packaged and amplified across various channels and programs to reach relevant audiences.

For instance, the insights from the sales deal can be packaged as a win or a loss story for entire sales, marketing and engineering groups. These stories in turn influence the sales playbooks, marketing campaigns and product roadmap.

Leveraging AI to scale human potential

AI creates value not by replacing human thinking, but by amplifying it. Instead of treating AI as a glorified summarizer, leading organizations are using it as a reasoning layer on top of their people’s expertise, capturing the “why” behind decisions, not just the “what” in the documents.

Subscribe now

Leadership Is a Mindset, Not a Title

Leadership is not confined to roles or ranks, it’s a mindset. We emphasized that true leadership begins with intent. Whether you’re mentoring a peer, guiding a team, or influencing a community, your ability to listen deeply, act with empathy, and remain curious defines your impact.

It reminded us that leadership is about being the change maker, not just managing tasks. It starts with asking “why,” listening before reacting, and embracing the journey of learning rather than the illusion of knowing it all.

Ancient Wisdom: Leadership Lessons from Lord Ganesha

We begin our journey with Hindu mythology, which is a rich tapestry of stories, each designed to teach enduring values. Specifically, we explored Lord Ganesha, the beloved deity known as the “Remover of Obstacles,” who embodies several symbolic traits every leader can learn from:

· Large Head – Wisdom and Learning: Adopt a “learn-it-all” mentality instead of “know-it-all.” Continuous learning and broad vision are foundational for high-impact leadership.

· Big Ears – Listening: Listen more than you speak. True understanding and responsiveness come from genuinely hearing team members and stakeholders.

· Small Eyes – Focus: Mindful focus—knowing when to zoom in and out—ensures goals aren’t lost amid everyday distractions.

· Trunk– Adaptability: Be flexible. Just as a trunk adjusts to its needs, leaders must adapt their style to each situation.

· Large Belly – Embrace Experiences: Digest both the successes and the setbacks, turning challenges into learning opportunities.

· Lotus – Purity of Intention: Lead with integrity and clear purpose. Actions should be rooted in pure intentions.

· Four Arms – Mind, Intellect, Ego, and Consciousness: Balance these four aspects for holistic leadership; self-awareness underpins meaningful influence.

· Modak (Sweet) – Recognition: Regularly recognize and reward effort—small gestures of appreciation motivate teams.

· Broken Tusk – Embrace Imperfection: Turn flaws into strengths. Imperfection is an opportunity for growth and innovation

Modern Wit: Leadership Lessons from Ted Lasso

Next, we connected the heartwarming humor of Ted Lasso with timeless leadership principles. The show, centered around an American football coach managing a struggling English soccer team despite having no experience in the sport, offers a treasure trove of leadership insights that resonate deeply in today’s workplace.

1. Everyone Is Worthy of Your Attention

Ted Lasso’s first interaction with Nate, the kit man, sets the tone. Instead of walking past him, Ted stops to greet him, signaling that every role matters. Arbindo emphasized that leaders communicate their values through what they choose to pay attention to. Inclusion begins with acknowledgment.

2. Learn People’s Names

Ted doesn’t just hear Nate’s name, he remembers it and uses it. This simple act of recognition builds trust and shows respect. It highlighted that remembering names is a deliberate act of care, reinforcing identity and belonging.

3. Invest in Relationships

Ted asks Rebecca, the team owner, about her favorite concert, not to gain leverage, but to connect. It pointed out that non-transactional relationships foster deeper trust. Leaders who take time to understand their team’s backstories build influence organically.

4. Leverage Rituals to Build Camaraderie

Ted’s daily biscuit ritual with Rebecca becomes a bonding moment. I shared how our own team has a 30-minute daily ritual to share stories and connect. Rituals create cultural glue, fostering psychological safety and team cohesion.

5. Family First—Always

When Ted’s family arrives mid-practice, he prioritizes them without hesitation. We linked this to Microsoft’s leadership principle: Model, Coach, Care. Leaders who visibly prioritize family empower their teams to do the same, reinforcing a culture of work-life balance.

6. Strong Teams Learn and Win Together

Ted benches a star player for unsportsmanlike behavior, even at the risk of losing the game. It emphasized that team culture trumps individual brilliance. Humility and shared growth are the engines of sustainable success.

7. Make Others Look and Feel Good

Ted invites Nate to a gala, making him feel special. When Nate arrives early and feels awkward, Ted eases the tension by asking for his advice. It noted that leaders should see potential in others, not just performance. This builds future leaders and inclusive teams.

8. Be Open—Good Ideas Can Come from Anywhere

Ted encourages Nate to share a strategy idea, despite his junior role. Arbindo connected this to Lord Ganesha’s big ears, symbolizing the importance of listening. Leaders must create a level playing field where ideas are valued regardless of hierarchy.

9. Play Like a Child

Ted joins kids in a playground challenge and lets a waiter choose his dinner. It emphasized that playfulness fuels creativity. Leaders who embrace curiosity and experimentation foster innovation and joy. “Leadership without play is merely management,” he said.

10. Don’t Take Yourself Too Seriously, Be Vulnerable

Ted admits to not knowing about relegation during a press conference. His honesty disarms a hostile journalist. Arbindo highlighted that vulnerability builds authenticity. Leaders who say “I don’t know” open the door to learning and connection.

Subscribe now

Be curious, not judgmental

We ended the session with a short video from Ted Lasso. If you have made this far, I highly encourage you to invest 4 minutes of your time in watching this video. It wraps up the session with one simple lesson - Be curious, not judgmental!

Click on this link for the video - Ted Lasso - Be Curious, Not Judgmental

Click on this link for the video - Ted Lasso - Be Curious, Not Judgmental

These lessons from Ted Lasso are not just entertaining, they are transformative. This session showed how storytelling can illuminate leadership principles that are both practical and profound. By embracing empathy, curiosity, and humility, leaders can build cultures where people thrive.

Practical Strategies for Leaders

The following were the key takeaways:

· Pause Before Responding: Strategic thinking starts with reflection.

· Use the “Five Whys”: Dig deeper to uncover root causes.

· Create Psychological Safety: Intentional rituals and permission to fail build trust.

· Balance Speed and Ethics: Innovation must align with values.

· Model, Coach, and Care: Microsoft’s leadership principle in action.

Closing Reflections

Leadership is not about knowing it all, it’s about learning it all. Whether through ancient stories or modern sitcoms, the principles remain the same: listen deeply, lead with purpose, and never stop growing.

This session reminded us that leadership is a journey of curiosity, empathy, and courage. By bridging ancient wisdom with modern wit, we can cultivate practices that resonate across cultures, generations, and industries.

Let’s be the leaders who inspire change, build trust, and make others feel seen.

Subscribe now

A provocative claim often heard: “LLMs have made metadata obsolete.”

Metadata has never been more crucial, especially for architects tasked with building robust, scalable systems where deterministic output matters.

Deterministic vs. Probabilistic AI: Why Output Consistency Is a Business Need

LLMs, by their nature, are probabilistic. They thrive on ambiguity, context, and creativity, offering a spectrum of answers to open-ended questions. This flexibility is powerful in scenarios such as customer engagement, content generation, or exploratory analysis.

Yet, business processes demand more than creative variance, they require predictability, repeatability, and traceability. From contract management to regulatory compliance, deterministic output—where identical inputs always yield identical outputs—is essential for:

• Auditable decision-making

• Reliable automation

• Clear accountability

• Regulatory compliance

Architects building agentic AI systems must reconcile the tension between the creative flexibility of LLMs and the stringent requirements of enterprise-grade applications.

How Metadata Enriches and Controls LLMs

Metadata acts as a scaffold that introduces structure to unstructured data and workflow processes. Extracting metadata from unstructured sources—contracts, emails, policies—lets you tag, filter, and control workflows, turning agent responses from guesswork into actionable, traceable steps.

Scenario 1: Contract Management Automation

Imagine a legal team that relies on software agents to review thousands of contracts. If an agent only “reads” contracts for legal queries, its output will vary every time. However, extracting explicit metadata—contract amount, expiry date, counterparty, renewal terms—transforms contracts into structured datasets.

• Automated workflows can trigger reminders when expiry dates approach.

• Analytics dashboards can aggregate spend data by region.

• Compliance audits gain a transparent trail of terms and conditions.

Hybrid extraction approaches—combining AI automation with human review—provide scalable, accurate results in high-stakes environments. The net result: deterministic outputs from a probabilistic system.

Subscribe now

Scenario 2: Enterprise Content Organization with Knowledge Agents

Microsoft’s Knowledge Agent demonstrates how AI can enrich, organize, and structure content at scale using metadata.

• Documents are automatically tagged with relevant metadata for improved searchability.

• Agents proactively flag and repair outdated links, stale policies, or missing tags, ensuring organizational knowledge remains discoverable and compliant.

• With libraries grounded in rich metadata, Copilot (and similar agents) can deliver business-impactful insights, not just conversational answers.

Architects can deploy agents that use metadata to drive workflows, automate labeling, and enable advanced analytics—all triggering actions based on metadata rules.

Scenario 3: Enabling Reliable Analytics in Data-Driven Workflows

Consider financial operations or supply chain analytics. Without metadata, LLM-powered agents risk inconsistent aggregation, losing track of data integrity. By enforcing metadata standards, ownership, lineage, business and technical context, architects enable enterprise-wide confidence and actionable reporting.

Architecting Metadata-Enabled Agentic Systems: Best Practices

• Automate Metadata Collection and Enrichment: Rely on AI-powered tools to scan, extract, and tag metadata continuously across content sources, reducing manual effort and errors.

• Enable Data Lineage Tracking: Maintain traceability for compliance and troubleshooting, especially as workflows span multiple systems and clouds.

  Share

The Future: Hybrid Architectures

The evolving landscape points toward hybrid agentic architectures—systems combining deterministic logic and probabilistic reasoning.

LLMs interpret, summarize, and infer from complex content; metadata extraction and structuring enforce predictable, auditable, and actionable outputs.

The result: adaptable AI systems that meet enterprise needs for flexibility and control.

Key Takeaways for Architects

• Metadata is foundational for deterministic, auditable, and business-aligned agentic AI systems.

• Hybrid extraction and enrichment strategies harness the strengths of both humans and machines, delivering accuracy at scale.

• Knowledge agents and metadata-driven workflows empower robust automation, analytics, and compliance.

Metadata is the silent driver of reliable agentic AI.

Watch this short 3 minutes video to witness metadata driving deterministic response

Multi-modal large language models (LLMs) can now interpret and reason over both text and image, helping us gather insights from sources like engineering drawings, construction plans, financial charts, and architectural designs.

In this example, I’ll use M365 Copilot (powered by GPT‑5) to analyze engineering drawings and extract insights that are not explicitly documented, such as:

• The finished product design

• Challenges in manufacturing

• Estimates of volume and weight

This is significant productivity enhancer because these capabilities reduce reliance on manual expertise, accelerate decision-making, and create new opportunities for innovation.

Instead of spending weeks analyzing designs or reconciling data, leaders can get near‑instant insights that improve speed, quality, and efficiency.

This is just one scenario, but it points to a broader shift: AI is becoming not only a tool for productivity, but a strategic partner in decision-making across industries.

PDF with sample drawings and prompts are shared below.

Engineering drawing - Sample from AutoCAD

Prompts

Identifying the finished product

Prompt:

What finish product does this shape represent?

Response:

The image in the file "CAD Sample ENG drawing.pdf" shows a technical drawing of a mechanical part, commonly used in engineering and manufacturing. This specific shape is a 3D CAD drawing, often referred to as a "bracket" or "mounting block."

It features:

• A rectangular base with a cut-out section.

• A cylindrical boss on top with a central hole (Ø17 mm) and an outer diameter of 35 mm.

• Various dimensions for precise manufacturing.

This type of part is typically used as a mounting or support component in machinery, fixtures, or assemblies. The central hole in the cylindrical boss suggests it could be used for mounting a shaft, pin, or fastener. The cut-out sections and overall shape are designed to fit into a larger assembly, providing structural support or alignment for other components.

In summary, the shape in "CAD Sample ENG drawing.pdf" represents a generic mechanical bracket or mounting

Highlighting potential manufacturing challenges

Prompt:

Highlighting potential manufacturing challenges for this finished product

Response:

Based on the technical drawing in "CAD Sample ENG drawing.pdf," here are some potential manufacturing challenges for this part:

1. Tight Tolerances on Small Features: The part has several small features, such as the 13 mm wide slots and steps. Maintaining precise dimensions and sharp internal corners in these areas can be challenging, especially if using traditional machining methods.

2. Deep Hole Drilling: The central cylindrical boss has a Ø17 mm through-hole that goes through a 25 mm height. Drilling deep holes with a consistent diameter and minimal deviation can be difficult, particularly if the material is hard.

3. Material Removal and Stability: The part has significant material removed from the base (slots and steps), which can reduce rigidity during machining and lead to vibration or deflection, affecting dimensional accuracy.

4. Surface Finish Requirements: If the part requires a specific surface finish (not specified in the drawing), achieving this on both the large flat surfaces and the cylindrical boss may require additional finishing operations.

5. Fixturing and Workholding: The complex geometry, with multiple steps and a raised boss, may complicate fixturing during machining. Securely holding the part without distorting it or interfering with tool paths will require careful planning.

Addressing these challenges may involve selecting appropriate manufacturing processes (e.g., CNC milling, EDM), using high-quality tooling, and possibly designing custom fixtures. If high precision is required, additional inspection and quality control steps will also be necessary 1.

Estimating volume

Prompt:

What is the volume of this shape?

Response:

To find the volume of the shape in the file "CAD Sample ENG drawing.pdf", we need to break it down into simpler geometric components and sum (or subtract) their volumes.

Step 1: Main Block

• Dimensions: 102 mm (length) × 60 mm (width) × 29 mm (height)

• Volume: 102 × 60 × 29 = 177,480 mm³

Step 2: Subtract the Two Cutouts

Each cutout (on the left and right) is:

• Dimensions: 13 mm (length) × 13 mm (width) × 16 mm (height)

  • Height of cutout = 29 mm (total height) - 13 mm (remaining height) = 16 mm

• Volume of one cutout: 13 × 13 × 16 = 2,704 mm³

• Total cutout volume: 2 × 2,704 = 5,408 mm³

Step 3: Add the Cylinder

• Diameter: 35 mm → Radius: 17.5 mm

• Height: 25 mm

• Volume: π × (17.5)² × 25 ≈ 3.1416 × 306.25 × 25 ≈ 24,034 mm³

Step 4: Subtract the Hole in the Cylinder

• Diameter: 17 mm → Radius: 8.5 mm

• Height: 25 mm

• Volume: π × (8.5)² × 25 ≈ 3.1416 × 72.25 × 25 ≈ 5,678 mm³

Step 5: Total Volume Calculation

Combine all the above:

• Main block: 177,480 mm³

• Subtract cutouts: -5,408 mm³

• Add cylinder: +24,034 mm³

• Subtract hole: -5,678 mm³

Total Volume = 177,480 - 5,408 + 24,034 - 5,678 = 190,428 mm³

Final Answer

The volume of the shape in "CAD Sample ENG drawing.pdf" is 190,428 mm³ (rounded to the nearest mm³) 1.

Conclusion

This reasoning using multi modal LLMs, such as M365 Copilot doesn’t stop at engineering. The same approach can be applied to financial charts, construction plans, architectural designs, and other technical documents.

If these documents are stored in SharePoint online, then any user with M365 Copilot access will be able to leverage these capabilities.

Share

While some argue that it does, this raises questions about whether we are advancing towards genuinely agentic AI solutions capable of transforming firms into frontier firms.

It appears that we proceeded to discuss building, optimizing, managing, and securing agents without first establishing a clear understanding of what an agent is within the context of artificial intelligence. Consulting colleagues or experts in AI regarding the definition of an AI agent may reveal that there is no universally accepted interpretation. Even prominent technology firms in the field of artificial intelligence lack consensus on the definition of an AI agent or on what does not qualify as one.

An AI agent offers substantial potential for enhancing productivity. However, without a shared understanding among stakeholders, we risk missing out on the full advantages that this transformative technology can deliver.

This article seeks to provide a comprehensive overview of AI agents by examining their definitions, analyzing key distinctions, outlining shared characteristics, and highlighting several prominent examples within the industry.

The objective of this article isn’t just to arrive at a single unified definition, but to ensure that leaders remain grounded in the promises that AI Agent brings, of becoming the frontier firm.

Definition of an Agent

Let us begin by defining an agent, a general-purpose agent.

As per Investopedia, an agent is a person or entity authorized to act on behalf of another individual or organization, often in business, legal, or other formal contexts. The key concept is representation—a principal entrusts the agent to perform tasks, make decisions, or negotiate transactions, either broadly or with specific limitations Source

According to this definition, an agent undertakes decision-making and other responsibilities on behalf of another individual or organization. It is important to highlight these essential components for further discussion.

Subscribe now

Definition of AI Agent

Next, let us review how leading technology and analysts define AI agents.

Microsoft

OpenAI

Anthropic

Gartner

Salesforce

Share

The following is my analysis of these definitions:

· Microsoft distinguishes between agents and assistants. This distinction is relevant because Microsoft provides M365 Copilot, which serves as an assistant, along with tools and platforms for developing, operating, and managing agents.

· OpenAI emphasizes that agents act autonomously and adapt their behavior.

· Anthropic is similar to OpenAI, but it emphasizes independent task execution. According to their blog post, Anthropic distinguishes between workflow and agent: workflows follow a predefined code path, while agents determine their own processes dynamically.

· Gartner notes that an agent may function in either digital or physical environments. For example, a self-driving car is considered an agent operating in a physical environment.

· Salesforce highlights that an agent could escalate, presumably to a human, for security and ethics.

A common theme across the definitions is that an agent operates autonomously or independently. Autonomy and independence are often considered key characteristics of an AI agent, distinguishing agents from other types of software such as workflows and microservices.

Key characteristics - Autonomous

This discussion will focus on autonomy. Autonomy means possessing the ability or right to make one's own decisions, act independently, and govern oneself without external control or interference. Autonomous entities determine their actions and direction freely, based on internal reasoning and goals.

In the context of AI Agents, autonomy refers to an agent’s ability to make decisions and take actions without direct human intervention or constant oversight.

For instance, in a workflow where expense reports are routed to a human approver. An AI agent could autonomously make approval decisions or supply relevant context about the expense along with a recommendation for approval or rejection to the human reviewer.

Autonomous execution brings another factor for consideration when compared to traditional workflows – probabilistic vs deterministic output

Subscribe now

Probabilistic vs Deterministic output

AI Agents are considered probabilistic because their outputs can vary for the same input, relying on models, statistical inference, and randomness to make decisions—unlike traditional deterministic workflows, which always produce the same result from the same input based on fixed rules.

This behavior of probabilistic output is like how a human would analyze certain inputs and decide. For instance, if multiple scientists are reviewing a research paper, each one would have a unique perspective based on their individual experiences and expertise.

Deterministic approaches excel in structured, critical contexts needing consistency and explainability, while probabilistic agents thrive in situations demanding adaptability, learning, and nuance amid uncertainty

Examples of AI Agents

AI Agent

Autonomy

Microsoft Researcher Agent (within Microsoft 365 Copilot)

Automates complex, multi-step research workflows by leveraging advanced AI reasoning models. It autonomously extracts key insights from massive internal and external data sources, integrates data from third-party platforms (like Salesforce, ServiceNow), and collaborates with other specialized agents to deliver comprehensive, context-aware insights. Demo video

Siemens Industrial Automation Agent

Siemens’ AI agents for industrial automation demonstrate autonomy by proactively executing complete workflows across the industrial value chain, moving beyond traditional query-based assistants to orchestrate multi-stage processes independently. These agents utilize a sophisticated orchestrator to coordinate specialized agents that understand user intent, adapt through continuous learning, and collaborate with both internal and third-party tools as needed. Source

DHL Enhanced Customer Solutions AI Agent

DHL Supply Chain recently announced its partnership with Boston Consulting Group to deploy two generative AI applications aimed at enhancing data management and analytics. This shift is intended to improve how the company assesses proposals and develops customer-specific logistics solutions Source

Microsoft Discovery (Enterprise Agentic R&D Platform)

A platform enabling collaboration with a team of specialized AI agents performing research, hypothesis testing, and iterative learning in R&D processes. Researchers configure agents with domain knowledge and processes, allowing them to collaborate autonomously to accelerate scientific discovery. Source

SharePoint agents:

SharePoint agents can analyze documents and answer questions about any site or library where users have permission, using natural language. Scenario video 1, video 2

Anthropic: Multi-agent research system

Anthropic’s multi-agent system (powered by Claude) is deployed in research, software development, and customer support. For example, enterprises use it for literature review, knowledge discovery, coding agents for GitHub issue resolution, and handling open-ended customer support scenarios that require both conversational flow and programmatic actions. Source

These cases illustrate autonomous agent capabilities like independent task planning, decision-making, tool use, and adaptive learning within complex enterprises and SaaS environments, exemplifying what defines agentic AI solutions today.

Summary

Agentic AI solutions show autonomy by independently perceiving environments, reasoning through complex tasks, and making decisions to achieve goals with minimal human oversight. They act flexibly and learn continuously, enabling examples such as autonomous vehicles, customer service bots, intelligent document handlers, and complex research assistants to operate effectively in dynamic real-world contexts.

This autonomy differentiates agentic AI from traditional AI tools by providing decision making and task completion capabilities, augmenting human behavior.

Subscribe now

Share

Subscribe now

The following is my conversation with M365 Copilot, using the Socratic method, to learn about Responsible AI initiative at Microsoft

Prompt Let’s use the Socratic method: You are a Socratic tutor. Please guide me through understanding the Responsible AI initiative at Microsoft by asking me a series of thought-provoking questions, and only provide explanations if I get stuck

Great approach! Let’s dive in Socratically. I’ll start with a foundational question:

Question 1: Why do you think a company like Microsoft would need a Responsible AI initiative in the first place?

(Take your time—there’s no single “right” answer, but think about risks, opportunities, and societal impact.)

Answer: To ensure that consumers of Microsoft's services feel confident and safe about using AI services

That’s a strong starting point! Building trust and safety is definitely central.

Follow-up Question 2: If trust and safety are the goals, what kinds of risks or harms do you think Microsoft is trying to prevent through Responsible AI practices?

(Hint: Think about both technical risks and societal impacts.)

Answer: One of the risks I see is where the AI goes rogue and tries to suggest actions that could be illegal or harmful to the user

Exactly, that’s a critical risk—AI suggesting harmful or illegal actions could have severe consequences.

Follow-up Question 3: If preventing harmful or illegal suggestions is one goal, what kinds of principles or guidelines do you think Microsoft might adopt to ensure AI systems behave responsibly?

(Hint: Think in terms of values like fairness, accountability, transparency, etc.)

Answer: One of the guidelines could include checks in the technology such that AI can identify when it is suggesting content that is not in line with the law, and stopping it from being responded

That’s a very practical guideline—technical safeguards are essential.

Follow-up Question 4: Beyond technical checks, what ethical principles do you think should guide AI development at Microsoft to ensure fairness and accountability?

(Hint: Think about values like inclusiveness, reliability, transparency, and privacy.)

Answer Ethical principles should include respecting user's privacy, being transparent on how the LLMs operate with the data, and being inclusive in the data that is used for training the LLM

Excellent! You’ve identified three of the core principles Microsoft emphasizes: privacy, transparency, and inclusiveness.

Share

Additional examples of Socratic prompts

Structural Prompt Example

· You are a Socratic study coach. Your goal is to help me deeply understand Microsoft’s Responsible AI initiative using the role-task-format approach: Ask me probing questions, encourage critical thinking, and only provide explanations if necessary. Let’s converse until I can explain the Responsible AI principles in my own words.

Opening the Dialogue

• "Let’s use the Socratic method: You are a Socratic tutor. Please guide me through understanding the Responsible AI initiative at Microsoft by asking me a series of thought-provoking questions, and only provide explanations if I get stuck."

• "Act as a Socratic tutor helping me to explore Microsoft’s Responsible AI principles. Begin with an open-ended question and continue the dialogue based on my responses, offering hints or follow-ups instead of direct answers."

Example Socratic-Style Questions

• "What are the core values outlined in Microsoft’s Responsible AI framework, and why might each be important for AI systems?"

• "How might fairness and inclusiveness be ensured when designing AI according to Microsoft’s policies?"

• "Why do you think transparency and accountability are included as guiding principles for Responsible AI at Microsoft?"

• "What are some real-world situations where reliability and safety could conflict with other Responsible AI principles, and how should such conflicts be addressed?"

• "How does Microsoft approach privacy and security in their Responsible AI practices, and why are these critical?"

Probing for Deeper Reflection

• "Could you ask me follow-up questions to challenge my understanding of how Microsoft balances ethical considerations in AI?"

  Leave a comment