In 2026, talking about AI "in the enterprise" has become easy. It's much harder to understand where it's truly creating productivity and where, instead, it remains an interesting chat but has little impact on daily work.
This is why it makes sense to start with usage data and how it changes over time. Not to chase statistics for their own sake, but to answer a concrete question.
where AI is already working as an operational lever and what is needed for that value to become stable and measurable.
Why we need AI data in 2026, not just anecdotal cases.
In recent months, many companies have accumulated micro-tests, demos, and internal experiments. The problem is that these signals are often contradictory.
In some teams, it seems that AI “works” and saves hours, while in others the perception is that “nothing changes.”
The point is that, without data and a minimum of methodology, we risk measuring the wrong thing. Use doesn't equate to impact. A team can use AI extensively and produce a lot of output, without it actually improving a process. Conversely, a less visible adoption can generate value because it fits in at the right point in the workflow.
In this regard, large-scale usage reports such as the Anthropic Economic Index are interesting. They observe how AI is used and how habits change with experience, also from a learning-by-doing perspective.
What we mean by real work
When we say real work we mean three things.
The first is recurring tasks. Activities that repeat every week, with variations, and that consume time because they require information research, reprocessing, standardization, and monitoring.
The second is the risk-benefit ratio. Not all activities are equivalent. In a regulated or sensitive context, a "well-written" response isn't enough. Consistency with updated rules, policies, and constraints is essential. Here, the value of AI emerges only if there is a clear framework and a way to reduce ambiguity and rework.
The third is ownership. A task becomes measurable when someone owns the process and can say, "We used to do it this way, now we do it this way." Above all, they can decide how to integrate AI without leaving it to individual initiative.
Here comes a key word that is often underestimated
Agent. Not in the technical sense, but in the operational sense. When AI stops being a conversation and becomes a component that serves a flow, with rules and boundaries, then it becomes easier to measure the changes.
What Usage Reports Show and Why It's Not the Same for All Jobs
The data helps us see two things.
The first is that usage diversifies over time. In its March 2026 report, Anthropic notes that requests on Claude.ai have expanded, with a lesser concentration on top tasks than in previous months.
This is consistent with a classic adoption dynamic
Initially, AI is used for a few high-value cases, then it expands to a wider set of activities, even more “everyday”.
The second is that certain types of work tend to migrate from conversational use to more structured automation and workflows, especially when APIs and integrations come into play. The report explicitly mentions the migration of coding to more automated API workflows.
This is an important track even for those who don't make software
When an activity lends itself to a repeatable and governable sequence, AI is more likely to move from “chat” to a workflow.
The point that many companies underestimate from chat to workflow and why the value changes there
As long as AI exists only as an individual aid, its value is fragile. It depends on the skill of the individual, their discipline, the time they dedicate to iteration, and the context they can provide.
But when AI is embedded in a workflow, three things happen.
The first is that the output becomes more standardized. Not because it becomes "cold," but because it follows shared constraints.
The second is that it becomes more measurable. Because the flow has a clear before and after, and because you can compare the time spent, rework, and repetitive errors.
The third is that it becomes more secure. Because a workflow brings with it minimal rules, access boundaries, and validation criteria, and therefore reduces casual use.
In short, the leap isn't about "using a better tool." It's about moving from AI as chat to AI as a part of the process.
Where AI tends to bring value first non-technical examples in regulated contexts
When the goal is to remain informative, it is useful to describe typical processes, not solutions.
Compliance and risk in banks or public administration
Imagine a bank manager or an executive working on tasks that require constant reference to updated rules, policies, and constraints. The key isn't "writing better" but reducing arbitrary interpretations and rework.
Here, AI adds value when it helps quickly retrieve the right rule, summarize it in an operational way, and streamline the verification process. The gain isn't just time, it's order. Less back-and-forth, fewer confusing versions, fewer late corrections.
B2B Sales Enablement
In many commercial contexts, the real challenge is maintaining consistency between materials, offers, factsheets, FAQs, and messages. AI can become an accelerator in preparing responses and materials, especially when the team works with "official" and up-to-date sources.
Value emerges when responses become more aligned, faster, and with less risk of saying things that are inconsistent with the actual offering.
Operations and support
The typical benefit here is reducing friction in repetitive and information-intensive tasks: sorting requests, summarizing, internal research, and preparing first operational drafts.
The point isn't to automate everything. It's to reduce the mechanical part to free up time for decisions and exceptions.
Because impact grows with expertise learning by doing and team standards
One of the most interesting signs from the data is that experience matters. In the report, Anthropic observes that older users have higher success rates in conversations and tend to use AI on more complex and higher-value tasks.
This is consistent with a simple idea
It's not enough to "have AI," you have to learn to use it well.
But there is a detail that makes the difference in the company
If this skill remains individual, its impact remains unstable. If it becomes a team standard, then it becomes scalable.
Standard means very practical things
clear use cases, shared framework prompts, minimum testing criteria, and above all ownership.
What to do in the company five light moves to move from experimentation to practice
- Understanding the real need
Before the tool, it's important to clarify the friction. Slow search. Rework. Inconsistency. Too many steps. Too much variability. - Drawing the human as is in a timely manner
We don't need an endless project, but we do need to understand where we're wasting time today, where we're making mistakes, and where we need to double-check. - Set realistic expectations
AI accelerates and organizes, it doesn't replace responsibility. If we treat it as a "source of truth," it becomes a risk. If we treat it as an accelerator, it becomes an ally. - Put in place a minimum infrastructure of trust
Compliance, security, transparency. Not as bureaucracy, but as a prerequisite for using AI without blockages or incidents. - Taking usage from individual to workflow
Even with just light automation and shared rules. The leap in value occurs when the process is independent of the individual user.
Mini checklist ready to copy
- Choose only one process with real pain
- appoint an owner who owns it and measures it
- define a simple before and after baseline
- establish minimum rules on sources and data
- decide on a primary and a secondary metric
- do a short, iterative test, then standardize what works