M-Files AI: from Intelligent Metadata to AI Agents

For years, the primary role of artificial intelligence in enterprise content management (ECM) has been to analyse documents.

This involved extracting information, classifying documents, suggesting properties, and identifying similar content or patterns.

To be honest, that was already quite useful.

For years, M-Files has offered capabilities such as Text Analytics, Matcher, Smart Metadata, and Smart Classifier. These intelligent services analyze document content and existing metadata to help users classify and enrich information.

However something has changed. Whether you’re skeptical or enthusiastic, one thing is certain: advances in artificial intelligence are creating new possibilities.

M-Files understood it perfectly, and its latest AI capabilities are particularly interesting.

With Aino, Metadata Agent, Metadata Agent at Scale, Search Agent, Custom Agents, Custodian Agent and the move toward MCP, M-Files is evolving from a document management platform with AI features into something much closer to an AI-enabled information platform.

So, what are all these capabilities actually doing?

And perhaps more importantly:

How are they different from the AI capabilities M-Files already had?

M-Files & AI, a long love story

AI didn’t start with ChatGPT and other Large Language Models.

M-Files has offered “Intelligence Services” for almost ten years.

For example, when a document arrives, M-Files might help determine:

• what class the document belongs to;
• which customer it relates to;
• what project it belongs to;
• what keywords are relevant;
• which existing objects might be related;
• or what other metadata should be assigned.

Technologies such as Text Analytics, Matcher, Smart Metadata and Smart Classifier belong to this generation.

They are still valuable because they reduce the effort required from workers and solve a fundamental ECM problem: Users are not very good at consistently describing the information they create.

This is especially important because M-Files relies on metadata for views, searches, relationships, workflows, security, and lifecycle management.

These features had already established M-Files as both a visionary and a disruptor of traditional enterprise content management (ECM) systems.

But the new generation goes considerably further.

The shift: from intelligence services to AI agents

The easiest way to understand the new capabilities is not to view them as a list of features.

Instead, look at them as the different roles that AI can play with enterprise information.

That is a very different proposition from simply adding a chatbot to an ECM system.

Let’s take a look at them!

Aino: AI that understands

Aino, launched in late 2023, is undoubtedly the feature that users will first notice: the integrated conversational AI experience.

It can summarize documents and answer questions about their content. Depending on the configuration, users can ask questions about a specific document, related content, or broader vault content. Aino respects M-Files permissions when generating responses and can display the source documents used.

This changes the way users interact with information.

Traditionally, to understand a 50-page contract, I might:

• Find the document.
• Open it.
• Search for keywords.
• Read several sections.
• Compare information.
• Build my own conclusion.

With Aino, the interaction becomes:

“What are the termination conditions in this contract?”

Or:

“Summarize the main obligations for the supplier.”

Or:

“What financial risks are mentioned in this document?”

The important point isn’t that AI can summarize a document. Anyone can do that now.

What’s interesting is where the AI gets its context from.

Unlike other systems that simply throw documents into a generic chatbot, Aino works with the information available in the M-Files environment. According to M-Files, Aino uses vault content and applies user permissions when generating answers.

Metadata Agent: AI that understands what a document is

Metadata has always been one of the most powerful and frustrating parts of ECM.

We know metadata is important.

However, users don’t particularly enjoy filling it in.

This is where the M-Files Metadata Agent comes in.

The M-Files Metadata Agent uses large language models to analyze document content and suggest relevant metadata. Administrators can configure which properties to suggest and provide descriptions to guide the AI toward relevant information.

Imagine uploading this:

Supplier Contract – ACME Corporation.pdf.

Instead of asking the user to manually enter:

• Class = Contract
• Customer = dbi services
• Contract Type = Supplier Agreement
• Effective Date = 01/01/2026
• and Expiry Date = 12/31/2028.

The system can analyze the document and suggest these values. The user would then validate them.

This is a subtle but important change.

The traditional approach was:

• Human reads → Human interprets → Human enters metadata.

The AI-assisted approach is:

• AI reads → AI proposes → human validates.

This doesn’t eliminate governance. It just changes where humans spend their time.

Instead of doing mechanical work like extracting information, humans can focus on exceptions and decisions.

This is also where M-Files’ older intelligence services and new AI capabilities start to overlap.

So, what’s different from Smart Metadata?

I expect this to be a question asked by M-Files administrators and consultants.

After all, M-Files already has Smart Metadata and Smart Classifier.

Why do we need another AI capability?

The answer is not simply: “Because the new one uses an LLM.”

A more interesting difference is the type of reasoning that can be applied.

Traditional intelligence services are useful when defined patterns, classifiers, rules, training data, and matching logic are available.

Generative AI offers another possibility: You can describe what you want the system to identify using natural language.

Rather than trying to encode every possible variation of a contract, for instance, you can give the AI instructions that describe the meaning of the property and the information it should look for.

This doesn’t render traditional intelligence services obsolete. It provides administrators with another tool.

In many cases, the best solution will probably be a combination of both.

Metadata Agent at Scale: AI starts paying back information debt

Now, we’ll discuss a capability that I find particularly interesting from an ECM perspective: Metadata Agent at Scale.

There is a huge difference between: “Help me classify the document I am uploading.”

and: “Help me classify the 300,000 documents we already have.”

M-Files describes the Metadata Agent at Scale as a capability that enriches large volumes of existing content in the background. It is specifically designed for legacy archives, migrations, and metadata quality initiatives and can run enrichment jobs on a recurring schedule.

This changes the problem completely.

Consider a typical ECM implementation.

You have: 500,000 documents.

And perhaps:

• 80,000 missing a customer
• 60,000 with incomplete metadata
• 30,000 using an old classification model
• thousands containing information that was never extracted
• relationships between objects that were never created

Traditionally, this becomes a data cleanup project. Data cleanup projects are rarely anyone’s favorite.

Now, imagine being able to define what good metadata looks like and letting AI progressively enrich the existing content.

This is much more than automatic indexing.

It is information remediation.

There is a concept I have written about before that fits perfectly here: Information debt. We constantly talk about technical debt.

However, organizations also accumulate enormous amounts of information debt. This includes information that is poorly classified or described, difficult to find, or disconnected from the surrounding context.

Metadata Agent at Scale provides a mechanism to start paying that debt back.

This could be one of the most important AI use cases in ECM.

It’s not about creating more information. It’s about making the information we already have more useful.

Search Agent: stop searching like a database

Search has always been one of the most underrated features of an ECM system (Oh, I also wrote something about that!).

Traditional search assumes that users know how to articulate what they are looking for.

• They enter keywords.
• They apply filters.
• Then, they inspect the results.
• They refine the search.
• Repeat.

However, humans don’t always think in keywords. They think in questions.

For example:

“I need the contract we signed with the German supplier for the equipment delivered to the Basel project around 2024.”

This is a reasonable way to describe what someone wants.

However, it’s not an effective keyword query.

Aino Search introduces natural-language search, which allows users to describe what they are looking for instead of entering isolated keywords. The Aino Search Agent can then ask follow-up questions to help refine the results. According to the current M-Files documentation, this capability is still in limited beta.

This is an important evolution.

Traditional search:

• Query → Results

AI-assisted search:

• Describe → Results → Clarify → Refine → Find.

The difference might seem small. It isn’t!

It shifts the focus of searches from database interactions to conversations.

Once again, metadata becomes extremely important.

The better M-Files understands: customers, projects, contracts, people, dates, document types, and relationships, the easier it becomes for AI to understand what the user actually means.

This brings us back to one of my favorite ECM topics: Context.

Custom Agents: AI that can actually do something

• Aino answers questions.
• The Metadata Agent suggests metadata.
• The Search Agent helps you find information.

But what happens when AI participates directly in a business process?

This is where Custom Agents become interesting

M-Files introduced Custom Agents as AI-powered workflow automation capabilities. They can be configured with natural-language instructions and operate within M-Files workflows. An agent can read a document and the related business context, make a determination, and update only the properties it is authorized to change. M-Files currently describes Custom Agents as a public beta capability.

Consider the following invoice workflow.

Traditional case:

Invoice received → Employee opens invoice → Employee reads invoice → Employee checks contract → Employee decides → Employee enters metadata → Employee routes invoice

With an AI Agent:

Read invoice → Read related contract → Read supplier information → Reason → Set appropriate metadata → Route according to result → Employee handles exceptions

That’s no longer simply document intelligence. That’s process intelligence, and this distinction is important.

A traditional automation rule might say:

• If invoice amount > CHF 10,000, send it to the manager.

An AI agent can potentially reason over more complex information:

• “Check whether the invoice corresponds to the contractual terms, identify discrepancies, explain the discrepancy and route the invoice accordingly.”

That’s a completely different type of automation.

The governance question becomes much more important

This is also where I become more cautious about AI.

If AI provides an inaccurate summary, it’s frustrating. However, if an AI agent assigns the wrong keyword, I can probably fix it.

However, if an AI agent changes metadata, routes a document, triggers a workflow, makes a recommendation, or takes an action, then we have entered a different world.

We need governance.

M-Files’ approach is interesting here. For custom agents, M-Files states that the agent is constrained regarding which properties it can modify and that values set by the agent can include reasoning and source information to support an auditable trail.

This is exactly the kind of thing enterprise AI needs.

The question should not be, “Can AI do this?”

It should be: “Under what conditions are we willing to allow AI to do this?”

These are very different questions.

Custodian Agent: AI that looks after information

The next step is perhaps even more interesting.

What if AI not only helps users and executes workflows, but also helps maintain the information environment?

This is the idea behind the Custodian Agent.

One of the main challenges of ECMs is determining if a document is still relevant months after its creation.

Should we keep it in primary storage and pay the high price associated with it?

Can we optimize performance by reducing inactive content in our vault?

Should static documents, like PDFs, take full advantage of the capabilities offered by AI agents?

These features are useful, of course, but they also come at a cost in terms of tokens per query. Optimizing that cost is also part of the overall strategy.

Note: this feature is not yet officially released.

MCP: what happens when AI doesn’t live inside M-Files?

I also appreciate this about M-Files. Even though they provide incredible solutions, they don’t fear competition.

The MCP (Model Context Protocol) server is a perfect example of this. Unlike the Aino or Metadata agents, which are integrated into M-Files, the MCP server is an interoperability mechanism. It differs from the AI modules we’ve seen previously!

Today, we are still at the beginning of what AI can do for us. Often, the capability is limited to the app we are running.

But imagine tomorrow when we can use one AI agent that works across all enterprise apps (ERP, CRM, ECM, etc.).

This agent would need access to governed enterprise information.

M-Files has stated that it is developing support for emerging interoperability standards, including the MCP, to extend the trusted enterprise context to broader AI platforms and agents.

Note: This feature is still in development and will be available soon.

The real shift is from documents to information

There is a deeper point here. For decades, ECM systems have been built around documents.

We created:

• document classes
• metadata
• workflows
• permissions
• versions
• views
• retention rules

All of this was designed to manage documents.

However, AI doesn’t care about documents as files.

AI cares about:

meaning.

It wants to understand:

• Who is involved?
• What happened?
• What is this related to?
• What obligations exist?
• What decisions were made?
• What should happen next?
• What information is missing?

This is precisely where the M-Files approach becomes interesting.

Rather than focusing solely on the physical location of a file, M-Files has built its information model around metadata, relationships, and context.

As AI becomes more prevalent in ECM, this approach becomes increasingly valuable.

This is because AI needs context.

What does this mean for ECM consultants?

An important, even vital, question: Will AI replace my job?

Obviously, the answer is no, and, as with many other professions, we’ll have to adapt. But is that really surprising if you work in IT?

For years, our role as consultants was to ask:

• What metadata should we create?
• Which workflow should we configure?
• Which properties should be mandatory?
• What classification should we use?

Those questions aren’t disappearing. But new questions are arriving:

• What information can AI safely infer?
• What decisions can AI make?
• Which decisions require human approval?
• What changes should an AI agent be allowed to make?
• How do we measure AI accuracy?
• How do we audit AI decisions?
• What happens when AI is wrong?
• How can we design information models that AI can understand?

We have a crucial role now more than ever because we are designing more than just ECM solutions.

We are increasingly designing the environment in which humans and AI agents collaborate.

I find that much more exciting.

AI is changing the role of the ECM

The interesting thing about M-Files’ latest AI capabilities isn’t a single feature, but rather the direction in which they’re headed.

We started with AI that could help identify information.

Then, it could classify information.

Next, it could answer questions about the information.

Now, AI can enrich information at scale, conduct conversational searches, participate in workflows, and perform actions.

The next step is connecting that information to AI agents outside the ECM itself.

This represents a very different vision of enterprise content management.

Discussions and concerns about AI remain intense (and rightly so). Can we really go against progress? Or should we choose to embrace it and make the most of it? – I’ve made my decision!

We have reached the point at which we can begin designing an information environment in which humans and AI agents can safely collaborate.

If that is where ECM is heading, then metadata, relationships, context, and governance are not becoming less important. They are becoming the foundation.

The future of ECM may not be about using AI to manage documents. It may be about managing information for an AI-powered organization.