Key Points at a Glance

• • The greatest benefit comes from structured sources: Documentation based on code, flows, and configurations can be generated with minimal additional work.

• • Least useful when content still needs to be created: concepts, presentations, and proposals. The bottleneck here isn't the task itself, but the lack of context.

• • Code: The time savings are real, but part of them is converted into technical debt that only becomes apparent during operation.

• • No review, no enterprise deployment: Humans remain part of the loop—even if AI helps with the review itself.

• • The next lever is context: preconfigured templates, accessible company knowledge, and agents turn what is currently the weakest area into the greatest source of potential.

 

Presentations: Quick Start, Noticeable Follow-Up Work

A complete set of slides—complete with structure, titles, and an argumentative framework—can be created in minutes instead of hours. The greatest benefit is for people who don’t create presentations every day: there’s no need to start from a blank document. Once a template is set up, the AI also works directly within the corporate design.

The downside: Many drafts are compelling at first glance but are lacking in detail. Context is missing, the line of reasoning is disjointed, and individual statements don’t hold up. In addition, the text is often recognizable as AI-generated—with repetitive sentence patterns, lists of three, and a strong emphasis on balance without taking a stance. The editing process eats up a large portion of the savings. Still, the bottom line is a clear gain.

Context: Researchers from BetterUp Labs and the Stanford Social Media Lab describe this phenomenon as a “workslop”—AI output that looks good but doesn’t actually advance the task. About 40 percent of respondents had received such output; the rework takes an average of just under two hours per case.

 

Concepts and Documents: Structure, Yes; Attitude, No

The outline, chapter structure, and first draft are quickly in place; the blank page is no longer an obstacle. Existing material—such as notes, meeting minutes, and email threads—can be reliably consolidated into a rough draft. Formal completeness, structure, and transitions are well addressed.

Follow-up work is required whenever technical details come into play:

• • The depth of detail and the connections between the chapters must be established manually.

• • Terms are used correctly, but not always in the right context—something that often goes unnoticed except by experts.

• • The AI determines when a clear recommendation is needed.

• • Customer-specific details, figures, and references must be verified: Fabricated information is phrased in a plausible manner.

Context: A widely cited study by Brynjolfsson, Li, and Raymond (NBER) finds that generative AI leads to an average productivity gain of 14 percent—but as much as 34 percent for employees with little experience. This aligns with our observation: Those who write infrequently benefit the most.

 

Documenting Solutions: The Most Obvious Benefit

This is where our assessment is most clearly positive. The reason: The source material is already structured. The AI doesn’t have to invent anything—it just has to translate. Code, PnP templates, Power Automate flows, Power Apps, and configurations are transformed into neat, complete documentation—with significantly less follow-up work than is required for concepts or presentations.

This makes something that is often neglected in many projects economically viable: documentation that reflects the actual status of the solution. And there’s a second benefit: What is documented today will serve as the context from which agents, support, and onboarding teams draw tomorrow. Especially for solutions on the Power Platform and in process digitization, this is a tangible advantage for operations and further development.

Context: Google Cloud’s DORA research program identifies high-quality, accessible internal documentation as one of the most powerful drivers of AI benefits and, as early as 2024, found a measurable correlation between AI adoption and improved documentation quality.

 

Code: Written quickly, paid for later

What's already working today

• • Initial Setup: Project structure, boilerplate code, configuration, and test framework are created reliably and quickly.

• • Individual, clearly defined modules and features are implemented to a high standard.

• • Problem-Solving and Troubleshooting: AI is a powerful sparring partner for hypotheses.

• • Proof of concept and prototypes, where speed is of the essence and the code has a limited lifespan.

When the bill comes later

• • The larger and more complex the codebase, the more errors the AI introduces.

• • This results in classic technical debt—which only becomes apparent during operation and when making enhancements.

• • Code needs to be reviewed and corrected. AI helps with this, but it does not replace the review.

• • Without a human in the loop, AI-generated code is not business- or enterprise-ready.

 

Context: According to the 2025 Stack Overflow Developer Survey, about 84 percent of developers use AI tools or plan to do so, but only about one-third trust the accuracy of the results. Two-thirds cite solutions that are “almost right, but not quite” as the biggest time-waster. GitClear has analyzed over 623 million code changes and has observed an 81 percent increase in duplicate code blocks since 2023, while the proportion of refactoring has declined sharply. And in a 2025 study by METR, experienced developers using AI were 19 percent slower—but believed they were 20 percent faster. So there’s a disconnect between perception and measurement.

 

Making Information Useful: An Underestimated Use Case

Collecting, consolidating, and tailoring information to the intended audience is a strength of AI that is still rarely used systematically in many companies. One example from our field is Microsoft 365 updates: the roadmap, Message Center, and release notes are continuously analyzed and consolidated into a brief overview for each customer or topic.

The same pattern applies to partner products, security advisories, and market monitoring. The effort involved is a one-time setup; after that, the benefits are recurring and require no additional staff.

 

Agents with Company Knowledge: From Generic Assistants to Colleagues

The leap from a generic assistant to an AI that “knows the company” comes from context—not from a better model. Typical use cases we’re currently evaluating:

• • Sales Agent: Prepares an initial draft proposal based on previous offers and service descriptions.

• • Product Agent: Answers questions about partner products based on manufacturer documentation.

• • Operations Agent: Accesses operations manuals, tickets, and documentation, and provides support for technical support and onboarding.

This requires structured company knowledge—and this is exactly where good documentation pays off a second time. Agents target the areas where AI is currently least effective without context: quotes, proposals, and recurring product questions. We show how to set up these agents in Microsoft 365 on our page about SmartAI365—AI and Copilot in Microsoft 365.

Context: The MIT study “The GenAI Divide” attributes the failure of most AI pilot projects not to model quality, but to a lack of integration into workflows and a lack of access to internal knowledge.

 

Outlook: From Prompt to Graph

The many terms related to AI agents can be viewed as stages of the same process. The first two determine what the model sees, and the next three determine how it works:

• • Prompt: better wording, examples, format—applies only to the current call.

• • Context: what appears in the context window—company knowledge, documents, history, memory.

• • Agent and Harness: The model is allowed to act—using tools, an execution environment, and permissions.

• • Loop: how long and how thoroughly—step budget, check loops, termination criteria.

• • Graph: multiple agents in a fixed sequence with branches and handovers.

[Image placeholder: Graphic “From Prompt to Graph” (5 steps) – will be uploaded to WordPress and replaced]

Our experience: The bottleneck is rarely the orchestration; it’s almost always the context. Those who invest here get more out of every stage.

 

What we conclude from this—and what we recommend to you

• • Start where the source material is already organized: documentation first—it pays off twice over.

• • Setting up templates: for presentations and documents, so that even the first draft is created in line with the corporate design.

• • For concepts and proposals: AI for structure and rough drafts; reasoning and tone are crafted by hand.

• • When it comes to code: small, well-defined units rather than grand schemes. Review remains mandatory; proof of concept (PoC) and troubleshooting are the strongest use cases.

• • Plan for rework: Include it as a fixed part of estimates, rather than treating it as an exception.

• • Evaluate agents systematically: Start with a real-world workflow and measure the benefits.

 

The core information at a glance

Core benefits

A quick initial estimate across all areas; most pronounced for documentation from structured sources and for individuals who rarely perform a task

Borders

Technical depth, independent perspective, and accuracy of figures and facts; in code, growing technical debt without consistent review

Key to success

Context Instead of a Model: Templates, Curated Company Knowledge, and Agents Embedded in Real-World Workflows

Target group

Companies in Switzerland that want to use AI and Microsoft 365 Copilot in a productive and controlled manner as part of their daily work

 

Conclusion

AI has become part of everyday work—as a tool to speed up initial results, not as a replacement for expertise and judgment. Those who realistically assess its benefits, plan for follow-up work, and invest strategically in context will get significantly more out of it than by simply using the latest model. People remain part of the loop; the only question is where their time makes the biggest difference.

 

Using AI Productively – with innobit

As Microsoft 365 specialists, we guide companies from their first use of Copilot all the way to agents that access your company’s knowledge. In our Copilot workshop, we identify the use cases with the greatest impact in your company; using SmartAI365, we implement them with technical precision and in compliance with governance requirements. To ensure the implementation is successful in day-to-day operations, we also support you with change and adoption management. You can also find an introduction to Microsoft Copilot in our article “Getting Started with Microsoft Copilot.”

 

Request a consultation now – We’d be happy to share our experience with you in a one-on-one conversation.

 

Sources and further reading: BetterUp Labs / Stanford Social Media Lab (“Workshop,” 2025); Brynjolfsson, Li, Raymond: “Generative AI at Work” (NBER); Stack Overflow Developer Survey 2025; METR (2025); GitClear Maintainability Gap Report (2026); DORA – State of AI-Assisted Software Development; MIT NANDA: “The GenAI Divide” (2025)