Every enterprise software company has an AI story now and most of them sound identical. It starts with a press release that goes out, the word “revolutionary” appears somewhere in the second paragraph. A dashboard gets a new tab but when the customer actually uses it, the question is much simpler: what problem did this solve for me?
Too often, the answer is not obvious. That does not mean the technology is not impressive because it is but the problem is that AI commentary has consistently run ahead of the usefulness customers actually experience in the products themselves.
Research published by Gong in April 2026 found that 58% of companies surveyed had stalled AI projects, with trust concerns including data security, explainability and model transparency emerging as significant barriers to adoption. I think that trust gap is solvable but not in the way most vendors seem to think.
What actually changed
When people explain why AI is more useful now than it was two years ago, the conversation usually focuses on capability. Models are larger, reasoning is better and outputs are more reliable. But capability is not the change that matters most for people building enterprise products. More important is the economics as capabilities that once demanded specialist teams, significant infrastructure and long development cycles are becoming far more accessible. That changes the economics of which problems are worth solving and for whom.
For enterprise software, this is significant because capabilities that previously only made sense for large organisations with substantial AI budgets are now within reach of mid-market businesses. The question has shifted somewhat from “can we build this?” to “which problems are actually worth solving?” Those are very different questions and the second one is a more interesting place to spend time.
Start with the problem
I have spent more than twenty years in B2C and B2B building digital products, from mapping and navigation to EV charging and today, fleet payments. The discipline that experience teaches you, above everything else, is to start with the customer problem and work backwards. Now that might sound obvious but most teams don’t do it.
The temptation with AI is strong to start with the technology. The capability is genuinely impressive, and the demos are compelling. But it’s easy to build something that looks remarkable in a controlled environment and then later discover, when real users try it under real conditions, that it solves nothing they actually needed solving, or is just ‘interesting’ but so what?. More insights for insights sake doesn’t help solve a problem.
A product that earns lasting adoption is usually one where the technology disappears and isn’t even noticed. Nobody should be thinking about the AI layer when they use a well-designed tool. They should be thinking about the decision they just made faster, the anomaly they caught before it became expensive, or the report they got in plain language instead of having to extract it from a spreadsheet. The technology earns its place by becoming ‘invisible’.
Trust is not a communications problem
Today many enterprise AI vendors talk about trust a great deal. Most treat it as something to address in messaging rather than build into the product from day one. However, in practice, trust in an AI-enabled enterprise tool comes down to two questions. Can the user see the source to show why a recommendation was made? And can they choose to accept or override it? If the answer to either is no, the product does not deserve to be trusted as it’s simply been built to look confident. How many times have you used AI to find questionable output and the response from the GPT is “You’re right. I did miss that…”.
Deloitte’s 2026 State of AI in the Enterprise report, based on responses from 3,235 senior leaders across 24 countries, found that only one in five companies has a mature governance model for autonomous AI agents. As AI moves from producing information to taking action, questions around human control, auditability, legal and compliance, and oversight become more important.
Now in a consumer app, a wrong recommendation is an inconvenience but at the enterprise level, where a tool touches business operations, financial decisions, or regulatory compliance, the bar is considerably higher. The design philosophy, the governance model and the oversight mechanisms all have to be built from the start. Now I’d argue that showing your working is not a limitation at all. A tool that explains what it detected, why it flagged the issue and what the user can choose to do about it is more valuable than one that produces an output and asks to be trusted. This builds confidence through transparency and reduces the ask of users to make a leap of faith that most enterprise buyers will not take.
Control matters more than intervention
There is a strand of AI commentary that treats trust as a simple choice between human control and automation but I argue that misses the point. The question is not whether a person needs to approve every action but more whether they understand what the technology is authorised to do, why it is acting and where the boundaries sit.
For routine, low-risk tasks, that may mean allowing technology to act without interrupting the person responsible. As confidence develops, the scope for delegation can grow too. But autonomy should be earned rather than assumed, with clear permissions, transparency and the ability for customers to change those boundaries when they need to.
Overall, the goal is to remove work that does not require human judgement while keeping people in control of the outcomes that matter. Done properly, that creates more time for the decisions where experience and judgement genuinely add value.
What the next generation of enterprise AI actually looks like
The enterprise AI tools that are adopted and last the course of time will be specific about the problems they solve, honest about what they do and do not do and will make the human’s role in the system visible. They will also earn trust through consistent, accurate, useful output over time and do not overclaim what the technology will eventually be capable of.
However, the shiny ‘game changers’ are those built on the assumption that the announcement is the product. Where the story told externally runs significantly ahead of what a user actually experiences and where the trust conversation happens in press releases and not in the interface. But this is solvable as the economics now exist to build genuinely useful AI capability at enterprise scale. The question is whether teams are willing to start with the harder, less glamorous work: defining the customer problem precisely, building the transparency mechanisms that create real confidence and resisting the temptation to announce a future the product is not yet ready to deliver. Or the customer to use.
The test worth running
Before shipping any AI-enabled capability, the question worth asking is does this solve a problem the customer actually has, can they see why it is making the recommendation it is making and can they choose to do something different? If the answer to all three is yes, the technology has earned its place. If not, spend more time on the problem before spending more time on the solution.
