Karthikeyan
Devadoss

In early 2024, I was asked to lead the Microsoft Copilot rollout across a squad of 18 people at NRG Energy's Berlin office. We were part of a 3-million-customer energy platform going through digital transformation. Here's what actually happened.

The first 30 days: resistance was real

Nobody asked for Copilot. It landed from the top — leadership had signed an M365 Copilot licence and wanted adoption metrics by Q2. My job was to make it stick. The first thing I learned: people don't adopt tools they don't trust. Half the team's first question was "is it recording everything?" We spent two weeks just on trust — data boundaries, what Copilot sees, what it doesn't, how GDPR applies to AI-generated meeting summaries.

What actually saved time

The use cases that worked weren't the fancy ones. They were boring and repetitive. Meeting summaries from Teams recordings. First drafts of status reports. Summarising long Confluence pages before a planning session. Rewriting ticket descriptions into proper acceptance criteria. By month three, the team was saving 4–6 hours per week per person. Not on complex thinking — on formatting, summarising, and first-draft work.

The GDPR blocker nobody planned for

We hit a wall in month two. One use case involved summarising customer interaction notes. Legal flagged it immediately — those notes contained personal data under GDPR Article 4. We had to pause, run a Data Protection Impact Assessment (DPIA), get sign-off from the DPO, and redesign the use case with anonymisation built in. It delayed that use case by six weeks. The lesson: GDPR review should happen before you build, not after.

What I'd do differently

Run a use case identification workshop in week one. Get the team to map their own repetitive tasks. People adopt tools they shaped, not tools imposed on them. Also: build a simple Power BI dashboard from day one tracking active users and weekly saves. Nothing drives adoption like showing someone "your team saved 47 hours last month."

Everyone talks about AI transforming industries. Fewer people talk about which specific use cases actually shipped in a real enterprise programme, survived a GDPR review, got adopted by real users, and delivered measurable hours back. Here are five from our programme.

1. Meeting summary automation (Copilot)

Teams meetings auto-summarised with action items and decisions. Simple. Boring. Saved 45 minutes per person per week across 18 people. First use case live, first one to hit 100% adoption. The key was making it opt-in and giving people the ability to edit the summary before it was shared.

2. Requirements document analysis (RAG pipeline)

Our platform has thousands of pages of specs, PRDs, and legacy requirement docs. We built a RAG pipeline on Azure AI Foundry so the team could ask natural-language questions against these documents. Reduced requirements review time by 40%. Three document types in production within four months.

3. Status report first drafts

Project managers spent 2–3 hours every Friday writing status reports. We built a Copilot prompt flow that pulled from Azure DevOps work items, recent Teams messages, and a structured template. First draft in 3 minutes. Edit and send in 15. This one had the clearest ROI of anything we shipped.

4. Ticket description rewriter

Developers hate writing acceptance criteria. A Power Automate flow with an Azure OpenAI call would take a rough ticket description and rewrite it as a proper user story with Given/When/Then criteria. Adopted by the whole squad within a week. Zero resistance — it made their jobs easier immediately.

5. AI governance use case screener

As the pipeline of AI use case ideas grew, our governance committee was meeting monthly to review them. I built a lightweight screening tool that ran each proposed use case through a structured prompt — GDPR risk, EU AI Act category, effort vs impact score. It cut governance meeting time by half and gave us a consistent framework for saying yes or no.

Running AI projects in a German enterprise in 2024 was nothing like the demos. Here are the three things that will slow you down — and what we learned from each.

1. GDPR is not a checkbox

Every AI use case that touches customer or employee data needs a Data Protection Impact Assessment. In Germany, with a 3M-customer dataset, that means the DPO is involved from day one — not day 60. We built GDPR review into our use case intake form from month two onwards. Before that, we had three use cases that had to be redesigned or paused after development had started. That's expensive. The fix: a simple two-page GDPR screening checklist runs before any AI use case gets approved. It flags personal data, data residency questions, retention periods, and model training risk. If it scores red, it goes to the DPO before engineering starts.

2. Hallucination is a product design problem

LLMs hallucinate. Every AI practitioner knows this. What most people underestimate is that enterprise users will lose trust the first time they see it — and they won't come back. We had one incident where a Copilot-generated summary included a decision that was never made. One person in the team saw it, mentioned it to two colleagues, and suddenly the whole floor was sceptical. The fix wasn't better prompting. It was human-in-the-loop design: every AI output is labelled "AI draft — verify before use," and critical use cases have a mandatory review step. Trust recovers slowly. Design for it from the start.

3. Change management is 60% of the work

The technology is the easy part. Getting 18 people to change how they work — even when the change saves them time — takes longer than you expect. We ran a monthly "AI wins" session where people shared what saved them time that week. We created a simple internal Slack channel for tips. We made early adopters into internal champions. None of this is in any AI delivery framework. But it's what made the difference between tools that got used and tools that got abandoned.