Stop Paving Cow Paths - Why Process Analysis and Optimization Should Precede AI Implementation
Article created on 7/28/2026
Few boardrooms nowadays are free of the question "what is our AI strategy?". The pressure to act is real, the FOMO is high: competitors are announcing pilots and making headlines with AI, vendors are promising transformation, and nobody wants to be the company that missed the wave. Under that pressure, a familiar pattern emerges - organizations treat implementing AI as the project itself: select a tool first, and only afterwards go looking for a place to apply it.
The evidence - from economics, information systems research, and a growing body of post-mortems on failed AI initiatives - points in the opposite direction. The organizations that capture value from artificial intelligence are not the ones that adopt it fastest, but the ones that prepare the ground for it: rethinking their business, mapping, analyzing, and optimizing their business processes before a single model goes into production. A maxim often attributed to Bill Gates captures the underlying logic: "(...) automation applied to an efficient operation magnifies the efficiency, while automation applied to an inefficient operation magnifies the inefficiency. The attribution may be folklore; the mechanism, as we will see, is well documented.
This article makes the case - with the receipts - for a sequence that disciplined operators have always followed: analyze first, optimize second, automate third.
The Uncomfortable Arithmetic of Enterprise AI
Start with the outcomes. A 2024 RAND Corporation study based on interviews with experienced data scientists and engineers estimated that more than 80% of AI projects fail - roughly twice the failure rate of IT projects that do not involve AI [1]. A year later, MIT’s Project NANDA reported an even starker figure for the generative AI era: despite tens of billions of dollars in enterprise investment, about 95% of GenAI pilots were producing no measurable impact on profit and loss, with only around 5% of custom enterprise tools ever reaching production at scale [2].
Survey data tells the same story from a different angle. S&P Global Market Intelligence found that the share of companies abandoning most of their AI initiatives jumped to 42% in 2025, up from 17% a year earlier, with an average of 46% of proofs of concept scrapped before ever reaching production [3]. Gartner had predicted precisely this shake-out, forecasting that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025 [4]. And none of this is unique to generative AI: back in 2019, before ChatGPT existed, a global survey by MIT Sloan Management Review and Boston Consulting Group found that seven out of ten companies reported minimal or no impact from their AI investments to date [5].
The important finding, however, is not the failure rate - it is the failure causes. When RAND’s researchers categorized the root causes, the leading ones had almost nothing to do with algorithms: misalignment between business leaders and technical teams about what problem the AI was meant to solve; lack of the data needed to train adequate models; a focus on using the latest technology rather than solving real user problems; and inadequate infrastructure and data governance [1]. MIT’s report reached a complementary conclusion: pilots fail because tools do not fit into actual workflows and do not learn from the context of the work - a “learning gap” rooted in operations, not in model quality [2].
In other words: enterprise AI mostly fails on problem selection, data, and process integration. All three are exactly what process analysis exists to get right.
We Have Seen This Movie Before
None of this should surprise anyone who remembers the last time a general-purpose technology swept through the enterprise. In 1990, after a decade of heavy corporate IT spending had delivered disappointing productivity results, Michael Hammer diagnosed the cause in one of the most influential management articles ever published: companies were using technology to mechanize their old ways of doing business, leaving inefficient processes intact and letting computers speed them up. “It is time to stop paving the cow paths” he wrote - instead of embedding outdated processes in silicon and software, organizations should first redesign the work itself [6].
The economists then quantified what Hammer had intuited. Studying firm-level data through the 1990s, Erik Brynjolfsson and Lorin Hitt showed that the returns to IT investment were highly uneven, and that the difference was explained by complementary organizational investments: firms that redesigned processes, decentralized decisions, and invested in skills captured large productivity gains from the same technology that did little for firms that simply installed it [7]. Two decades later, the same research tradition formalized this as the “productivity J-curve”: general-purpose technologies - electricity, computing, and now AI - deliver their benefits only after a lag, because firms must first build intangible complements such as new processes, new skills, and reorganized workflows. Skip the complements and the technology underperforms indefinitely [8].
AI is the sequel to this movie, with larger budgets and faster hype cycles. The technology has changed; the economics of complementarity have not.
This is why Mourinho Solutions is the right partner to guide you on the path to unlocking efficiency in your organization and maximizing the return on your AI investment. We defend that business comes first - technology (and AI) should follow. And there are 5 good reasons for that.
Reason 1: Automation Amplifies Whatever Process It Touches
An algorithm deployed on a process inherits that process’s logic - including its pathologies. This was Hammer’s original warning about IT [6], and the research community issued precisely the same caution when robotic process automation (RPA) boomed in the late 2010s. Because RPA robots replicate the way work is currently done - clicking through the same screens, following the same rules - automating without first examining the process risks pouring digital concrete over inefficiency, making a bad process faster, more rigid, and harder to change [9]. A systematic review of the RPA literature likewise identified the selection and suitability of processes - rule-based, standardized, mature - as one of the central determinants of success, and poor process selection as a recurring cause of failed initiatives [10].
The lesson transfers directly to AI, and arguably with higher stakes. A language model drafting responses inside a broken approval chain simply fills the bottleneck’s queue faster. An ML model predicting demand for a planning process nobody follows produces predictions nobody uses. Amplification is the whole point of automation - which is why the thing being amplified deserves scrutiny first.
Reason 2: Data Quality is a Process Outcome - and It Is Worse Than You Think
Every AI system is downstream of data, and data is downstream of processes: it is created, keyed in, transformed, and handed off by the everyday operations of the business. The state of that data is sobering. In a study in which managers assessed samples of their own organizations’ records, only 3% of companies’ data met basic quality standards, and nearly half of newly created records contained at least one critical, work-impacting error [11]. RAND’s interviewees ranked the lack of adequate, usable data among the top causes of AI project failure, noting that organizations chronically underestimate the data preparation the technology requires [1].
The consequences of ignoring this are well documented at the practitioner level. A study of AI practitioners across high-stakes domains found that 92% had experienced “data cascades” - compounding, downstream failures triggered by upstream data quality issues, typically invisible until they surface late as broken models and abandoned deployments [12]. The study’s title alone is a diagnosis of the industry’s sequencing problem: “Everyone wants to do the model work, not the data work.”
Here is the strategic point: you cannot fix data quality with a data-cleaning sprint, because bad data is continuously re-produced by the processes that generate it. Process analysis locates where data is born, duplicated, re-keyed, and corrupted; process optimization fixes the defect at the source. An organization that optimizes its processes first is, quietly and simultaneously, building the data foundation its future AI will stand on.
Reason 3: Without Process Visibility, You Will Pick the Wrong Use Cases
Deciding where AI belongs is a portfolio decision, and it cannot be made sensibly without knowing how work actually flows. Davenport and Ronanki [13], surveying 152 cognitive technology projects, found that ambitious “moonshot” deployments routinely disappointed, while the successful adopters took an incremental approach: examining their processes task by task, and matching specific AI capabilities to specific bottlenecks in specific workflows.
That task-level visibility is exactly what the business process management (BPM) discipline provides and one of our core competences. The standard BPM lifecycle runs through process identification, discovery, analysis, and redesign before implementation and automation - a sequencing distilled from decades of practice in which automating an unexamined process is a category error [14]. Modern process mining strengthens this further: by reconstructing processes from the event logs of existing IT systems, it replaces opinion-based process maps with evidence - exposing the real number of process variants, the rework loops, the deviations, and the bottlenecks that interviews alone never reveal [15]. It is, in effect, an X-ray taken before surgery - and it frequently shows that the process executives describe is not the process the organization actually runs.
The classic automation success stories were built on precisely this selection discipline. At Telefónica O2, one of the earliest and most-cited RPA deployments - ultimately yielding a return on investment in the hundreds of percent - automation targets were chosen deliberately: high-volume, standardized, rule-based processes, stabilized before software robots were applied to them [16]. The success was engineered in the selection, not the software.
Reason 4: Sometimes the Best “AI Project” is Deleting the Step
Process analysis has an inconvenient habit: it keeps finding problems that do not need AI at all. A duplicated approval, a report nobody reads, a handoff that exists because of an org chart from 2014, data re-entered manually between two systems that could simply be integrated - classic industrial-engineering doctrine says eliminate, combine, and simplify before you automate, because automating waste only produces faster waste. This was the core of Hammer’s “don’t automate, obliterate” argument: much of what technology was being asked to speed up should not have existed in the first place [6].
On a personal note, I am quite fond of the motto often attributed to Leonardo da Vinci — "simplicity is the ultimate sophistication." Making a process simple is demanding work: it requires strong analytical competencies, the ability to relate inputs from different organizational areas, and the integration of knowledge from different fields. It is precisely the craft we have built Mourinho Solutions around.
For an organization about to fund an AI program, this is not a detour; it is free money and risk reduction. Steps that are eliminated cost nothing to automate, require no model maintenance, create no new failure modes, and start paying back immediately. And the process work yields a second dividend: a measured baseline. One of the defining findings of the MIT NANDA report was that most GenAI pilots could show no measurable P&L impact [2] - but impact can only be measured against a known starting point. An optimized process comes with documented cycle times, error rates, and costs per transaction; the subsequent AI deployment inherits, for free, the ability to prove its own ROI.
Reason 5: AI Value is Realized through Organizational Change, Not Installed With Software
Finally, the research on where AI value actually comes from converges on an uncomfortable truth for tool-first strategies: the value materializes at the level of business processes, and only when the organization changes around the technology. A systematic literature review on AI and business value found that first-order effects appear as process efficiency and improved decisions within processes - and that capturing them depends on organizational enablers and deliberate business process transformation, not on the sophistication of the model [17]. Interview studies of AI adopters similarly identify organizational readiness - strategic alignment, data quality and governance, process know-how, and culture - as the precondition that should be assessed before adoption decisions, not discovered after them [18].
Practitioner research agrees. McKinsey leaders writing in Harvard Business Review reported that the biggest obstacles to scaling AI are not technical but cultural and organizational, and that success requires redesigning workflows and running AI initiatives as business-led process change, with interdisciplinary teams - not as IT installations [19]. And in the MIT Sloan/BCG global study, the companies that actually captured value with AI were distinguished by the way they combined strategy, organizational change, and technology, rather than treating the technology as the strategy [5].
Process analysis and optimization is precisely this organizational work, done in the right order. It aligns leadership on which problems matter (attacking RAND’s number-one failure cause [1]), builds the process understanding that workflow redesign requires, and creates the stable, standardized operations onto which AI can be deployed with confidence.
Mourinho Solutions brings you the “Process First"
None of the previous reasons argues for years of analysis before anyone is allowed to touch a model. It argues for a short, disciplined sequence - essentially the front half of the BPM lifecycle [14] - executed in weeks, not quarters, and scaled to the ambition of the AI program it precedes.
Mourinho Solutions has a bespoke consultancy approach, based on state-of-the-art knowledge and years of experience, that takes your organization through the entire BPM lifecycle methodically, so that you get the greatest return on your investment.
There is also a compounding effect worth naming. The organizations in the successful minority - MIT’s 5%, the “pioneers” of the MIT Sloan/BCG studies - did not merely avoid failure; they built the organizational muscle that makes each subsequent AI deployment cheaper and faster than the last [5], [2]. Process clarity, clean data flows, and measurement discipline are infrastructure. They appreciate.
Conclusion: AI-Ready Is Process-Ready
The rush to implement AI treats process analysis as a delay. The evidence says it is the opposite: it is the highest-leverage risk reduction available, attacking the documented root causes of AI failure - wrong problem, weak data, unintegrated workflows - before they can compound into an abandoned pilot. Paving cow paths will lead to virtually no value creation - and it can actually decrease it. Thirty-five years of research, from business process reengineering [6] through the economics of complementarity [7], [8] to the current generation of AI post-mortems [1], [2], tells one consistent story: technology multiplies the organization it lands in. Optimize the organization first, and AI multiplies something worth multiplying.
The question for leadership teams, then, is not “how fast can we implement AI?” but “which of our processes, once understood and optimized, would AI genuinely transform?” Answering it in that order is what separates the 5% from the 95%.
Mourinho Solutions makes sure your organization belongs to the 5%.
At Mourinho Solutions, this sequence is how we work. We help organizations map and analyze their processes, optimize them for measurable gains, and only then design and implement AI where it demonstrably pays off. If your company is planning an AI initiative, start with a conversation about your processes - contact us for a process assessment.
References
[1] J. Ryseff, B. F. De Bruhl, and S. J. Newberry, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed: Avoiding the Anti-Patterns of AI. Santa Monica, CA, USA: RAND Corporation, 2024. [Online]. Available: https:/
[2] A. Challapally, C. Pease, R. Raskar, and P. Chari, “The GenAI divide: State of AI in business 2025,” MIT NANDA, Cambridge, MA, USA, 2025. [Online]. Available: https:/
[3] S&P Global Market Intelligence, “Voice of the Enterprise: AI & machine learning” survey, 2025; findings reported in “AI project failure rates are on the rise: Report,” CIO Dive, Mar. 2025. [Online]. Available: https:/
[4] Gartner, “Gartner predicts 30% of generative AI projects will be abandoned after proof of concept by end of 2025,” Press release, Jul. 29, 2024. [Online]. Available: https:/
[5] S. Ransbotham, S. Khodabandeh, R. Fehling, B. LaFountain, and D. Kiron, “Winning with AI,” MIT Sloan Management Review and Boston Consulting Group, Cambridge, MA, USA, Res. Rep., 2019. [Online]. Available: https:/
[6] M. Hammer, “Reengineering work: Don’t automate, obliterate,” Harvard Business Review, vol. 68, no. 4, pp. 104–112, Jul.–Aug. 1990. [Online]. Available: https:/
[7] E. Brynjolfsson and L. M. Hitt, “Beyond computation: Information technology, organizational transformation and business performance,” Journal of Economic Perspectives, vol. 14, no. 4, pp. 23–48, 2000, doi: 10.
[8] E. Brynjolfsson, D. Rock, and C. Syverson, “The productivity J-curve: How intangibles complement general purpose technologies,” American Economic Journal: Macroeconomics, vol. 13, no. 1, pp. 333–372, Jan. 2021, doi: 10.
[9] W. M. P. van der Aalst, M. Bichler, and A. Heinzl, “Robotic process automation,” Business & Information Systems Engineering, vol. 60, no. 4, pp. 269–272, 2018, doi: 10.
[10] R. Syed et al., “Robotic process automation: Contemporary themes and challenges,” Computers in Industry, vol. 115, Art. no. 103162, Feb. 2020, doi: 10.
[11] T. Nagle, T. C. Redman, and D. Sammon, “Only 3% of companies’ data meets basic quality standards,” Harvard Business Review, Sep. 11, 2017. [Online]. Available: https:/
[12] N. Sambasivan, S. Kapania, H. Highfill, D. Akrong, P. Paritosh, and L. Aroyo, “‘Everyone wants to do the model work, not the data work’: Data cascades in high-stakes AI,” in Proc. 2021 CHI Conference on Human Factors in Computing Systems (CHI ’21), Yokohama, Japan, 2021, Art. no. 39, doi: 10.
[13] T. H. Davenport and R. Ronanki, “Artificial intelligence for the real world,” Harvard Business Review, vol. 96, no. 1, pp. 108–116, Jan.–Feb. 2018. [Online]. Available: https:/
[14] M. Dumas, M. La Rosa, J. Mendling, and H. A. Reijers, Fundamentals of Business Process Management, 2nd ed. Berlin, Germany: Springer, 2018, doi: 10.
[15] W. M. P. van der Aalst, Process Mining: Data Science in Action, 2nd ed. Berlin, Germany: Springer, 2016, doi: 10.
[16] M. C. Lacity and L. P. Willcocks, “Robotic process automation at Telefónica O2,” MIS Quarterly Executive, vol. 15, no. 1, pp. 21–35, Mar. 2016. [Online]. Available: https:/
[17] I. M. Enholm, E. Papagiannidis, P. Mikalef, and J. Krogstie, “Artificial intelligence and business value: A literature review,” Information Systems Frontiers, vol. 24, no. 5, pp. 1709–1734, 2022, doi: 10.
[18] J. Jöhnk, M. Weißert, and K. Wyrtki, “Ready or not, AI comes — An interview study of organizational AI readiness factors,” Business & Information Systems Engineering, vol. 63, no. 1, pp. 5–20, 2021, doi: 10.
[19] T. Fountaine, B. McCarthy, and T. Saleh, “Building the AI-powered organization,” Harvard Business Review, vol. 97, no. 4, pp. 62–73, Jul.–Aug. 2019. [Online]. Available: https:/