Between 2023 and 2024, as Director of Digital Initiatives at an educational publisher, I led the adoption of Microsoft Copilot and other large language model (LLM) platforms across three functions: project management, content development, and software testing. The results included time returned to staff, faster initial drafts, and more proactive quality assurance on the development side. I did this without direct managerial authority. Applying AI tools to real deadlines was the primary way to prompt adoption.

Although generative AI is strong at summarizing and drafting text, its appropriateness for customer-facing vs. internal-facing content varies. Focusing on internal-facing work provides more room to get used to outcomes from the tools. Business documents that inform or capture decisions are good areas to look. The examples below come from supporting-content creation, project communication, and initial software acceptance review.

Once a set of opportunities are identified, review each on these four dimensions:

• Achievable. Look for delegable tasks — sometimes called “Quadrant 3” work in time-management frameworks: urgent but not high skill, and therefore easy to hand off. A bonus: these are usually the boring-but-necessary tasks nobody wants to keep doing.
• Focused. Confine the work to one area, or to a single, encapsulated sub-process.
• Measurable. Track time, cost, or volume against a real baseline, and hold the output to the team’s existing quality bar.
• Visible. Pick work that somebody is already waiting for. Success should visibly speed up or improve something colleagues regularly receive.

Put together, the result has to matter to the people doing the work — the ones who know the hands-on baseline and what “good” looks like.

These criteria especially matter if you're leading without direct authority. Rather than directing another team to change how it operates, start by identifying a delegable task that's necessary but avoided because of its size.

Offer to take it off the team's plate for one cycle, and show them what comes back. From that delivery, you can discuss how to adjust the process to take advantage of it. Typically, you can point to gains in turnaround time, volume processed, and consistency of output.

Content labeling. A WCAG 2.1 AA accessibility initiative covering more than 25,000 educational resources needed captions and descriptions. I worked with the editorial team on prompt engineering against our house style guide and vocabulary until the output matched what an editor would have written by hand. The AI vendor used that work to train its model to the task in our house style. The volume wasn't achievable through manual work alone, and the deadline made the case for the tool. A side benefit was consistency: once a concept's labeling convention was set, the output didn't vary — the only remaining question was controlling false positives and negatives.

Software testing. We already regularly analyzed technical support records — monthly email and chat transcripts — to find the issue types generating the most volume, and routed that analysis to a standing cross-functional fulfillment and onboarding team to address. LLMs helped us identify trends with word clusters, without complicated steps such as regular expression crafting and scripting.

We also had to run pre-production tests on student data visualizations. A second application generated synthetic student activity data, which the team used two ways: to exercise reporting systems with realistic data, and to debug student-facing interactive features.

Project management. An initial, obvious application is generating meeting minutes. Additionally, much time was spent pulling use cases, requirements, and other expectations/promises out of various slide decks and minutes so that work-package definitions and expected outcomes could be corralled and reviewed. Project managers had always done this by hand. With an LLM producing the first draft, their time shifted from hunting to clarifying.

The most value comes when staff and AI reinforce each other.

By 2026, meeting transcripts and AI-generated summaries are common. But using them well still requires coaching. For example, in a meeting, AI can perform real-time speech to text and summarize. But I recommend treating these as ephemeral. Current AI output tends to be highly detailed on one dimension and almost wholly devoid on other key dimensions: it captures words nearly perfectly and loses nearly all tone, body language, and timing.

An effective meeting still needs a human lead: someone who tracks which topics are open, which are hard for the team to discuss, and what's left to do. A human chair can also verbally reconfirm action items and get the assignee's spoken agreement to the task. For a simple post-meeting action list, a human facilitator produces something shorter, more accurate, and more effective — because it's accepted out loud, in front of the group.

The application, although simple, meets the criteria. Managing minutes is a chore, but it’s a key communication method in business. It’s both achievable and typically the duty of one particular group or type of employee at a firm. Minutes are visible because pretty much everyone in the business uses meeting results. The speed and consistency of them is measurable. When the business gets used to same-day minutes that are similarly formatted, many common misunderstandings, and meetings to unwind them, will disappear.

Identify a few possible projects: Do they play to your AI tools’ strengths? Are they clearly defined? Which among those are achievable, focused, measurable, and visible? Once you've picked one, write up what you found and start a conversation with the people who'd be involved.