AI tourists or AI operators? Why enterprises are struggling to prove AI ROI

Artificial intelligence is often sold with the promise of speed, efficiency and transformation. But inside enterprises, the first encounter with AI is not always magical. It can be uncomfortable. It can expose weak processes, poor data ownership, unclear workflows, brittle systems and decisions that were never formally documented. In this interaction with Dataquest, Sanchit Vir Gogia, Founder and Chief Analyst, Greyhound Research , says that AI should not be seen merely as a technology upgrade. It is beginning to change how work is performed, supervised, priced, trusted, audited and governed. That makes the question of ROI far more complex than licence cost, productivity gains or pilot success. For Sanchit, serious enterprises are now moving beyond the excitement of pilots. They are asking harder questions about ownership, governance, integration, accountability and measurable business outcomes. The distinction, he says, is becoming clear: some organisations remain AI tourists, while others are slowly becoming AI operators. You describe AI as a mirror. What is it really showing enterprises? AI is proving to be a very effective detector of organisational theatre. It finds the process that exists in the policy document but not in practice. It finds the data owner who owns very little in reality. It finds customer journeys that depend on multiple handoffs and no clear accountability. It also finds productivity metrics that measure activity rather than outcomes. That is why the mirror is not always flattering. Many enterprises expected AI to simplify work immediately. Instead, it is first showing them the complexity that was already there. It is surfacing undocumented exceptions, fragmented systems, inconsistent policies, weak data lineage, informal approvals, brittle handoffs and human judgement that was being treated as process. This is uncomfortable, but it is also useful. AI behaves less like a productivity wand and more like a diagnostic instrument. It tells the enterprise where the work is actually broken. Is this one reason why CIOs and CFOs are not always aligned on AI investments? The CIO-CFO tension around AI is completely rational. The CIO sees AI as a capability build. The CFO sees rising spend, unclear attribution, soft returns and pilots that sometimes continue without moving into full-scale deployment. The CIO is often funding future execution capacity. The CFO is waiting for evidence that the future has actually paid rent. Both are right. The disagreement is usually about timing, measurement and ownership. Greyhound Research CIO Pulse 2025 shows that CIOs are prioritising customer support and IT operations because these areas offer more measurable use cases and faster time to value. Our finance-sector advisory work shows a similar pattern with CFOs. Financial exception handling is attractive as a use case, but pilots often stall when integration with ERP systems, policy rules and exception workflows becomes difficult. The economics may exist, but the operating design is often not ready. Should AI ROI be assessed like a standard technology upgrade? No. That is the wrong frame. A standard technology upgrade improves an existing system. AI begins to change how work is performed, supervised, priced, trusted, audited and governed. A conventional upgrade has familiar economics. Replace the system, consolidate licences, reduce maintenance, improve uptime, remove some manual steps and book the saving. AI does not work so neatly. AI sits across data, workflow, decision-making, compliance, infrastructure and human judgement. It consumes compute continuously. It introduces probabilistic output into processes that were built around deterministic assumptions. It also creates new responsibilities around validation, monitoring, model evaluation, escalation, auditability and failure recovery. That is why AI is not just a software refresh. It is becoming a governed execution layer, and that has financial consequences. What are enterprises underestimating when they calculate AI ROI? Technical debt is one of the first warning signs. AI can produce code, content, process maps, test scripts, decision notes, policy summaries, operational recommendations and customer responses at extraordinary speed. But speed is not the same as system health. AI-generated code, for instance, can increase output while worsening maintainability, review burden, reliability risk, test coverage, security exposure and hidden dependency issues. Engineering may look faster in the short term, while the repair bill moves downstream. That is not a productivity miracle. It is technical debt appearing in a more confident form. The same applies to scattered pilots, long payback periods, old pricing models, vendor heterogeneity and unclear accountability. These are not side issues. They are central to the ROI story. Where are enterprises going wrong in AI buying? I see a version of the Gruen effect playing out in enterprise AI buying. In simple terms, the Gruen effect refers to what happens when shoppers enter a store with one clear purpose but get distracted by the design, choices and displays around them. In AI, something similar is happening. Organisations enter the market with a sensible business question and leave surrounded by too many tools, platforms and promises. There are copilots, agents, model catalogues, orchestration layers, vector databases, governance tools, prompt libraries, inference options, private AI claims, sovereign AI pitches and pricing calculators. Choice is useful, but uncontrolled choice creates confusion. The original problem can get buried. A use case becomes a platform discussion. A platform discussion becomes an architecture debate. An architecture debate becomes a governance workshop. Somewhere in the middle of all this, the customer still wants the invoice processed, the claim resolved, the code reviewed or the service ticket closed. This is where enterprises need discipline. The question should not be, “Where can we use AI?” The better question is, “Which business system deserves intelligence, and what economic result must it produce?” Are enterprises now moving from AI hype to AI fatigue? AI fatigue is real, but it is not always bad. In some cases, it is the first sign that an organisation has moved from fantasy to friction. That friction forces better questions. Which workflow matters? Which outcome is measurable? Which risk is acceptable? Which model is sufficient? Which vendor is accountable? Which control is non-negotiable? Which pilot should be closed? These are not pessimistic questions. They are mature questions. The better-run enterprises are becoming more sober after their first contact with reality. They are narrowing use cases, strengthening governance, redesigning workflows, assigning business owners, building cost telemetry, classifying workloads and measuring outcomes rather than tool usage. That is a healthy shift. It means the organisation is moving from announcement mode to operating discipline. You make a distinction between AI tourists and AI operators. What separates the two? The split between AI tourists and AI operators is becoming more visible. AI tourists are not necessarily unserious. Many are intelligent, well-funded and supported by capable vendors. But they still treat AI as something to be adopted. AI operators treat it as something to be governed, measured, placed, priced and woven into the economics of work. The tourist organisation asks, “Where can we use AI?” The operator asks, “Which business system deserves intelligence, and what economic result must it produce?” The first question produces pilots. The second produces architecture and operating discipline. Experimentation is useful at the beginning. But once real money, real customers, real regulators and real employees are involved, experimentation is not enough. AI has to move into governed execution. Why is AI ROI difficult to calculate in isolation? AI ROI cannot be calculated properly by looking at the model alone. A model may perform well in isolation and still produce poor returns inside a weak enterprise system. I would explain this through a simple farm analogy. Looking at AI ROI only through the model is like studying a single grain and assuming you understand the whole crop. But enterprise value depends on the larger field around it. The yield changes with the quality of the soil, the weather, irrigation, labour, machinery, storage, timing and market conditions. AI works in a similar way. The model matters, but it is only one part of the field. Its value depends on data quality, workflow design, adoption behaviour, integration depth, compliance controls, human escalation, infrastructure placement, vendor contracts, regulatory exposure, model routing, security posture, business ownership and financial discipline. Remove any one of these and the outcome changes. Sometimes it collapses. That is why AI ROI is not simple project accounting. Some use cases create direct cost savings. Some improve revenue conversion. Some reduce compliance exposure. Some compress cycle time. Some improve customer experience. Some reduce fraud. Some improve decision quality. Some improve working capital. Some build strategic capability without an immediate profit-and-loss impact. The narrow ROI model fails because it isolates the AI deployment from its environment. The stronger model asks about the system around it. What data feeds the AI? Which workflow absorbs the output? Which human verifies it? Which policy constrains it? Which system does it touch? Which cost centre pays for it? Which business owner owns the outcome? Which regulator can question it? Which customer is affected? Which vendor is accountable when the result fails? Those questions are not decorative. They are the ROI model.