The Spectrum of Trust (And the Questions That Were Never Asked)
A major enterprise software company recently published the results of its annual mainframe survey. The data is substantial, the sample credible, and the findings genuinely interesting. It deserves serious attention — not because of what it says, but because of what it assumes.
The headline finding: nearly one in three mainframe professionals say they trust artificial intelligence to complete operational tasks autonomously. Not just advise. Not just alert. Complete.
For anyone who has spent decades in mainframe environments — where a misplaced comma in JCL can take down a national payment system — that number demands examination. Not dismissal. Examination.
The Data Is Real. The Question Is What It Means.
The survey asked hundreds of mainframe professionals about their willingness to trust AI across three operational areas: checkpoint and commit pacing, problem diagnosis, and backup decisions. For each, respondents chose from four levels: no trust at all, alert only, recommend an action, or complete the action autonomously.
The results followed a remarkably consistent distribution. Roughly 9 percent said no trust at all. About 25 percent were comfortable with alerts. Around 37 percent accepted recommendations. And 28 to 32 percent — depending on the task — said they would let AI complete the action independently.
These numbers are worth taking seriously. They suggest a genuine shift in how mainframe professionals think about automation. But they also raise questions that the survey, by its nature, cannot answer.
The most important one is this: what does "trust" mean when it's expressed in a survey versus when it's tested at three o'clock in the morning?
The Gap Between Survey Trust and Production Trust
There is a well-documented phenomenon in organisational behaviour: stated preference versus revealed preference. What people say they would do in a hypothetical scenario and what they actually do under pressure are often remarkably different things.
A mainframe professional sitting at a desk, answering a survey during a quiet Tuesday afternoon, may genuinely believe they trust AI to complete a checkpoint adjustment. The same professional, watching transaction volumes spike during end-of-month batch processing while a severity-one incident is open, may reach for manual control instinctively — and rightly so.
This isn't a criticism of the respondents. It's a recognition that trust in technology operates differently under cognitive load, time pressure, and personal accountability. A pilot may trust autopilot completely — until turbulence hits and their hands move to the controls.
The survey measures attitude. Production measures behaviour. The distance between the two is where most automation projects quietly fail.
What would be genuinely valuable — and what no vendor survey can provide — is data on how many of those respondents have experienced AI making an incorrect autonomous decision in a production environment. Because trust that has survived failure is fundamentally different from trust that has never been tested. One is resilience. The other is optimism.
"Free the Human" — Free to Do What, Exactly?
The survey report introduces an evocative phrase: "Free the Human." The idea is compelling. AI handles the repetitive, rules-based, time-sensitive tasks — checkpoint tuning, log analysis, backup scheduling — and the human professional is freed to focus on architecture, modernisation, and strategy.
It's an attractive vision. It's also incomplete.
The question that "Free the Human" assumes but never asks is: does the organisation actually value the work the human is freed to do?
In many enterprise environments, the answer is no. The organisational structure, the incentive systems, the budget allocation, and the management culture are all optimised for operational delivery — keeping the lights on, hitting SLAs, processing transactions. The people who do that work are measured, funded, and rewarded for doing it.
When AI takes over operational tasks, the human is theoretically free to think strategically. But strategic thinking requires a different organisational contract: time to explore, tolerance for ambiguity, budget for experimentation, and leadership that values diagnosis over delivery speed.
Without that contract, "Free the Human" becomes "Make the Human Redundant." Not because the technology failed, but because the organisation never created the role that was supposed to absorb the freed capacity.
The most honest version of "Free the Human" would include a second sentence: "...provided the organisation is willing to invest in the work that remains." In most cases, that investment doesn't exist — and the human isn't freed. They're eliminated.
The Generational Shift: Knowledge Transfer or Knowledge Loss?
The survey documents a dramatic demographic transformation. In 2018, Baby Boomers represented 28 percent of the mainframe workforce. By 2025, that figure has dropped to 5 percent. Millennials now constitute 51 percent. Gen Z has risen from 1 percent to 15 percent.
The report frames this as cultural readiness for AI adoption. Younger professionals grew up with technology, trust automation instinctively, and see AI as an enabler. That framing is probably accurate.
But it overlooks something critical: what leaves when the Boomers leave.
The report suggests that generative AI can absorb documentation, system logs, and historical records to replicate institutional knowledge. The image is appealing — a digital repository of decades of accumulated wisdom, searchable and always available.
The problem is that the most critical institutional knowledge was never documented.
Anyone who has worked in a mainframe environment for more than a decade knows this. The reason a particular batch job runs in a specific sequence isn't in the documentation — it's in the memory of the person who discovered, fifteen years ago, that reversing the order caused a deadlock that took three days to resolve. The reason a certain CICS region has an unusual configuration isn't in the change log — it's in the head of the systems programmer who implemented the workaround during a crisis and never had time to write it down.
This knowledge doesn't exist in logs. It doesn't exist in documentation. It exists in conversations, in instincts, in the hesitation before someone says "don't touch that, I'll explain later." AI cannot absorb what was never externalised.
The generational transition isn't a smooth handoff. It's a race against time — and in many organisations, time has already won. The 5 percent of Boomers who remain are the last living documentation of systems that process billions in transactions daily. When they leave, that knowledge doesn't transfer to AI. It simply disappears.
The survey's optimism about generational synergy is understandable. But the operational reality is that most organisations have already lost more institutional knowledge than they realise — and AI, for all its capabilities, cannot recover what was never captured.
The Structural Question That Was Never Asked
There is a pattern in enterprise technology surveys that appears so consistently it deserves its own name. The pattern is this: the company that sells the solution conducts the survey that validates the need for the solution.
This is not dishonesty. It's structural incentive.
A mainframe software company surveys mainframe professionals about their readiness to adopt AI. The results show growing trust in AI. The company sells AI products for the mainframe. The survey becomes marketing evidence. The cycle closes.
The data may be entirely accurate. The methodology may be rigorous. The findings may be genuinely useful. But the questions that are asked — and more importantly, the questions that are not asked — are inevitably shaped by the interests of the entity asking them.
The survey asks: "Would you trust AI to complete this action?" It does not ask: "Has AI ever made a mistake that affected your production environment?" It does not ask: "Does your organisation have a governance framework for autonomous AI decisions?" It does not ask: "If AI freed you from operational tasks, does your organisation have a funded role for the strategic work you would do instead?"
These are not hostile questions. They are diagnostic questions. They are the questions that would complete the picture rather than frame it.
What Would Complete Diagnostic Data Look Like?
If someone were to design a survey that genuinely assessed enterprise readiness for autonomous AI in mainframe environments, it might include questions like these:
How many autonomous AI decisions has your organisation implemented in production in the past 12 months? Of those decisions, how many required human override or rollback? What percentage of your critical operational knowledge is documented in a form that AI could access? When a human operator is freed from operational tasks, what specific funded role absorbs their capacity? What is your organisation's governance framework for autonomous AI decisions, including accountability for errors?
These questions are harder to ask because the answers might be uncomfortable. But they are precisely the questions that separate survey optimism from operational readiness.
An organisation where 32 percent of staff trust AI to complete actions, but zero percent have experienced AI completing actions in production, has an attitude. An organisation where 32 percent of staff trust AI because they've seen it work correctly — and recover from errors — has a capability.
The difference is everything.
Trust Is Not a Spectrum. It's a Diagnostic.
The survey introduces a useful metaphor: the "Spectrum of Trust," progressing from alerts to recommendations to autonomous action. It's an elegant framework and probably an accurate description of how adoption evolves.
But trust in enterprise systems isn't a linear progression. It's a diagnostic outcome — the result of specific conditions being met.
Trust requires transparency: can the AI explain why it acted? Trust requires accountability: when AI makes a mistake, who is responsible? Trust requires survivability: has the AI demonstrated the ability to fail safely? Trust requires governance: are there guardrails that the AI cannot override? And trust requires organisational readiness: is the human side of the equation prepared for what changes when AI takes over operational decisions?
Without these conditions, moving along the "spectrum" is not building trust. It's building dependency on a system that hasn't been stress-tested.
The mainframe community has always understood this intuitively. You don't promote code to production because the developer trusts it. You promote it because testing, review, and validation have earned that promotion. The same rigour should apply to autonomous AI — and in most organisations, it doesn't yet.
DiagnosticMind Perspective
None of this means AI shouldn't be trusted on the mainframe. It should — eventually, conditionally, with evidence.
The mainframe is, paradoxically, one of the best environments for autonomous AI precisely because of the discipline the community already practices. Change control, audit trails, rollback procedures, testing protocols — these are the guardrails that make autonomy safe.
But adopting AI because a survey says people are ready is not the same as adopting AI because the organisation is ready. People are ready when they say yes on a questionnaire. Organisations are ready when they have governance, documentation, fallback procedures, funded roles for displaced capacity, and — critically — experience with AI failure and recovery.
The survey data is a starting point, not a conclusion. It tells us that attitudes are shifting. It doesn't tell us that organisations have done the work to make those attitudes safe.
That's the gap. And filling it isn't a technology problem. It's a diagnostic problem.
The BMC 2025 Mainframe Survey provides genuinely useful data about workforce attitudes. The analysis above is not a criticism of the survey but an attempt to add the diagnostic layers that vendor research, by its structural position, cannot provide. The full survey is publicly available and worth reading.