In July 2026, when Microsoft UK apprentice Kasheef McLennon scrolled past his CEO’s LinkedIn post celebrating Accenture’s 740,000-seat Microsoft 365 Copilot deployment, he didn’t see just a corporate milestone. He saw a question no one was asking: What does this mean for the 18-to-25-year-olds who will actually use it? That question led to the first early-career AI hackathon co-hosted by Microsoft, Accenture, and Avanade in London—a hands-on event that may offer a blueprint for turning Copilot licenses into real workforce skills.
From a LinkedIn scroll to a real-world AI lab
McLennon, an Account Technology Strategist at Microsoft UK, spends his days helping customers understand Copilot. But staring at that post, he realized the industry’s conversation about the massive rollout had skipped a generation. “I thought, how are apprentices and graduates going to learn this?” he later told Microsoft UK Stories. Instead of leaving the thought in the comments, he reached out to Chidinma Iroegbu, a technology apprentice at Accenture. Together, they built something that had never been tried: a cross-company AI event designed entirely for early-career workers, not adapted from an executive workshop.
With backing from Microsoft’s MOSAIC early-career inclusion group and practical expertise from Avanade—the long-standing Accenture-Microsoft joint venture—the pair pulled together a one-day hackathon in London. About 40 participants aged 18 to 25 showed up. Sessions were kept short and pre-event study requirements minimal; the day prioritized building over listening. Teams created AI agents, generated presentations, and pitched solutions to a panel of specialists from all three organizations. Two projects stood out: an agent that could draft a full Request for Proposal (RFP) response, and a Client Template Agent that pre-populates documents with customer data.
Why building beats watching
This wasn’t a passive training session. Corporate AI education often starts with centrally produced videos, compliance modules, and product demos. Those materials can raise awareness, but they rarely answer the critical question: Where does AI fit into my actual job? A hackathon reverses the order. Participants begin with a problem they recognize, decide whether AI is appropriate, build something, and defend it in front of judges. That experience surfaces technical limitations and business questions that a polished demo conceals.
McLennon and Iroegbu deliberately avoided a lecture-heavy format. They discussed attention spans and cognitive overload, but the underlying principle was simple: a young professional who spends a day building a working agent will leave with more capability than one who simply watches a feature walkthrough. The event aimed to produce reusable artifacts—a prompt, an agent concept, a workflow tweak—not just good survey scores. For early-career employees, that hands-on confidence can be career-altering. You don’t just learn that Copilot can summarize meetings; you learn how to decide which meetings are worth summarizing, which data sources to ground the summary against, and when a human must still review the output.
The governance catch: when a good prototype goes rogue
The London teams operated in a controlled environment, but the prototypes they built highlight real enterprise tensions. The RFP agent concept, for instance, could dramatically speed up bid responses by pulling approved material from a governed library, classifying questions, drafting answers, and assembling sections. An RFP response can contain binding commitments, security claims, prices, and regulatory statements. A fluent draft is not necessarily an accurate or authorized one. The strongest implementation treats AI as an orchestration and drafting layer—not an autonomous bidder. Legal, commercial, and security owners still need to approve every section.
Similarly, the Client Template Agent’s value depends entirely on data quality. If account records are outdated or contradictory, the agent fills a polished template with wrong information. AI readiness is often data readiness in disguise. These lessons matter for both IT leaders and the young workers who may one day build such tools. Enterprises need what I’ll call an “authority ladder” for agents: at the bottom, the agent retrieves or summarizes information; higher up, it executes limited, reversible actions; at the top, it operates with delegated authority. Each rung increases potential value—and risk. A hackathon prototype fits safely at the first two levels. Production deployment demands formal architecture, security review, testing, and ongoing ownership.
How we got here: Accenture’s Copilot journey
Accenture’s deployment didn’t happen overnight. The company began piloting Copilot in 2023 with selected employees and senior leaders, expanded to roughly 20,000 users, and then scaled to 740,000 seats—Microsoft’s largest Microsoft 365 Copilot enterprise agreement to date. Those numbers made headlines, but as McLennon recognized, a license is not the same as productive use. Accenture has reported strong monthly active usage and internal surveys showing employees complete routine tasks faster. Yet self-reported time savings don’t automatically translate into audited financial gains, and a quicker task may require more review if AI introduces an error.
Still, sustained voluntary use across hundreds of thousands of workers is a stronger signal than a splashy executive demo. The problem hidden inside that success is fragmentation: there’s no single “Copilot user.” An apprentice, a marketing director, a security analyst, and a finance specialist have different data, risks, and definitions of value. Central programs provide governance and consistency, but they can’t translate AI into the language of every role. That’s where small, audience-specific events become crucial. The London hackathon asked young workers to discover how AI could help them, rather than assuming a generic adoption campaign would answer the question.
What you can do now: a hackathon playbook for the rest of us
If you’re an IT leader, a Microsoft 365 admin, or a business owner looking to replicate this model, don’t just order pizza and prizes. Start with a bounded business purpose. “Use AI to improve the company” is too vague. Pick a domain—reducing time to prepare customer meetings, improving access to policy documents, cutting rework in a document-heavy process—and define a measurable outcome. Ban certain data: live customer records, privileged legal material, source code, and security information should be off-limits for a rapid experiment.
Design your judging criteria before the event, and don’t let a polished pitch beat a sounder idea with rougher edges. A balanced scorecard should consider problem clarity, user value, data readiness, human oversight, security, feasibility, scalability, and originality. Judges should provide written feedback, not just winners. Every prototype, even a failed one, can teach something about architecture or risk.
Most importantly, build a path from prototype to production. Without one, good ideas die when participants return to billable work. An effective funnel looks like this:
- Participants submit problem statements before the event.
- Organizers screen for data sensitivity and feasibility.
- Teams build with approved tools and controlled information.
- Judges assess value, risk, and viability.
- Selected projects get a business owner and technical mentor.
- Security, privacy, and compliance teams review the design.
- A limited pilot tests the workflow against success criteria.
- Proven solutions move into a managed catalog or production service.
This preserves the energy of a hackathon while acknowledging that enterprise deployment demands sustained ownership.
A checklist for your own Copilot hackathon
- Pick a specific, measurable challenge. “Reduce the time to assemble a standard customer proposal by 30%” beats “use AI better.”
- Limit data access. Use synthetic or appropriately classified data. Never point a prototype at production SharePoint sites with broad permissions.
- Involve cross-functional judges. Technical polish isn’t everything. Include someone from legal, compliance, or a business leader who will inherit the tool.
- Require documentation. Teams should record their intended purpose, data sources, assumptions, and known limitations. This forces clarity and helps later governance reviews.
- Plan for day two. Before the hackathon, identify who will review promising prototypes and what resources they’ll need. A great idea without a follow-up is a missed opportunity.
Why every Windows and Microsoft 365 admin should care
On the surface, a London hackathon for apprentices looks far removed from your daily work. But the tools those young workers used—Word, Excel, PowerPoint, Outlook, Teams, SharePoint, OneDrive, and Copilot—are the same ones your users touch every day. Copilot doesn’t operate in a vacuum; it respects existing permissions. If a SharePoint site has overly broad access, or sensitivity labels are inconsistently applied, Copilot will surface that information more efficiently than a manual search. Before expanding AI use, audit your information governance: site ownership, sharing links, group memberships, sensitivity labels, retention policies, and legacy repositories.
A well-managed Windows device is essential for endpoint security, but Copilot readiness also requires work inside Entra ID, SharePoint, OneDrive, Teams, Purview, the Microsoft 365 admin center, and possibly the Power Platform. A fully patched PC doesn’t compensate for an ownerless SharePoint site with wide-open access. And young employees may arrive with extensive consumer AI experience that creates confusion. They need to understand that enterprise Copilot is governed by identity, tenancy, permissions, citations, and data classification—not just a chat interface.
Outlook: will the London experiment catch on?
The most striking part of this story isn’t that an apprentice attended an AI workshop. It’s that an apprentice conceived the event, coordinated senior stakeholders, managed a budget, and hosted the day. McLennon’s journey from casual LinkedIn discovery to three-company hackathon captures a truth about workplace technology: transformation rarely happens through licensing alone. It happens when someone notices a practical question, finds collaborators, and creates space for experimentation.
Participants’ most common request was for more time to build and network. That suggests the format generated real engagement. Future events could stretch into multi-day sessions, connect participants across countries and industries, or deliberately mix early-career and experienced workers—not as teachers and students, but as holders of different knowledge. The London hackathon was small, but it tests whether large organizations can give junior employees meaningful autonomy without abandoning operational discipline. That balance may determine whether enterprise AI becomes an empowering layer across the workforce or another centrally purchased platform that only a minority uses well.
For an apprentice, a degree-level assignment just got a lot more real. For the rest of us, it’s a reminder that the most valuable AI training might not come from a software giant’s headquarters, but from a room full of young people who were simply asked: what would you build?