During the annual IT budget review, the CIO of a food manufacturer finds three contracts she did not know about. Finance pays for a tool that reads supplier invoices. Legal uses a different one to review contracts. The quality department has its own solution that checks supplier certificates before a batch is released. All three are document AI, bought from three vendors, by three budget holders, in the space of eighteen months.
Nobody did anything wrong. Each team had a real problem, found a solution that works and built a reasonable business case. But the organisation now maintains three integrations, three security reviews, three sets of models to monitor and three audit trails for what is, underneath, the same piece of work: reading a document, extracting the right fields, checking them against something you already know, and sending the exceptions to a person.
The 4,980 use cases in the MetroMinds Library show how often that happens. They also show why document AI keeps getting bought again, what the real cost of that is, and how you can turn it into one shared capability without taking the decisions away from the teams that need them.
Document work hides in every function
Look at how use cases are described and documents turn up everywhere. In the MetroMinds Library, 1,120 of the 4,980 use cases (22.5 percent) start from documents of some kind: contracts, invoices, certificates, letters, attachments or scanned files. Only 53 of them sit in the Library's document understanding cluster. The rest are filed under other names, from compliance monitoring to agentic workflows.
Those 1,120 use cases cover all 12 business functions at the top level of the capability model, all 12 and 168 different . Operations takes the largest share, but supply chain, risk and compliance, customer management, legal and finance all have dozens of their own.
Some document types are especially widespread. Contracts appear in 353 use cases, spread across every one of the 12 functions. Invoices appear in 64 use cases across nine functions, including procurement, finance, sustainability and customer management. Within a single industry the pattern repeats: in manufacturing, document-based use cases sit in 11 different functions and 11 AI domains.
This matches what European companies report. According to Eurostat, analysing written text was the most widely used AI technology among EU enterprises in 2025, used by 11.75 percent of them, while 19.95 percent used at least one AI technology [1]. Reading text is often where an organisation's AI journey begins, which is also why it tends to begin in several places at once.
Why the same capability gets bought again and again
There are good reasons why document AI multiplies. Budgets sit with functions, so each function solves its own problem. Vendors package their products around one type of document, such as invoices, contracts or claims, which makes each product look different from the others. And the use cases themselves end up in different AI categories, so a portfolio list rarely shows that five of them do the same thing.
The labels make it worse. Of the 1,120 document-based use cases in the Library, 386 sit under content discovery, 137 under agentic AI, 99 under augmentation and 90 under knowledge intelligence. A CIO scanning a portfolio by AI type sees separate initiatives. A CIO who looks at what each initiative actually does with a document sees one recurring pattern.
The pattern is easiest to see when you put three examples next to each other.
What the Library says about the value
The Library makes the shared nature of document AI unusually visible. All 60 use cases in the document understanding cluster carry exactly the same , based on real-world implementations: a 25 to 45 percent reduction in manual review time, a 15 to 30 percent reduction in rework and a 10 to 25 percent increase in documents that pass through cleanly.
Those 60 use cases cover 33 different L3 capabilities and 60 different sub-industries. They include supplier invoices in grocery, aircraft maintenance log pages, lease scans in real estate, benefit eligibility evidence in government and batch production records in process manufacturing. The documents could hardly be more different, and yet the expected improvement is the same.
That tells you where the value comes from. It comes from the capability of reading and checking documents, applied to a volume of work. What differs between departments is the volume, the baseline time per document and the cost of an error, which is exactly what each business case should measure. The underlying improvement is not something you need to buy separately for each document type.
It also means you should be careful when several document use cases land in the same team. The benchmark records in the Library come with explicit warnings, such as not adding saved review time again through a reduction in staff. We explain why in our guide to double counted AI savings.
The real cost of buying it five times
The licence fee is the visible part of the cost, and often the smallest. Each separate solution needs its own integration with the systems that hold the reference data, its own security and privacy review, its own monitoring and its own audit trail. Each one needs people who understand how it works and who can retrain or correct it. When the vendor changes its model, each one needs testing again.
Those costs grow with every new document type. A shared capability has a higher starting cost, because the platform and its controls have to be set up properly once. After that, each new use case mainly needs configuration: the document types, the fields, the reference rules and the routing. The difference becomes clear when you put numbers next to each other.
There is a cost that does not show up in this calculation: inconsistency. Five tools read the same supplier's documents in five different ways, apply five different confidence thresholds and leave five audit trails in five formats. When an auditor asks how an invoice was checked, or a regulator asks how personal data in documents is handled, you have five answers to give instead of one.
What a shared document capability looks like
A shared capability does not mean one central team doing everyone's document work. It means one set of building blocks that every function can use, with the business rules kept where they belong. The use cases in the Library describe those building blocks in remarkably consistent terms.
First, intake and classification: recognising what a document is, whether it arrived by email, upload or scan. Second, extraction: pulling out the fields that matter, with a link to the page each value came from. Third, validation: checking those fields against reference data such as an order, a contract playbook or a specification. Fourth, routing: sending what does not match to the right role, with thresholds that decide when a specialist is needed. Fifth, an audit trail that records who accepted, corrected or rejected each result and why.
The data underneath deserves attention too. Scans, handwritten notes and photographed documents are a hard starting point for any AI. Gartner predicts that through 2026, organisations will abandon 60 percent of AI projects that are not supported by AI-ready data, and found that 63 percent of organisations either do not have or are unsure whether they have the right data management practices for AI [2]. For document AI, that means scan quality, consistent reference data and clear ownership of document types. Solving that once for the whole organisation is far easier than solving it five times.
Where to start
You do not need to stop every team and rebuild their tools. A practical route takes four steps.
Start by listing every AI use case that starts from documents, whatever its label, and link each one to the business capability and the role whose work changes. That alone usually reveals more document initiatives than anyone expected. Then group them by the steps they share. Most will fit the five building blocks above, with different reference rules.
Next, choose a first use case that is a good foundation for the others: high volume, clear reference data and a team willing to work with you. Matching invoices against delivery notes and orders is a classic example, because the reference data already exists and errors are easy to measure. Finally, agree how the capability is funded. A common model is central funding for the platform and its controls, with each function paying for the configuration of its own use cases and keeping ownership of its business rules.
Some document use cases also involve an agent that acts on what it reads, such as preparing a payment or a carrier action. For those, the same capability sits underneath, with extra approval rules on top. Our guide on whether a use case really needs an AI agent explains how to decide.
Questions to ask your teams
- Which of our AI initiatives start from documents, regardless of what they are called?
- How many separate tools, vendors or models do we use to read documents today?
- Which steps do those initiatives share, and which business rules are really specific to each team?
- Who owns the reference data each tool checks against, and is it the same data in several places?
- Can we explain to an auditor, in one answer, how a document was read, checked and approved?
- What would the next document use case cost on a shared capability compared with a new tool?
How MetroMinds helps you see the pattern
MetroMinds links every use case to a business capability from L1 to L3, to the role whose work changes and to an AI domain and capability cluster. In the Library you can filter by business function, process, capability cluster, and data readiness, so document use cases from different teams appear side by side instead of in separate lists.
In your own portfolio, you see which capabilities and roles your initiatives touch, where they overlap by L3 capability, and what each business case is worth with benchmark ranges and scope, adoption and . That gives you the facts to decide which document initiatives to combine, which to keep separate, and which to put on hold until a shared capability is in place.
Want to see how many document AI initiatives sit in your portfolio? Book a demo
Sources
- Eurostat, "Use of artificial intelligence in enterprises", Statistics Explained, data extracted December 2025.
- Gartner, "Lack of AI-Ready Data Puts AI Projects at Risk", press release, 26 February 2025.
- MetroMinds Library, 4,980 AI use cases with benchmark ranges based on real-world implementations, August 2026.
