AI Trends 2026: The Era of Workflow Automation

Everyone is talking about artificial intelligence (AI), but few companies have content a model could actually use. In 2026 the focus moves from chatbots to workflow automation, and our platform starts precisely from the file where the documents live. This analysis walks through the five trends published by OpenText, sets them next to the Eurostat figures and says what they mean for a company of 20–150 people.

The problem is not the model, it is the content you feed it

The models have matured, but most organisations do not have data ready for them; the gap between ambition and readiness is the problem of 2026. Alison Clarke of OpenText opens the predictions for 2026 with a simple formula: AI is only as smart as the content you feed it. What is becoming clear, she writes, is that real AI progress is not about speed but about readiness.

The figures she cites come from two studies. Harvard Business Review Analytic Services found in 2024 that only 10% of organisations consider themselves fully prepared for AI. Drexel LeBow and Precisely found, also in 2024, that just 12% claim their data is AI-ready.

In Europe, adoption is recent. Eurostat shows that 20.0% of European Union (EU) enterprises with at least ten employees used AI in 2025, up from 13.5% in 2024. The most common uses are analysing written text (11.8%), generating pictures, video or sound (9.5%), generating language (8.8%) and speech recognition (7.2%).

What links the two sets of figures is content. Companies use AI to read text, which is exactly what they already have in emails, contracts and quotes, but that text sits scattered across mailboxes and personal folders. A good model with an incomplete file gives an answer that is confident and wrong.

At a company with a few dozen employees, the obstacle is not the technology. It is the product catalogue with three codes for the same item, the scanned contract with no date and the procedure only one person knows. No tool compensates for that; it makes it visible.

Five trends move AI from search to execution

OpenText describes five shifts for 2026: content moves to the boardroom, assistants execute, agents get rules, data stays put, documents read themselves. None of them is about a new model. All of them are about what happens around it: who decides, what it reads, what it is allowed to do and where the data sits.

The first is that content readiness becomes a board-level requirement, not an IT project. Whoever decides what gets automated also decides which documents must be complete and findable. At a small company, that means the owner, not a consultant.

The second is the move from search to execution, the shift we see most clearly. An assistant that finds the 2024 quote is useful; one that updates it with the new prices and sends it for approval is work done. The difference between the two is not in the model but in the steps someone has defined.

The third is that agentic AI scales only with guardrails. OpenText cites Gartner's 2025 estimate: over 40% of agentic AI projects will be cancelled by the end of 2027. The causes named are cost, unclear value and risk controls.

The fourth is that data stays where it is: multi-cloud architectures push towards zero-copy access and data sovereignty. For a small company, that means you do not move everything into a new system to use AI. You connect the existing ERP (Enterprise Resource Planning system), email and files, and give the agent access to what it needs.

The fifth is intelligent document processing, which becomes foundational to AI readiness. Invoices, requests and delivery notes enter the workflow read automatically, not typed in, and manual entry becomes the exception.

Workflow automation wins because it connects systems, not because it writes text

The value of AI in 2026 is not generated text but the link between email, ERP and approval, each step triggering the next without operators. The classic example is the invoice. It arrives by email as an attachment, and an AI Agent extracts the supplier, the amount and the due date.

The ERP checks it against the order and the stock, a person approves, and the payment order is generated; no step is spectacular. What matters is that they connect and that the workflow has a way back. When validation finds missing data, the invoice returns to the agent, not to an operator searching through email.

OpenText takes one more figure from IDC (International Data Corporation) FutureScape: by 2027, 80% of agentic AI use cases will require real-time, contextual and ubiquitous access to data. If the price list is in a local spreadsheet and the contracts are in email, the agent has nothing to connect. It is not blind because of the model, but because of the file.

The practical consequence is that automation is designed around systems, not texts. The question is not “what can AI write for us” but “which system receives the result and who checks it”. The answer to the second question draws the workflow.

Agents need rules before they get access

An AI Agent that executes needs the same rules as a new employee: who approves what, what data it sees, its limits, who is accountable. Of the three causes Gartner names for cancelled projects, the first two are solved by choosing the workflow well; the third is solved before granting access, with written rules.

The rules are not complicated. An AI Agent that proposes paying an invoice has an amount limit above which a person approves. An AI Agent that sends quotes sees only the customers it was set up for.

Every action leaves a trace in the log, with what it read and what it did. Exceptions, such as an invoice without an order, always go to a person. And the data stays where it is: the agent reads it in the ERP and in the file, not in a copy nobody updates any more.

Order matters: the rules are written before access, because after the first incident they are written in a hurry and too strictly. An AI Agent started with clear limits can be given more rights as the log shows it uses them well. One started without limits is stopped at the first mistake and never restarted.

In practice: a distributor with 60 employees prepares its content in three months

A distributor with 60 employees wants an AI Agent on invoices and quotes; the first three months touch catalogue, contracts and email, not the model. The company is fictitious; the situation is the one we see at every start. The distributor sells 4,200 items to a few hundred customers, with prices negotiated per customer, and receives invoices daily from 52 suppliers.

The audit in the first month shows why it cannot happen yet. The catalogue has the same product under three codes, because each warehouse clerk entered it as they saw fit. The supplier contracts are scanned, but 14 have no expiry date recorded anywhere, and the payment terms sit in emails from 2022.

The receiving and returns procedures exist in two people's heads and in no document. Quotes are made from three price lists, one of which is on a sales rep's laptop. None of this is about AI, and all of it is what the agent would read.

Months two and three are content work, not AI work. Duplicate codes are merged and the catalogue is exported from the ERP as the single source; the per-customer price lists go into the system, in one version. The contracts get an expiry date and an owner, and the receiving and returns procedures are written on one page and approved.

What remains is the archived email, in three employees' mailboxes, with commercial terms nobody ever moved into a contract. It is handled last, because it is the largest and the least structured. Only after that does the AI Agent start on invoices, with the amount limit and the approver set in month two.

At the end of month two, the preparation file looks like this:

Where it does not apply

The five trends describe companies with repeatable processes and documents that move between systems; in three situations, workflow automation is not the next step.

• When the workflow does not repeat. A company that makes ten quotes a year, each different, has nothing to automate. An AI Agent on a rare workflow costs more than it produces, and its rules never get tested.
• When the content does not exist yet. If the procedures live in two people's heads and the catalogue is a personal spreadsheet, the first investment is digitisation, not the agent. Without a file, automation copies the chaos faster.
• When the ERP cannot be connected. An ERP without an application programming interface (API), or a vendor that does not give access to the data, lets the agent read but not execute. Then you have an assistant, not a workflow.

What next

Do the content audit before any demo: for the workflow you want to automate, list the documents it needs and mark each one complete, partial or missing. What you cannot find, you cannot automate. If the list has more red lines than green ones, the guide on digitisation before automation shows in which order to fix them.

Frequently asked questions

Sources

1. AI is only as smart as your content: five predictions for 2026 · 2025
2. 20% of EU enterprises use AI technologies · 2025