95% of AI Investments Generate Zero ROI
But also, 74% deliver positive ROI. Two reports, contradictory findings. What can we learn?
In the past week, two reports on AI adoption seem to come to strikingly different conclusions. The first, which received a lot of coverage in my corner of X/Twitter at least (relentlessly bullish on AI), was the Wharton Business School 2025 report Accountable Acceleration: GenAI Fast-Tracks into the Enterprise. The second, which seems to me to have had more coverage on LinkedIn (generally more bearish on AI), was an MIT NANDA[1] paper The GenAI Divide: State of AI in Business 2025.
The headlines are contradictory. Where Wharton report 74% of investments in AI deliver positive return on investment (ROI) already, MIT report 95% of organisations are getting zero return on their AI investments.
Because the Wharton report says things that are mostly in our interest at Cassi to promote, I offer only the shortest of summaries. I was more interested in what we can learn from the more negative report, hopefully ensuring we (and you, now I have decided to make what was to be an internal memo a blog post) avoid mistakes others have made.
Wharton’s report tells us that:
“Accountability is now the lens” - structured, ROI-linked metrics e.g. profitability, throughput, workforce productivity - measurable outcomes, not just adoption.
“Long-term optimism in Gen AI is increasingly strong” 88% expect increased spend on AI in next 12-months, +16pp YoY; growing trend of cutting to fund: 11% (+7pp YoY) reallocating funds, mostly from legacy IT + HR and Workforce Programs.
74% see positive RoI already from AI investment; 80% expect it within 2-3 years; larger enterprises are less successful despite bigger budgets: 34% report ’too early/neutral outcomes); Digital enterprises most successful, 88% positive; banking and professional services next 83%. Efficiency and productivity the top benefit, then quality, creativity and security.
Human factors key. Deliberate change management needed to produce RoI; overcoming bottlenecks - human capital, talent, training - essential to success. [I’d add incentives but the report doesn’t say this!]
The report also calls for Chief AI Officers, which is something I would want to reflect on further, and perhaps see more evidence of, before endorsing – I worry appointments like this can be performative, and also that it creates new internal rivalries that might not be conducive to adoption e.g. the Chief AI Officer wants to own and drive AI transformation for their own career success, but their rivals around the board table for bonuses and promotion have an incentive to either (a) own their own AI initiatives and outperform the CAIO or (b) thwart or limit the success of the CAIO’s initiatives – especially if they compete for resources from within their domain. Still, Wharton says the evidence/testimony of many Execs shows they are a good idea.
If you only read the Wharton report, you might conclude ‘more AI spending = higher ROI’ this would be wrong, as both the MIT and Wharton report conclude.
Both agree that accountability, assessment vs operational objectives, and measurement vs operational outcomes, are the key to success. Funding more pilots, or scoring ‘adoption’ alone are unlikely to generate ROI.
But the MIT report gets into the reasons for failure more deeply.
The headline finding in MIT’s report is probably the most important thing for those funding, or potentially funding, AI investments to know. 95% of those investments fail, as assessed by MIT across ‘300 public implementations’, on which MIT conducted interviews, surveys and their own analysis.
But those same decision-makers should be more curious – the headline could mislead, failing to read deeper means ceding a chance to learn from others’ mistakes. The report also finds that “5% are extracting millions in value” and offers a diagnosis of what works and why. If 95% of organisations use this report to stop or reduce AI investment, as their AI sceptics are emboldened, they are increasing the risk their business is eaten or defeated by the 5% that get their investment right.
There are other bearish indicators in the report, for example, so far:
Tech & Media are the only two of nine industries seeing significant disruption.
No sign of the needed deep structural shifts associated with past general purpose technologies e.g. new market leaders, disrupted business models, or measurable changes in customer behaviour. Substitution, pilots, experimentation - not transformation - rule
Why have 95% or AI Investments failed to Generate ROI?
The report says only 5% of custom enterprise AI tools reach production. We suggest the Valley of Death, and innovation theatre are likely major causes – more so than product or pilot failure.
There is a lot of incentive for organisations, and the individuals within them, to experiment with AI. It is career enhancing to show yourself to be forward thinking, and have some stories for your boss, board and customers. But the unpopularity of disruption, which always creates winners and losers is likely a major factor in killing off so many AI investments at the pilot stage. In contrast, showing the courage to ‘fail fast’ and kill of your own initiatives, or even more so those of others, can be career enhancing. The incentives in many enterprises are pretty terrible. As the report says, the fact that only 5% of custom enterprise tools make it from pilot to production ‘explains why most organisations remain on the wrong side of the divide.’
Perhaps supporting our analysis, the report notes that larger enterprises launch more pilots but have the lowest pilot to production ratios. Larger organisations are harder to align incentives within, have more competing fiefdoms, and often require much larger and more disruptive change management programmes since there are so many more interconnected and interdependent workflows.
Another reason for limited ROI is that 40% of the 5% of GenAI tools that make it from pilot to production are chatbots, and the inability of chatbots to learn means they add little value, and cause limited disruption. Like the pilots themselves, they sit on the periphery of staff workflows.
Pilots fail for more predictable reasons - the unwillingness of users to adopt new tools, and concerns around trusting model output.
Enterprise CIOs should also consider the more neutral findings that to date, GenAI adoption has primarily enhanced individual productivity, not P&L performance. Why might that be?
Secret Cyborgs. One reason for this is likely that employees are not passing on their gains in productivity to the business, or at least not in measurable ways. We suggest that in most organisations, there is no shortage of work, or at least ‘make-work’ in the form of meetings and coordination costs and emails to reply to - many, perhaps most, are overwhelmed, not under-employed. Employees are likely reinvesting time saved in things they might not otherwise have got to – but of course often that just generates more internal work for everyone else.
Furthermore, there is a significant shadow AI economy – comprised of those Ethan Mollick at Wharton called ‘secret cyborgs’ – that is thriving: 90% of employees use GenAI for work - many using external models. We suggest a primary reason for this is model-lag: by the time enterprises adopt a model, it is usually out of date, and far better models are available via personal subscription to ChatGPT, Gemini, Claude. Second, Co-Pilot, in particular, is just not as good as external tools (in my view), yet seems to be what most are using at work and basing their opinions of GenAI on. Supporting this, the report notes a generalised distrust of internal tools even from avid GenAI users.
MIT note that forward thinking organisations turn Secret Cyborgs into advantage - they don’t punish but learn from shadow usage, surveying, understanding and adapting by procuring enterprise alternatives.
Speed & Crossing the Valley of Death. Success in AI investments is more frequent in mid-market companies, which generally invest less, and have fewer pilots, but are much faster, running at an average of 90-days from pilot to production, and take a greater proportion of their pilots through to production
In contrast enterprises take 9-months or greater. Nine months vs current rates of AI progress is far too slow, unless the tool is being continuously updated and upgraded as new models drop. If staff find your tools inadequate and have to hide what they are doing, it is unlikely you’d see gains on your P&L from their use of AI – their incentive is to hide the gains, along with the use.
What can we learn that is more positive from the MIT report?
Reassurance?
AI is not (yet) coming for your job. Notwithstanding all we have written on this blog on the likely impact of AGI on employment, MIT find that for now, AI adoption has not led to workforce reduction. Displacement, per a wider MIT analysis cited in the report, is likely, but will be graduated through discrete displacement vice huge lay-offs. If we get to AGI on current timelines, I don’t think this is correct, but still, for now, we can all take some reassurance from the data.
How Buyers that Achieve ROI Succeed
Strategic Partnerships with External Vendors Double the Chance of Success.
“External partnerships see twice the success rate of internal builds” Employee usage rates are also twice as high for externally built tools vs those built internally. This success rate is particularly notable since internal build dominates within the sample case studies, but is rarely successful unless in partnership.
Buyers who succeed demand process-specific customisation and evaluate tools on business outcomes rather than software benchmarks. They expect systems that integrate with existing processes and improve over time.
The most effective AI-buying organisations don’t wait for perfect use cases or central approval. They drive adoption through:
1. Distributed Experimentation.
2. Vendor Partnerships / Strategic Partnerships (x2 likelihood of success) – “procure external tools, co-develop with vendors”.
3. Benchmarked tools on operational outcomes.
4. Partnered through early stage failures.
5. Sourced initiatives from frontline managers (vice top-down or central lab generated) – but paired bottom up sourcing with senior exec accountability. This likely accelerates adoption through operational fit and career incentives for adoption.
6. Decentralised authority and clear ownership.
7. Clear Accountability.
8. Harnessing power users to the cause.
9. Likewise individuals and managers on the customer side were critical to success.
Highest ROI for now comes from substitution - from eliminating previous business process outsourcing, notably in document processing costs, and ~30% reducing external agency use and consulting spend. ROI was highest in back-office functions. Financial services benefited cutting ~$1M on previous out-sourced risk management. Successful buyers also saw improved customer retention and sales conversion, through things like automated outreach and follow-up. That ROI did not come from AI transformation, is perhaps because very few organisations seem to be undertaking it yet.
How winning vendors behave
Trust in the vendor is the most important factor for executives awarding contracts.
Vendor trust isn’t always good news for start-ups like ours – many enterprises prefer to wait for existing vendors to offer the new tools than take a risk on evaluating ‘emerging’ vendors. Plus, the report notes, existing Business Process Outsourcing suppliers have the advantage of already understanding customer workflows. Consequently:
“Product quality alone is rarely sufficient. Referrals, prior relationships, and VC introductions remain stronger predictors of enterprise adoption than functionality or feature set.”
Vendor trust is the key priority, but deep understanding of customer workflows is the second highest, requiring domain expertise and a willingness to learn.
Relatedly, a third key priority is that AI products succeed when they integrate with minimal disruption to current tools following (people don’t want to learn a new UI/tool, no matter how good). Vendors succeed because they customise deeply and embed themselves in workflows, adapt to context, and scale from narrow but high-value footholds.
Tools that succeed therefore have low configuration burden and immediate, visible value. Tools requiring extensive input tended to stall at the pilot phase. Tools with complex internal logic, opaque decision support, or that are ‘optimised based on proprietary heuristics’ tended to fail.
Agentic AI adoption is very nascent. Too early to assess. The report argues this could lead to a more fundamental shift and much higher ROI. But also, MIT NANDA is a community for those building AI Agents, so while I think they are partly right (there will be much higher ROI for GenAI too, Agentic implementation a complement, and the revenues from each as inextricable as the two technologies), we should allow for some bias in their judgement.
So What?
Both reports are right in their own way. MIT’s exclusively looks back whereas Wharton’s is more forward looking (asking ‘what do you expect?’, as well as ‘what has happened?’). They agree on what is needed for investments to succeed. They agree that ROI is nascent, and organisations are moving from pilot to production, from adoption metrics to proving value vs operational outcomes.
In the end, they measure the same things in different ways. Wharton captures self-report data, likely sweeping up gains the MIT report misses, and allowing for promising minimally ROI-positive implementations that are ‘felt’ to have delivered internal ROI that doesn’t necessarily show-up on the top or bottom line, and/or that are expected to grow, or are growing, to be weighted for future impact. MIT captures verified transformation and financial impact and is the more rigorous. Both are right within their frames—both track experience, one weights expectations more heavily, the other current execution.
Both offer caution - and useful lessons.
Neither really address the medium to longer term on AI that this blog usually focuses on - the forecasts on the arrival of AGI/ASI being most likely, dependent on the preferred definition, 2026-2033. But for leaders betting on AI, they are useful guides as to how to get your organisation AGI ready - how and where to start your transformation.
If your intent is to optimise for success applying the latest in decision science to optimally allocate resources, minimise risk and maximise opportunities, you need us. Find us at www.cassi-ai.com
[1] NANDA: Networked AI Agents in Decentralized Architecture – an MIT project, a mix of technical protocols and community of interest https://www.linkedin.com/pulse/nanda-internet-ai-agents-ramesh-raskar-211ve/ seeking to “…explor[e] how artificial intelligence can evolve into a truly democratic and distributed ecosystem.” By “…address[ing] critical architectural components such as privacy, incentives, orchestration, or user accessibility… …designing a framework where billions of agents can collaborate autonomously while preserving privacy, scalability and innovation at the edges.” It is perhaps unsurprising the report concludes, in part, that the real ROI will come from the deployment of Agents – an assertion I think is partly true, probably sincerely held, but clearly a conclusion NANDA was always likely to arrive at.