Organisations keep buying artificial intelligence and failing to use it. The evidence increasingly points away from budgets and licences, and towards something harder to purchase: digital literacy, governance, and a workforce that is allowed to think.
There is a comfortable version of the artificial intelligence story in which the only thing separating a struggling organisation from a transformed one is money. Buy the licences, hire the vendor, wait for the productivity.
The evidence does not support it, and neither do the people implementing these systems for a living.
Writing in Mexico Business News on 10 August 2026, Jaime Castro Palma, General Manager at BPF, put the argument plainly: the main obstacle to making AI work is cultural, not technological or budgetary. His column is worth reading in full. This piece takes his framework, tests it against the available research, and asks a question his column does not: if the barrier is culture rather than cash, then who inside an organisation actually gets the training, and who just gets the automation?
The failure rate nobody wants to underwrite
The most cited number in enterprise AI right now comes from MIT research published in 2025 and widely reported since: roughly 95% of enterprise generative-AI pilots produced no measurable impact on profit and loss. The study drew on interviews with business leaders, a survey of employees and an analysis of public deployments.
That figure deserves handling with care, and most coverage has not handled it carefully. It measured whether a pilot delivered rapid financial return within roughly six months, and its sample leaned heavily towards sales and marketing use cases, which were also the lowest-return category in the study. “No P&L impact in six months” is not the same as “worthless”. A hospital that shortens documentation time has done something real even if it does not show up in a quarterly statement.
But strip out the overstatement and the underlying finding survives, and it is the same one Castro Palma describes from the implementation side: the projects that fail overwhelmingly fail for organisational reasons, unclear objectives, weak data foundations, no integration into how work actually gets done, no governance, rather than because the model was not good enough.
The same research found that deployments built with external partners succeeded roughly twice as often as ones built purely in-house. That is usually read as an argument for buying rather than building. It is at least as good an argument for something else: most organisations do not yet have the internal capability to run these projects, and buying a product does not create that capability.
Five things you cannot purchase
Castro Palma identifies five enablers that determine whether AI works inside an organisation. They are worth restating, because none of them appears on an invoice.
- Digital literacy. Not knowing which chatbot is fashionable, knowing what a model can and cannot be trusted to do.
- Organisational digital maturity. Whether the processes AI is being attached to are documented, measured and understood in the first place.
- Governance. Who is allowed to deploy what, on which data, with what oversight.
- Data integrity by design. Data quality built into how information is collected, not repaired afterwards by whoever drew the short straw.
- Critical thinking. The capacity, and the permission, to reject an output that is confidently wrong.
He also names three recurring mistakes: deploying with no clear objective because the technology is fashionable; automating processes that were never properly defined; and proceeding without sufficient digital literacy or critical thinking in the workforce.
The second one is the quiet killer. Automating a broken process does not fix it. It industrialises it, and removes the human friction that used to catch its worst outcomes.
Where the equity question actually lives
This is the point where the standard business analysis stops and the more important question begins.
If digital literacy is the binding constraint on whether AI helps an organisation, then the distribution of that literacy is not a training-budget detail. It determines who gains from these systems and who is simply processed by them.
In most workplaces, AI training flows towards people who already have the most autonomy: managers, analysts, knowledge workers with discretion over their own time. The people whose work is most exposed to automation, administrative staff, call-centre workers, schedulers, intake clerks, warehouse and logistics staff, and, in health systems, the front-line administrative layer that keeps patients moving, are frequently the last to be trained and the first to be affected.
Castro Palma’s own framing makes the stakes clear: AI amplifies capability, but it also amplifies error. Models fill gaps in what they know with fluent, confident invention, the failure mode usually called hallucination, and they reproduce and scale the biases already present in the data they were trained on.
Applied to sales forecasting, that is an efficiency problem. Applied to hiring shortlists, benefit eligibility, credit decisions, patient triage or risk scoring, it is a civil-rights problem wearing a dashboard. Historical data encodes historical discrimination. A system trained on who was hired, treated, approved or flagged in the past will, absent deliberate intervention, reproduce those patterns at a speed and scale no human bureaucracy could achieve, while presenting the result as neutral arithmetic.
Which is why the fifth enabler on that list is the one that matters most, and the one organisations are least likely to fund. Critical thinking is not only a skill. It is a permission structure. A worker who can identify that an automated decision is wrong, but who has no route to override it and no protection for raising it, does not functionally possess critical thinking at all. Their employer has bought the training and withheld the authority.
Accountability cannot be outsourced to a model
The one principle in Castro Palma’s argument that should not be negotiable is this: humans remain the decision-makers, and humans bear responsibility for the outcomes.
This sounds obvious and is routinely violated in practice, usually through what governance researchers describe as automation bias, the well-documented human tendency to defer to a machine recommendation, particularly under time pressure and particularly when overriding it requires justifying yourself to a supervisor.
The result is a system that is formally advisory and functionally binding. The clinician, caseworker or loan officer technically has the final say, and technically nobody has removed their judgment. But the default is the model’s output, the workload assumes the model’s output, and the burden of deviation falls on the individual.
Any organisation deploying AI in a decision that affects a person’s health, income, housing or liberty should be able to answer three questions in writing, before deployment rather than after a complaint:
- Who is accountable when this system is wrong, by name and role, not by department?
- What is the actual, used, resourced process for a person to contest an outcome?
- What is the override rate, and is it being monitored for suppression?
An organisation that cannot answer these has not deployed a tool. It has deployed a decision it does not intend to own.
What this means for smaller organisations
There is a genuinely encouraging implication in the “it’s culture, not budget” argument, and it should not be lost in the caution.
If the differentiator were capital, the outcome would already be decided: the largest players would take the entire benefit. Because the differentiator is organisational, clear objectives, documented processes, clean data, real governance, trained and empowered staff, a small clinic, a community organisation, a municipal service or a mid-sized firm can genuinely outperform a much better-funded institution that skipped all five.
For non-profits, community health providers and small employers, that suggests a specific and unglamorous sequence:
- Pick one process you can already describe end to end, and fix it before automating it.
- Define what success means in a number you can check in 90 days.
- Train the people who do the work, not only the people who supervise it.
- Write down who is accountable, and how someone appeals.
- Keep the human decision, and measure how often it is actually exercised.
None of that requires a large budget. All of it requires something more difficult: an organisation willing to look honestly at how its own work is done.
Frequently asked questions
Why do most enterprise AI projects fail?
Research and practitioner accounts converge on organisational causes rather than technical ones: no clear objective, automation of poorly defined processes, weak data foundations, absent governance, and insufficient digital literacy among staff. MIT research reported in 2025 found roughly 95% of enterprise generative-AI pilots produced no measurable profit-and-loss impact within about six months, though that metric is narrower than the headline suggests.
Is AI adoption really a budget problem?
According to Jaime Castro Palma, General Manager at BPF, writing in Mexico Business News, the main obstacle is cultural rather than technological or budgetary. The enablers he identifies, digital literacy, digital maturity, governance, data integrity by design and critical thinking, cannot be bought directly.
What is an AI hallucination?
A hallucination is when an AI system fills a gap in what it knows with fluent, confident but incorrect information. It is a structural property of how these models generate text, not an occasional glitch, which is why human verification remains necessary.
How can a small organisation adopt AI responsibly?
Start with one well-understood process, define measurable success, train front-line staff rather than only managers, document who is accountable and how decisions can be contested, and monitor how often humans actually override the system.
Sources and method
This article is based on the expert column “It’s Not Budget: Digital Culture Makes AI Work” by Jaime Castro Palma, General Manager at BPF, published by Mexico Business News on 10 August 2026 and available here. The five enablers, the three common mistakes and the principle of human accountability are his; the framing, the equity analysis and the recommendations are Raw POV’s. The 95% pilot-failure figure comes from MIT research published in 2025 and widely reported since, and is presented here with its methodological limits stated. Raw POV did not interview Mr Castro Palma for this article and has no commercial relationship with BPF or any AI vendor.
Read more from Raw POV: our Society coverage and how big data is reshaping football’s most beautiful moments.
