Balfour Beatty has committed £7.2 million to a multi-year, enterprise-wide deployment of Microsoft 365 Copilot, betting that generative AI will slash rework, tighten safety, and reshape how 27,000 employees work across its global infrastructure projects. But CIO Jon Ozanne’s most emphatic message is not about the technology. It is that the organizations most likely to succeed in the age of AI are those where HR and IT operate in lockstep.

The multinational construction, engineering and infrastructure group, which books roughly £10 billion in annual revenue, is no stranger to digital tools. Its long-standing partnership with Microsoft spans Office 365, Azure, Power BI and Power Apps. Yet the Copilot program signals a deliberate departure from a traditional IT-led rollout. Ozanne describes it not as a technology deployment but as “an employee-led business and cultural change phenomenon.” That framing explains why HR was embedded in the program team from day one.

What the £7.2 million actually buys

Microsoft 365 Copilot acts as an AI assistant woven into Word, Excel, PowerPoint, Outlook and Teams, and it surfaces insights from an organization’s own data estate—SharePoint, OneDrive, Teams conversations, and the wider Microsoft Graph. For Balfour Beatty, that translates into:

  • Instant document search and summarization across project archives
  • Automated meeting notes, action-tracking and follow-ups inside Teams
  • Natural-language queries over structured and semi-structured project data
  • Drafting assistance for inspection and test plans (ITPs)
  • High-level analysis of portfolio reports and dashboards

Crucially, Copilot operates within the enterprise’s tenant context. It respects user permissions, sensitivity labels, and data-loss prevention rules, which matters enormously when the datasets include safety-critical specifications, commercial contracts and client-sensitive information.

Beyond the horizon of first-wave productivity boosts, Balfour Beatty is co-developing “smart agents” with Microsoft. These agents move from analysis to action. They perform early-stage reviews of inspection and test plans, flag outdated templates, highlight missing approvals, and route issues to subject-matter experts automatically. This shift, Ozanne says, is where the heaviest operational leverage lies—cutting the chronic, expensive rework that afflicts major construction programs.

Early adoption: numbers that signal buy-in, not just benchmarks

The initial rollout covered 13,000 UK-based employees and produced survey results that are striking for a sector often stereotyped as digitally sceptical:

  • 75% of early users felt their work improved with Copilot
  • 78% reported stronger communication
  • 77% experienced reduced mental effort on mundane tasks
  • 66% said they would be more likely to accept a role where Copilot was available

These are subjective and early-stage, but they point to genuine user acceptance and motivation—variables that historically make or break large-scale IT programs. The company also reports neurodiverse employees describing greater confidence in meetings and presentations, a benefit that extends the value case beyond productivity and into inclusion.

Deployment was deliberately phased. “We don’t want to bet all our chips at once,” Ozanne told UNLEASH. The UK cohort served as pilot and proof-of-concept, giving the firm time to tune training, governance, and the change narrative before a planned global expansion.

Why Ozanne insists HR must drive the transformation

The CIO is unapologetically blunt: “The organizations that will thrive are those where HR and IT work in lockstep—aligning people, purpose and technology to build a confident, future-ready workforce.” In Balfour Beatty’s model, that meant HR owned:

  • The design and delivery of learning journeys for all levels, from frontline supervisors to executives
  • Job-role redesign and recruitment messaging to reflect AI-augmented work
  • Internal communications that answered the simple question: “What’s in it for me?”
  • Employee sentiment tracking and uptake KPIs

This people-first choreography vaulted adoption ahead. HR’s influence shaped where and how Copilot landed, ensuring that the tool was not just introduced but understood, embraced and embedded in daily rhythms. Ozanne stresses that HR leaders must not treat AI as outside their remit; they are central to how it scales and how it is perceived across a workforce with diverse digital confidence.

Operational tactics that make the HR-IT partnership real

Several concrete practices emerged from the program that other enterprises can replicate:

  • Cross-functional squads that co-locate HR business partners, learning and development specialists, security architects, and IT product owners.
  • Large-scale hackathons and hands-on bootcamps that create early champions and surface tangible use cases.
  • Measurement systems that blend qualitative data (surveys, storytelling) with quantitative metrics (time saved, actions automated, reduction in rework loops).
  • Sequenced regional rollouts that allow local customization while maintaining enterprise guardrails.

From a governance standpoint, HR also helps define the policies around acceptable use, performance standards, and reskilling pathways—all critical for maintaining workforce trust as AI becomes pervasive.

Safety, sustainability and the rework imperative

Rework is construction’s silent budget-eater. Industry data suggests that avoidable errors can consume a significant portion of project value, inflate timelines, and introduce safety hazards when crews return to sites for corrections. Balfour Beatty’s logic is that AI, by improving planning, enforcing correct templates, and providing faster access to historical lessons, will drive a material reduction in rework hours.

This links directly to sustainability. “If we achieve what we set out to with AI—reducing rework, streamlining inspections and optimizing resource use—the efficiencies gained and waste eliminated will far outweigh the energy costs,” Ozanne explains. The company is framing its AI investment through a carbon lens: energy consumed by generative models must be justified by the embodied carbon saved from fewer errors and less material waste.

It is a plausible argument, but its proof will require rigorous instrumentation. The firm must track hours saved, rework incidents before and after, materials conserved, and incremental cloud consumption. Only then can net carbon impact be validated rather than projected.

Security and governance: safeguards that can never be optional

Microsoft’s enterprise Copilot arrives with layered defences—identity-scoped access, Purview classification and DLP, encryption, and prompt-injection mitigations. For a company handling safety-critical designs and commercially sensitive data, these are non-negotiable. But they do not eliminate risk. Ozanne acknowledges that AI “brings with it advanced sophistication for criminals” and that data protection must become an even higher priority.

Practical governance essentials that Balfour Beatty and its peers should enforce include:

  • Strict sensitivity-label policies that prevent generative models from accessing top-secret commercial or personal data.
  • Endpoint DLP and monitoring to block users from copying classified content into unmanaged AI tools.
  • Logging and retention of prompts and responses for audit and eDiscovery.
  • Red-team testing of agent behaviour to catch prompt-injection or adversarial scenarios.
  • Vendor contracts that clearly delineate data processing boundaries, model training exclusions, and intellectual property protections.

For agents that execute tasks rather than simply suggest, the stakes are even higher. Balfour Beatty is adamant that any output touching safety, compliance or structural integrity must pass through a certified human reviewer. Autonomy ends where life-critical decisions begin.

Agents and the next AI frontier

Balfour Beatty’s focus on “smart agents” is the program’s most forward-looking element. These agents move beyond summarization and drafting into orchestration: they can triage data across systems, trigger secondary workflows like inspections or escalations, and maintain contextual state to surface exceptions proactively. On large projects, agent-style tooling has already been trialled to automate ITP reviews and to flag outdated templates or missing approvals—tasks that previously consumed hours of manual, inconsistent effort.

But agents introduce new challenges: observability, error handling, and the need for explicit human-in-the-loop checkpoints. The company treats agents as product features with version control, rollback plans, and incident-response playbooks, not as experimental add-ons. That product discipline will separate successful agent deployments from programmes that erode trust.

Reskilling the workforce for hybrid human-AI roles

Copilot’s presence is already shifting the employer value proposition. Two-thirds of early adopters said they’d be more likely to accept a job where Copilot was available—a potent signal for talent attraction in a tight labour market.

Reskilling priorities touch every layer:

  • Copilot operating skills: effective prompt design, output validation, and integration into decision workflows.
  • Data literacy: understanding data provenance, limitations and potential biases.
  • Domain-AI hybrid expertise: engineers and safety leads who can contextualize and qualify AI-generated recommendations.
  • Change-management capabilities: supervisors skilled in managing hybrid teams and setting realistic expectations.

Ozanne foresees deeper links with universities and training ecosystems, pairing construction’s longstanding apprenticeship tradition with modern AI literacy to build a sustainable pipeline of digital skills.

Risks every enterprise must manage

The program is not without vulnerabilities. Five stand out:

  1. Vendor lock-in. Deep integration with a single vendor’s Copilot and agent framework creates strategic dependence. Mitigation requires designing workflows to be cloud- and model-agnostic where practical, maintaining data export capabilities, and locking down contractual rights over custom agent code.
  2. Data governance gaps. Sensitive project plans or client data could be inadvertently exposed. Strict sensitivity labels, double-key encryption for high-sensitivity data, and continuous monitoring for anomalous prompt patterns are essential.
  3. Hallucination in safety-critical contexts. LLMs can generate plausible but incorrect advice. Every AI output that affects safety must be validated by a certified professional. Provenance, confidence scores, and traceable source links should accompany any technical suggestion.
  4. Workforce anxiety. Poorly managed change can stoke fear of job loss. Balfour Beatty’s narrative—role evolution, not elimination, with tangible examples of time reclaimed for higher-value work—must be consistently reinforced.
  5. Energy and carbon trade-offs. High-frequency AI use increases compute footprint. The cure is instrumenting incremental cloud consumption and comparing it against demonstrable reductions in rework and material waste.

A playbook for CIOs and HR leaders

The Balfour Beatty experience offers a replicable sequence:

  • Start with narrow, repeatable, mission-critical use cases tied to measurable results.
  • Embed HR business partners in the program team from the outset, co-owning adoption KPIs, training and role design.
  • Build a layered governance model covering data classification, DLP, retention, human oversight and agent approvals.
  • Pilot, measure and iterate in short sprints, adapting policy and training as confidence grows.
  • Invest in observable audit trails that capture prompts, responses and agent actions for compliance and post-incident analysis.
  • Treat agents as product features with versioning, rollback plans and defined error budgets.
  • Establish clear escalation pathways so that any AI output impacting safety or compliance goes to a certified professional for sign-off.

What success looks like

The ultimate test of this £7.2 million bet will be measured in hard operational metrics: reduced rework hours per project, fewer inspection delays, a sustained drop in safety incidents traceable to earlier detection, and a verifiable cut in material waste and carbon. Softer indicators—employee net promoter scores, the percentage of roles redesigned for AI augmentation, and the maturation of agent capabilities into trusted workstreams—will show whether the culture shift has taken root.

Large cheques grab headlines, but capital alone is not a strategy. Balfour Beatty’s approach reveals a programme that understands the difference: vendor co-development, phased rollout, HR-IT fusion, and a relentless focus on safety, sustainability and governance. Other enterprises, particularly those in field-service and infrastructure, can borrow its blueprint. But they must also internalize its unglamorous truth—that AI transformation is less about the licence fee and more about the long, disciplined work of aligning people, process and technology.

Jon Ozanne’s parting thought for HR leaders is a directive, not a suggestion: “Don’t treat AI as outside your remit.” For an industry about to be reshaped by generative agents, that one instruction may matter more than any line item in the IT budget.