Article | November 02, 2026
Discover the governance gap stalling enterprise AI, the four sets that let AI diffuse into culture and why resilience now belongs in every operating model.
By Hari Prasad Govindarajan, Chief AI & Transformation Officer, DXC Technology UK&I
The pilot works. The slide deck calls it a win. Then, someone asks what happens if the model gets something wrong in the early morning hours: Who notices? Who decides? Who answers for it?
The room goes quiet.
Call it the 2am test. Ten major research reports crossed my desk this quarter, including from KPMG, EY, McKinsey, Bain and the World Economic Forum (WEF), and every one of them, in its own words, measured that silence. Adoption is no longer the constraint: McKinsey found that 88% of organizations are already using AI regularly in some part of their business. What stalls is everything after. Bain’s executive survey captures the paradox in a single pairing: 80% of generative AI use cases met or exceeded expectations, yet only 23% of companies can tie them to revenue or cost outcomes. McKinsey found that fewer than one in ten high-value use cases ever makes it past pilot.
The technology performed, but the organization could not account for what it did. That is an accountability problem, not a capability flaw, and accountability is a governance question. That’s why the next phase of enterprise AI, which we at DXC call AI Reinvent, is not a technology program at all, but an operating model and culture program with governance at its core and technology as an enabler.
Diffusion is the point at which AI becomes part of how an organization works. It does not happen naturally; it has to be built in and governed from the start. In financial services and technology firms, we find that AI diffuses into the culture only when the organization addresses four sets simultaneously, each with a named owner and clear rules for how AI is used, checked and corrected:
Do people see AI as a teammate or a threat?
Left unmanaged, fear plus ambiguity produces shadow adoption. KPMG’s study of 48,000 workers found 57% of employees hide their AI use and present its output as their own. As a result, your AI usage dashboards look modest while AI is everywhere.
Is work redesigned around AI or AI is simply bolted onto old work?
Deloitte finds 37% of companies use AI with little or no change to underlying processes, so productivity gains evaporate before they reach the P&L, because the surrounding workflow still runs at the old speed.
Will people supervise AI, not just use it?
KPMG finds 66% rely on AI output without checking it, a skills gap masquerading as a trust surplus. That means errors are discovered downstream, by customers or regulators, rather than at the point of origin.
Are approved enterprise tools good enough so that people don’t go around them?
BCG finds 54% of employees will use unauthorized AI tools if the approved ones fall short. You’ve got a problem if your risk register describes a different technology estate than the one your people actually use.
Miss any one set and diffusion stops. I personally have reviewed programs with a world-class tool-set and an unaddressed mind-set: months of license spend, single-digit weekly active use and a workforce quietly running the same tasks through personal accounts. The pilot metrics were green, but the culture had voted no, and no one could say who would own the call at 2am.
DXC’s market frameworks arrive at the same destination from a different angle. Xponential, our AI operating model, puts governance, observability and compliance first, ahead of accelerators and automation. The Human+ approach designs AI around people, keeping humans in the loop from the outset. The four sets are a way to put that shared stance into practice in program reviews: people, process and technology, each extended for a world in which AI acts.
The boardroom instinct is that governance slows AI down. The evidence says otherwise: capability without governance can speed up individual pilots, but it cannot carry them into production. Governance is what turns capability into scaled, accountable value.
EY’s Responsible AI Pulse survey found 99% of large organizations have already taken financial losses from AI-related risk, averaging $4.4 million, and only 12% of C-suite executives could correctly identify the AI controls their own organization has in place. Deloitte finds three-quarters of companies planning agentic AI deployment within two years, while only 21% have a mature governance model for autonomous agents: decision boundaries, real-time monitoring, audit trails for the full chain of agent actions.
The WEF’s 2026 agent playbook makes the destination clear: every deployed agent carries an authorization profile setting out what it may decide alone, what decisions need human input, and who audits the trail, with oversight matched to its autonomy. Authority is defined before an incident, not during one.
Each of the ten reports reaches the same conclusion. The organizations pulling ahead are those that pass the 2am test on every agent they run, however many agents they deploy.
The 2am test has a harder version. It asks not only what happens when your AI gets something wrong, but what happens when someone else’s AI gets something right — against you.
That version became urgent this year, because near-frontier capability is no longer scarce. Speaking at the Building Societies Annual Conference in Edinburgh in April, I argued that AI resilience, not AI adoption, would become the defining board question for regulated firms. To many in the room it sounded premature. Within three weeks, the Bank of England, the Financial Conduct Authority and HM Treasury issued a joint statement warning that frontier models’ cyber capabilities already exceed what a skilled practitioner can achieve, faster, at greater scale and lower cost, and asking boards to understand and act on the risk. By August, the Bank’s Governor, writing as chair of the Financial Stability Board, was warning G20 finance ministers that frontier AI could change the speed, scale and economics of cyber risk across the whole system.
Four developments explain the shift:
The frontier jumped. In April, Anthropic’s Claude Mythos Preview and OpenAI’s GPT-5.5 made some of the largest leaps in AI cyber capability the UK AI Security Institute (AISI) has recorded. Anthropic withheld Mythos from general release and deployed it through Project Glasswing, a defensive coalition. Within weeks, roughly 50 partners had found more than 10,000 high- or critical-severity vulnerabilities in systemically important software. The bottleneck in security has moved from finding flaws to verifying and patching them as fast as AI finds them.
The open-weight gap narrowed. Open-weight models from Z.ai, Moonshot, DeepSeek and Alibaba now trail the closed frontier by about four months on general capability. On cyber, AISI measures a lag of four to seven months, down from six to ten months through most of 2025. Once weights are released, safeguards can be stripped and copies run beyond any monitoring, permanently. What today’s most restricted models can do, anyone may be able to do within two quarters. That gap is not a comfort. It is your preparation window.
The attacks are already priced in. IBM’s 2026 Cost of a Data Breach study puts the average breach at a record $4.99 million, with AI-driven attacks, led by deepfake impersonation, up 56%. Shadow AI accounted for 43% of security incidents, more than double the year before. Anthropic’s 2025 finding remains the marker: an espionage campaign in which AI agents carried out 80–90% of the operation.
Why this demands a new operating model, not a new security tool. Most AI operating models rest on three assumptions that no longer hold: that capability is scarce and held by a few vendors, that the vendor supplies safety and that risk can be reviewed on a quarterly cycle. When near-frontier capability can be downloaded, safety becomes something the enterprise must supply, and a threat that moves at machine speed cannot be governed on a committee calendar.
That is why AI resilience — the ability to anticipate, withstand, recover from and adapt to AI-related failure, yours or an adversary’s — cannot sit in a separate security workstream. It runs through all four sets:
People who assume the convincing voice on the phone, or the CFO on the video call, may be synthetic are rewarded for pausing to check.
Build workflows with rehearsed fallbacks for when a model is compromised, deprecated or simply wrong, and patch processesfor a wave, not a trickle.
Supervisors are trained to challenge AI output, and defenders who use AI to triage and fix vulnerabilities as fast as attackers exploit them.
Agents are under zero-trust controls; every model, open or closed, is managed as a supply-chain dependency with provenance checks and an exit plan; and aligns with the frameworks hardening into expectations, These are: the NIST AI Risk Management Framework, ISO/IEC 42001, UK operational resilience rules and the EU AI Act, whose high-risk obligations apply starting December 2027.
The first 2am test tells you who owns the call when your AI fails. The resilience version tells you who owns it when the failure was engineered by someone else’s AI. An organization ready for the world of AI uses AI well, and remains standing when AI is turned against it.
AI 2027, the scenario forecast from the AI Futures Project led by former OpenAI governance researcher Daniel Kokotajlo with Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, makes the same argument at civilizational scale. It maps a path to superhuman AI and posits two possible endings: one ending races ahead and loses control of a misaligned system; the other keeps control.
The capabilities involved in both endings are identical. The difference in the latter ending is a governance decision possible only because control mechanisms existed before they were needed: rollback options, monitoring, defined boundaries. The authors concede reality is running slower than their scenario, but the timeline was never the insight. The insight is structural, and it transfers exactly to the enterprise: misalignment of a model, or an agent, or a program emerges gradually under weak oversight, while the dashboards stay green. By the time the system shows you it is misbehaving, the evidence is ambiguous and the cost of stopping feels prohibitive.
Governance built after deployment explains what went wrong. Governance built before deployment keeps you in control.
Ask the sponsor of each of your three largest AI programs to complete one sentence: When this AI gets it wrong at 2am, ___ owns the call. If the sentence writes itself, scale. If it doesn’t, score the program against the four sets (mind-set, process-set, skill-set and tool-set), with resilience stress-tested through each. The review takes a week, not a quarter. Use it to identify what needs to change in your operating model, structure and culture to help existing AI pilots deliver value. This is where reinvention begins, not at the launch of another pilot.
Accountability that remains unwritten is not accountability. And an AI strategy without resilience is built for a world that no longer exists.
Hari Prasad Govindarajan is the chief AI & transformation officer for UK&I at DXC Technology, leading AI strategy, go-to-market, and delivery across the region. He has built and scaled several AI products and is a keynote speaker on practical AI diffusion using clarity methodology which about four dimensions Mindset, Toolset, Process Set, and Skillset. Connect with Hari at LinkedIn.
