The fever around AI agents — advertised as AI tools that perform actions with some degree of autonomy and agency — has taken the tech sector by storm. A few companies see them as potential for expansion — developing agents for their applications or enabling users to create their own agents. But the reality is that agents are a threat to the commercial software industry.
Why? Because agents expose the tech sector’s lies and, most outrageously, the disdain for its customers. And now it is payback time.
Let me show you the tech narratives that got us here.
Narratives About Users
Users do not know what they want
One of the favourite ways of tech to defend their ideas against customers’ interests is the apocryphal quote from Henry Ford, “If I had asked people what they wanted, they would have said faster horses.”
This patronising way of developing software is a declaration of intentions — we, those who design software, know better than the people using it. As a result, we decide what to build, which guardrails to apply (if any), and for how long we want to maintain it.
AI Agent Payback (AIAP): Users see agents as a way to finally get software to do what they have asked tech to deliver for decades.
Users are not clever enough
Is your application’s user interface an impenetrable maze, where customers get lost? Or have you advertised your tech product as “democratising” highly complex knowledge when, in reality, it can only be successfully used by people with a PhD and two post-docs?
The answer from tech is to blame it on their users, selling them remediation training and services.
AIAP: Users feel empowered by developing their own agents, which bypass tech companies as painful “intermediaries” between what they want and what they get.
Users are gullible
Tech has embraced the mantra that cybersecurity is a people problem. Has your organisation been hacked, a victim of ransomware, or scammed by a deepfake? Users — aka “human error” — are surely at fault.
As such, users are expected to become cyber risk detectors, while software companies are absolved of ensuring their clients’ safety and security.
AIAP: As users feel they are on their own to fend off cyberattacks, agents do not appear riskier than other software applications.
Users are too complicated
Tech has tried to convince us for years that humans are too complicated and that the only remedy is to create products for a very limited set of idealised standards called “user personas”.
Is the application inaccessible to some users? Biased against them? Or unable to meet their needs? Then those users are deemed to be the problem and left to cope alone. The real issue? The inability of user personas to capture the breadth of the human experience.
AIAP: As users expect to adapt to how applications work, rather than the reverse, clunky agents are not seen as a downgrade but a continuation of subpar tech experiences.
Narratives About Tech
Tech is reliable
Tech companies shower us with their reliability metrics: platform uptime, incident time to resolution, and penetration test results.
The reality is that the commercial tech sector has shown us how they can capitalise on our data (Cambridge Analytica Scandal), massively botch software upgrades (Crowstrike-Microsoft outage), and how their dependence on “free” tools can expose millions of users around the world to cyberattacks (Log4J vulnerability).
AIAP: Users do not feel worse off by the hit-and-miss of agents compared to what they perceived as unreliable service by software providers.
Tech sets the pace
We have made tech synonymous with innovation and progress. As a result, users are expected to upgrade their hardware and software at the pace dictated by tech, even if that upends their workflows or makes the tool suddenly inaccessible. They are also expected to suck it up when their tech provider changes or removes features they rely on for their job.
AIPP: Confronted with the lack of control over what they get from software vendors, agents provide users with a sense of agency.
Tech leaders are the experts
When computers took off in the 1960s, women became the programmers while men focused on the hardware, which was regarded as the most challenging work. As programming gained status during the 1980s, men pushed women out of those jobs.
That prompted a sharp increase in software developers’ salaries. It also manufactured the concept of the software developer as a (male) super-intelligent being. From Bill Gates, Steve Jobs, and Geoffrey Hinton to Mark Zuckerberg, Elon Musk, and Demis Hassabis.
As a consequence, tech has thrived on advertising coding as a mysterious craft only within the reach of the chosen ones.
AIPP: Agents have demystified coding for users, making it seem like a nice-to-have skill rather than a requirement.
Tech has the right to set the boundaries
The sector has espoused isolation and a lack of interoperability between applications and operating systems (Apple vs Windows). While it has been sold as a way to protect users, in reality, it has been weaponised against them, making them hostages of tools and platforms. It has also put the onus on customers to transfer data among applications and to make disparate applications work to get the outcome they desire.
AIPP: Agents promise to break the silos among different applications for users.
AI Agents Dark Side
AI agents provide a sense of empowerment with their promise to enable non-techie users, or techie users with limited resources, to bring to reality their ideas, navigate bad user interfaces, and automate the connection between non-interoperable applications — all that with a very low learning barrier, for free or a fraction of the price-of mainstream applications, and no worse experience and risk than usual.
But is that true?
Software development is a craft
I have been coding since the late 80s. I have also been working for software companies for 20+ years as Head of Training, Head of Contract Research, and Director of Scientific Support. One of the things I learned over the last two decades is that developing software is much more than coding.
Good software development is not only about achieving the goal — agents’ promise — but about the how — efficiency, reliability, guardrails.
Whilst many digital applications may not meet our expectations, it is important not to underestimate the know-how necessary to design, develop, and deploy applications at scale.
Maintenance is not sexy
We like boring. We want to be able to use our car and have it start every time, no matter the weather. Likewise, we want our phone, internet, and email to work flawlessly, with the option to call support if they don’t.
But who do you call when you are the one who created the tool? What happens when the large language model (LLMs) your agent uses gets updated? Or when some of the applications the agents access are deprecated?
Maintaining an application functional over the years is tedious work. With agents, users are — mostly unknowingly — accepting the task of fulfilling that job.
The inaccessibility avalanche
I still cringe when I think about how, six months after launching my website on diversity and inclusion in tech in 20218, an expert in disability asked me if it was accessible and pointed me to the post 10 Ways to make your blog accessible for people with a visual impairment on the site Life of A Blind Girl. Reading the article was transformative. It made it clear to me that, irrespective of my intention — promoting diversity and inclusion — my impact was the opposite: I’d potentially been frustrating and excluding millions of people with visual impairments who use screen readers from my website.
Fortunately, the awareness about the need to embed accessibility in tech products from inception has increased exponentially in recent years. This has also had repercussions at the regulatory level, with more countries establishing standards to ensure users with different accessibility requirements have a similar experience. It has also provided more visibility to accessibility experts.
With the rise of AI agents, accessibility professional expertise is likely to be overlooked, increasing the risk of creating workflows and applications that exclude other users.
The hidden costs
The typical business model for large language models (LLMs) is either free — Meta, Alibaba, Deepseek — or freemium — OpenAI, Anthropic, xAI, Mistral, Google — fostering the belief that there is no cost to running agents.
Whilst LLMs may appear free or cheap for the end user, they are not without cost. For example, the minerals and metals used to build the chips or the low wages paid to miners exploited to extract them. There is also the increase in electricity bills as a result of data centres’ power consumption. And the 40,000 tons of e-waste generated annually, of which 70% is toxic and only 12.5% is recycled.
What about agents? Unlike other tech applications, there is no promise of efficiency. Trial and error — disguised with terms such as “thinking” and “chain-of-thought” — is the norm. In other words, what was offered for free or at very low cost becomes expensive in the long run.
For example, as some users of the Hermes agent recently discovered, there is no free meal in the universe: token consumption can ramp up very rapidly, making the cost of running agents unsustainable for many users.
You don’t know what you don’t know
A key feature of agents is that they partly rely on non-deterministic capabilities. That is, there is a certain amount of uncertainty about how the same goal may be achieved every time. Moreover, they are more prone to crashes than traditional software applications because they are not typically thoroughly tested for all use cases.
AI agents have also been shown to compromise personal data at scale and to become easy targets for cybercriminals. For example, an AI agent that replies to your emails may be hacked and, as a result, add phishing links to your emails without your consent.
Finally, AI agents rely heavily on using LLMs for code generation. That means that the applications created with them will be easier to break because the code will be more predictable.
In summary, they make users more vulnerable.
Hobby vs Work
There is no proven business model for AI agents. Not even for LLMs. That means we should expect agent fees to change overnight, many providers to disappear, and zero commitment to fixing bugs or providing customer support.
In my book, that means AI agents should be treated as a hobby or for prototyping purposes rather than as a structure that underpins a business.
My take
The gap between AI agents’ promise and reality is well exemplified by the message conveyed by Peter Steinberger, creator of OpenClaw’s AI agent platform, in a TED Talk and before an AI engineering audience. During the first, he enthusiastically shared his breakthrough in creating the platform, how he felt persecuted by the LLM company Anthropic over trademark issues, his thoughts on how AI agents might reshape your ability to work, create, and build, and even had time to brush off cybersecurity concerns. In the second one, he disclosed the unprecedented number of security incidents for OpenClaw (1,142 security issues and vulnerabilities in 69 days, 99 of them critical) compared with other large open source projects such as curl and Linux. Moreover, he stated that “AI Agents are both the product and the attack vector”
There are powerful reasons why AI agents are dominating the tech conversation. They are the answer to our sector’s shortcomings and disappointments.
Whilst tech well deserves to be shaken to its core, it is a disservice to those users who feel disempowered to foster the expectation that AI agents are the free answer to their prayers for personalised software applications.
To move forward, tech should learn from this disruption that it cannot take users for granted and that it needs to show repentance and mend its ways.
How can tech do that?
- Demystifying tech. Whilst developing software is a craft, developers are not inherently more intelligent or know more about their users’ needs than the users themselves. Moreover, users deserve fit-for-purpose interfaces, useful documentation, and a focus on addressing their pain points.
- Increasing the transparency about how decisions are made. Software providers should embrace clarity about fixing bugs and implementing user requests, banishing the dreaded canned statement: “considered for a future release.”
- Most commercial applications are either out-of-the-box or “empty containers” that require enormous professional services budgets to meet customer requirements. This needs to change. It is imperative that tech provides alternative ways to give customers more agency and the ability to personalise their experiences.
- Leveraging the power of agents to enable customers to build rough versions of how they want to use the applications to solve their problems. Then it is for commercial tech to build robust versions of those workflows.
- Creating applications that make it easy for customers to operate with all their digital ecosystem, rather than striving to keep them captive.
AI agents are not a substitute for well-designed applications, whether deployed on-premises or via SaaS(software-as-a-service). However, it would be a mistake to dismiss them as a fad or a personal toy.
Tech traditional players must take users’ enthusiasm for AI agents as a clear signal of their customers’ massive dissatisfaction with the power asymmetry and a lack of control that has plagued commercial software.
Will the commercial tech sector finally learn the lesson?
WORK WITH ME
To cultivate thought‑leadership and boost your career prospects:
- [NEW] Women in Payments: Leadership Collective [Spring 2026 cohort]: I am delighted to be one of the coaches of this bold, four-module virtual experience designed to empower mid to senior-level professionals across the global payments ecosystem. I will be delivering the module AI for Leaders: Navigate the Future with Confidence. Key themes include:
- Understanding AI, machine learning, and generative technologies
- Gaining insights into the ethical, environmental, and societal impacts of AI
- Applying a decision-making framework to real-world leadership challenges
- Navigating digital disruption with strategic foresight and purpose