Agentic AI becomes part of everyday work: observations from 500 professionals

This report sets out what Cominty observed among 500 users of its platform across 10 client companies between July 2025 and July 2026. These users are already incorporating agentic AI into their day-to-day work.

The purpose of the study is to observe how these users learn to work with agentic capabilities and adapt their ways of working, using a portal designed for business users and nontechnical professionals.

One agentic portal, two ways of working#

An agentic portal provides a conversational interface where a nontechnical user states a goal in natural language. An AI agent acts as the orchestration layer: it turns that goal into actions, selects tools, works within a digital environment, observes the results and adjusts its plan until it delivers a usable outcome.

With a conventional chatbot, the prevailing pattern is:

Question → generation of an answer

With an agentic portal, the pattern becomes:

Goal → understanding → planning → actions → observation → adaptation → validation → outcome

A conventional workflow follows a predetermined path; an agentic system can choose its actions and tools at runtime in light of the goal, the context and intermediate results.

Two modes of interaction#

Cominty examines how 500 professionals use the agentic portal. Although all have access to the same orchestration capabilities, the way they interact with it varies by assignment:

• End-to-end delegation. The user defines the goal and desired outcome. The agent takes responsibility for planning, execution and delivery from start to finish. The user steps in at key approval points. Examples: document analysis, ongoing monitoring, report preparation.
• Iterative assistance. The user works with the agent step by step, refining the results and steering what happens next. The work takes shape through successive exchanges.

What changed in a year#

Overall, in just over a year, working with AI has changed in character. One-off interactions—a question followed by an answer—have given way to genuine work sessions: users assign a task, iterate on it and return to it. The portal is no longer merely an assistant to consult; it has become a workspace where users carry out projects.

What makes the shift tangible is continuity. In our observations, one in eight sessions (12.5%) now extends over several days, compared with 1.4% a year earlier: users reopen a thread to continue, refine or correct work already under way. Agent memory extends that continuity beyond any one assignment: persistent instructions, preferences and decisions can be reused without having to restate them. AI is no longer a brief detour; it is finding a place in how a day, a week or a case file is organized.

This sustained pattern of work is now the norm: in July 2026, 77% of requests occur within work sessions of at least five exchanges, versus 35% a year earlier. Yet what matters is not the length of those exchanges, but what they produce. The next two observations focus on precisely that: how far users delegate, and what AI actually produces for them.

Observation 01 — End-to-end task delegation to AI: ×6 in one year#

Entrusting AI with an assignment means letting it carry the work through to an outcome. That is the shift we observe here: users are no longer simply conversing with the agent; they entrust it with an objective to pursue from start to finish.

These project-like work sessions bring together an initial instruction, autonomous execution and a delivered outcome. Given a goal and constraints, AI agents organize and sequence the necessary operations—research, analysis, tool calls, verification and production—without users having to direct each step separately.

Between July 2025 and July 2026, the share of sessions containing tasks completed end to end by AI agents rose from 4.6% to 29%—from fewer than one in twenty sessions to nearly three in ten. That is a gain of 24.4 percentage points and an increase by a factor of approximately 6.

This shift suggests a greater willingness to delegate. In these sessions, users can still specify a constraint, approve a direction, review the outcome or ask for a correction. Agent autonomy does not mean the absence of human oversight: it means the system handles execution between these checkpoints.

This way of working reduces the need to direct every operation and may make projects easier to deliver.

Depending on complexity, tasks involve anywhere from 3 to more than 40 sub-agent calls#

To complete a task, the primary agent—the orchestrator—delegates different steps to specialized sub-agents. Each performs a specific function: analyzing structured data, finding information, reading a document, producing a deliverable, checking a result or synthesizing multiple sources.

In July 2026, we observed that the number of sub-agent calls varied with the complexity of the assignment given to AI in a request:

• Simple assignments (3 calls) – 6.4% of requests: for example, "summarize this article" or "extract the key dates from this document." Median: 3 calls.
• Moderately complex assignments (4–15 calls) – 32.7% of requests: for example, "analyze our three quarterly reports and identify the trends." Median: 9 calls. These assignments combine reading, comparative analysis and synthesis.
• Complex assignments (16 to 40+ calls) – 60.9% of requests, the most common category. For example: "synthesize our sales data, cross-reference it with budgets and regulations, recommend priority actions, generate a report, then email it to the sales team." Median: 37 calls. These assignments involve extraction, analysis across multiple sources, verification, synthesis, writing and actions taken in company tools.

Across all requests involving orchestration, the mean is 34.8 sub-agent calls, and the median is 21 calls.

What the usage patterns reveal#

This distribution points to an important pattern: the agentic portal tends to attract substantial tasks. Trivial requests—those a conventional chatbot could handle better and faster—account for just 6.4% of the total. By contrast, 60.9% of requests involve intensive orchestration, requiring 16 or more sub-agent calls.

That does not mean users default to asking for complex work. Rather, it suggests a natural process of self-selection: professionals who turn to an agentic portal bring problems that warrant such complexity. Simpler questions remain with other tools. The agentic portal finds its place in workflows where breaking work into multiple steps adds value.

Observation 02 — Eight in ten users entrust AI with producing deliverables#

Having measured how work sessions evolve and how delegation deepens, this observation examines what users ask AI to produce and how they guide that production.

Definition — a recorded output is an artifact flagged by the platform on a request marked "success". This measure identifies instances in which an interaction creates an output distinct from the conversation.

Top 10 most common output categories (July 2026)#

Three agentic AI user profiles#

Regular users are the largest group. Power users operate at a different scale: several sessions a day, outputs every week, and systematic orchestration of agentic capabilities.

The three enablers of agentic output#

Delegating production does not diminish the user's role; it changes it. The user defines the desired result, supplies the working materials and sets expectations. The agent then draws on the methods, resources and tools needed to carry out the work.

The July data point to three enablers the agent combines to produce an output: Skills, which provide a structured method; internal context (files and documents), which supplies material specific to the company; and the web, which brings the outside world into the process.

1 — Skills: turning expertise into a reusable method#

To produce an output, the agent combines two complementary elements:

• Tool → performs an operation: reading a file, finding information, running a calculation or creating a document.
• Skill → applies a working method: reusable instructions specifying the steps, rules, checks and expected form of the result.

Example — preparing a presentation: tools consult the source documents, analyze their contents and create the file. The Skill defines the presentation's structure, the hierarchy of messages, the checks to perform, citation rules and formatting principles.

This level of adoption shows that specialization is becoming a routine part of producing work with AI. By capturing the task-specific steps, criteria, checks and rules for presenting results, Skills turn expertise into a method that can be reused immediately.

They make execution more efficient and precise: the agent need not devise its working method afresh for every assignment, and results follow a more consistent framework. A proven method can be used again, applied across multiple assignments and refined over time.

Whether selected by the user or activated automatically by the agent, Skills make this specialization available without adding complexity to the user experience.

The user states the desired result; the agent applies the appropriate method to produce it.

This shift takes AI from a general-purpose capability to one equipped for a specific task. Value lies not only in a model's general knowledge, but also in how its actions are organized.

2 — Internal company data: grounding AI in real work#

This context appears chiefly through two signals: the agent reads a file or calls on company documents (RAG). It can also come through a direct connection to business applications (via protocols such as MCP), allowing the agent to draw on tools the organization already uses.

The presence of company context points to use grounded in existing work. Users are not simply asking AI to create from a blank page: they give it material to revisit, analyze, compare, transform or present in a new form.

The work session thus becomes a space for transformation. A document can be read, analyzed, enriched, turned into a presentation, compared with other data and then revisited after validation. AI becomes part of an established sequence of work rather than operating separately from the company's resources.

For the organization, this is a structural shift. The more relevant context agents have, the better they can tailor their output to the documents, rules and knowledge specific to the work. The portal thus connects AI's general capabilities with the particular demands of each assignment.

3 — The web brings the outside world into the production process#

Internal resources are not the only material agents can draw on. The web plays a complementary role here. It can supply up-to-date information, a public reference, a point of comparison, a regulatory source, or data missing from the documents already available.

The observed level of use suggests that agents turn to the web selectively, when a task calls for an external perspective. An output can thus combine company knowledge with a broader understanding of the environment in which the company operates.

What the usage data reveals#

The central finding lies elsewhere: 68.9% of outputs combine internal context and Skills. In other words, agentic output depends less on the model's built-in knowledge than on pairing company-specific material (its documents, files, and internal data) with structured instructions (Skills). The model is not the source of the content; it is the engine that applies a method to a given context.

This pattern points to AI use grounded in a particular setting: AI operates within the company's environment, with its data and its rules. The value of its output depends less on what the model generally "knows" than on its ability to bring the right context and method to bear at the right moment.

Six key takeaways for 2026#

Six lessons emerge from these data. Together, they describe the same shift from six angles: value is moving from the model to the system, from the answer to the result, from execution to judgment, and from the isolated prompt to the shared method.

Producing outputs becomes the real signal of adoption#

Conversation frequency alone is no longer enough to gauge how fully people have adopted AI. The turning point comes when an exchange produces something that can have a life beyond the interface: a document to share, an analysis to put to work, a presentation to use in making a case, or a file to reuse. That is now true for a majority of users. The question is shifting from how good an answer is to whether an intention can be turned into a usable result.

The autonomy taking shape is grounded in context#

This is the study's most counterintuitive finding. The most advanced use cases do not show AI working alone, detached from the realities of the task. They show AI closely connected to the user's documents, files, methods, and constraints. Internal company context carries several times more weight than web search in the outputs produced. The more concrete the task, the more central context becomes. Agentic autonomy is, first and foremost, the ability to act within a working environment.

The model is no longer the only measure of performance#

The value of an agentic system no longer depends solely on the capabilities of the model generating a response. It rests on the entire chain the system can draw on: context, methods, tools, specialized resources, sub-agents, and execution mechanisms—with a single task involving anywhere from a few to several dozen sub-agent calls, depending on its complexity. The decisive question is no longer just "Which model should we use?" but "What system should we build around the model to carry out the task reliably and repeatably?" Orchestration is becoming as consequential a capability as the intelligence of the model itself.

Domain-specific methods become a reusable asset#

Formalized methods mark a profound change: a way of working no longer has to remain locked in an exceptional prompt or in the know-how of a single advanced user. In practice, these methods are already present in a substantial share of the outputs observed. A formalized method brings together steps, rules, sources, and a specified output format—making it, by design, something that can be retained and passed on. An organization's agentic maturity will therefore also depend on its ability to turn its practices into executable, shared methods.

The human role shifts toward higher-value decisions#

As the system takes on more of the intermediate work—with end-to-end delegation increasing more than sixfold in a year—the user is less involved in steering every step. Their role increasingly centers on defining the objective, selecting the context, setting constraints, weighing trade-offs, and validating the result. AI does not eliminate human responsibility; it makes the ability to frame the problem well, assess the output, and decide how to use it more important than ever.

Simplicity becomes a property of the system#

A simple interaction does not mean the work behind it is simple. It means the system can absorb, organize, and make sense of the complexity needed to carry it out. This is the central paradox of the study: a request expressed in a single sentence can set dozens of coordinated operations in motion. Progress, then, is not about removing the steps, but about sparing the user from having to coordinate them one by one. The user states the desired result; the agentic architecture assembles the means to achieve it.

The path taking shape#

What we observe is a gradual move toward AI becoming part of the company's production process. It no longer merely supplies information: it helps organize work, apply methods, draw on resources, and produce usable outputs—while leaving direction and final judgment to people.

In 2026, agentic maturity is judged not only by what AI can say, but by what an organization can enable it to accomplish—within its own context and under its control.

Scope of the results#

Source: all users and all sessions in July 2026 on the Cominty.ai portal. Depending on the metric, percentages are based on 500 active users in July 2026, all sessions, or sessions with an output.