This article is written by Stuart Lauchlan on Diginomica on February 6, 2026
Summary:
The author of ‘Crossing the Chasm’ and ‘Into the Tornado’ offers a candid commentary on AI adoption trends as 2026 becomes a year of more pragmatic awareness of the challenges and opportunities.
Geoffrey Moore
When you sit down for an overdue catch-up with Geoffrey Moore, author of the seminal ‘Crossing The Chasm’, still the definitive tome on how products evolve from creation through early adoption to mainstream acceptance in a series of stages, there is one question that you really don’t want to be clichéd enough to ask him – and of course, you always do!
That was me this week when Moore joined me for a chat about AI and the state of that particular nation, and I found the words tumbling out of my mouth:
So, has agentic AI ‘crossed the chasm’?
Fortunately, Moore is well-used to this sort of thing. It’s also not his first rodeo when it comes to AI in its various incarnations over the years. Using the vernacular, predictive AI is well settled on ‘Main Street’, while generative AI is right ‘Inside the Tornado’ as we speak.
As for agentic AI, that’s in the process of ‘crossing the chasm’, but it’s still early days, and there are a number of barriers to adoption that need to be addressed. Moore’s analysis:
There are some beachhead markets, but when you ‘cross the chasm’ it’s when one function completely re-engineers itself. Coding is beginning to feel a bit toward that, Customer Service in call centers is a bit toward that, but I don’t think in either case, there’s enough trapped value in the process to cause the complete takeover of the thing.
Moore identifies the ‘Whole Product’ as a crucial concept in product adoption, this being an end-to-end solution that meets the expectations of a pragmatic customer. It is the hardest thing to form and so the most prized, he notes. As for agentic AI in this context, that’s some way off.
That being so, the root question to ask is if there is a business process in a particular industry that is in such dire need of being re-engineered that the industry is predisposed to a chasm crossing here. If so, can you then create a ‘Whole Product’ for that process? Moore picks out coding as a for instance:
I know that coding is desperate for this. That’s the problem. I think coding can use it, but we keep on talking about generating backend programmatic code. Testing and QA are the bottleneck, not coding. So if you said, ‘I have QI-enabled QA’, that might do more. And then the question is, ‘What would that look like?’
Another use case might be the organizational need to refactor code, he suggests. A frightening number of institutions still run mainframe code under emulation and processing the algorithms for key business processes:
Then you say, ‘OK, do you have an AI system that can read auto-code or object-code and figure it out?’. In that situation, the whole product will come into being where you have enough of a repeat of a particular process in a particular industry.
Metaphorically speaking
OK, so if that’s where we are with agentic AI, what can we do to speed up the progress towards ‘Main Street’? Interestingly, Moore is of the opinion that:
The narrative of agentic AI is still struggling.
One of the issues here is that there are a lot of wrong metaphors/analogies being bandied around about the tech, such as job killer, cheat tool, rogue loose cannon, thought expert that’s not as expert as you think, etc., etc. These have one thing in common, notes Moore:
They’re all based on, in some sense, on viewing AI as something you would outsource to, keep an arm’s length from, and you would not have a relationship with. The right metaphors have to embed relationship in the metaphor…Salesforce talks about managing a shared workforce of human agents and AI agents and the interactions between them. We’ve got to find some metaphor which says we create a relationship that is more mutual and less less almost xenophobic.
Such as a research librarian, entry-level buddy system tutor, a reliable concierge/personal assistant, that sort of thing? It’s a useful idea. The tutor angle is a particularly helpful discussion point, as a genuine fear around AI is that if it takes over all the entry-level jobs, as some predict with a little too much relish, then how do organizations build their expertise and human capital for the future? Viewed from that perspective, there has to be a case to argue that over-reliance on AI is creating a skills chasm to fall into rather than cross over.
Moore concurs that if you outsource, you cognitively decline, citing an example that must resonate with many of being entirely dependent on GPS systems in cars today, that the idea of having to find your own way to somewhere seems like a massive challenge. When it comes to organizations and their entry-level workforces:
The idea is you need to learn what the AI is doing. If you’re an entry-level person, your first job is actually to replicate what the AI is doing. Until you can replicate what the AI is doing, we can’t promote you because you don’t know the basics of our craft. Once you demonstrate that you do know the basics of our product, then we’ll let AI do it for you, because we want you to do other things. But I think, I think there has to be some sort of a training ground or a proving ground going forward, and the faster you can do this, as far as I’m concerned, the better.
But, bots?
A second linguistic barrier to agentic adoption might also lie in the habit of referring to agentic AI using the terminology of chatbot tech. For most people, bot is a three-letter four-letter-word, such is the loathing of so much chatbot tech to which people have been exposed to date. Such comparisons don’t help the agentic cause.
Part of the problem here is that agents are akin to a Pushmi-Pullyu, the rarest animal in the world according to Doctor Doolittle, with a head at each end. Agents can also be looked at from both ends. Or as Moore puts it:
The back half of an agent is a bot, but the front end of an agent is not a bot.
He expands:
Its front end is a conversational user interface that looks a lot like gen AI, its back end is a deterministic action which looks a lot like a bot….What we dislike about bots is they don’t have a front end, they’re headless, so we find them very hard to negotiate with. What we don’t like about gen AI is that it tends not to be able to take action.
Getting over that chasm
OK, so what is going to get us past this current position? At diginomica, we’ve always insisted that use cases are the key – what an end user’s organization’s peers are doing in practical terms with any tech is far and away the single most powerful marketing message that any vendor can point to, no matter how slick their sales pitch or charismatic their CEO. Moore agrees:
Transformative use cases are where it’s going to happen, but the most transformative ones are often in regulated industries, which are going to have their additional issues to build slow adoption. If you could imagine in Healthcare, or in Social Services, or Education, or Law Enforcement, any of those things, that’s where we really do need the agentic power boost, because workers are swamped.
He adds:
In the buying decision in corporations, there’s the economic buyers who are above the fray, and they are looking at the social issues and the liability issues or whatever. Then there’s the process owners who are kind of stuck with, ‘I need to do something to make my process more productive’. They’re gonna be much more receptive. And then the end users who are both excited and perhaps intimidated by, ‘What is this going to do to my workflow going forward?’.
The narrative here needs to be around releasing trapped value, a universal challenge for organizations regardless of sector. Moore explains:
In any industry, in any company, in any profession, where’s the trapped value? Because that’s where you’d like to apply technology to release the trapped value…Where’s the trapped value in your industry’s operating model? Why isn’t your company twice as productive as it is today? What’s stopping it? There’s obviously something stopping it, so where’s the serious constraint? Why aren’t you applying AI there? And the answer is,’ Well, I don’t know how’. But the point is, AI can’t answer every problem, but particularly after all this digital transformation, we know that the trapped value probably can be impacted.
And value doesn’t equate to cost savings, he adds:
That’s a conservative mindset. Cost reduction is always the least risky. It’s a Productivity Zone value, it’s not a Performance Zone value. Performance Zone wants to change the world; Productivity Zone wants to save money and reduce risk. And you can do that, but for the adoption life cycle, the Productivity Zone is part of ‘Main Street’, but to get things across the chasm, no, it’s not going to work.
SaaSpocalypse No!
Moore then makes a significant comment given the ludicrous so-called ‘SaaSpocalypse’ madness that gripped the more gullible parts of the tech sector this week:
My bet is that half the trapped value in the world you could solve with the data that’s in your system of record. Why? Because you have 25 instances of SAP, because you’ve acquired 16 companies! It’s not magic, right?
But surely, as OpenAI’s Sam Altman told us long before the latest Anthropic-led collective hysteria and doommongering, AI is going to spell the death of traditional software companies? The ‘SaaSpocalypse’ has already been foretold, has it not? Moore treats this delusional fantasy with the derision it deserves:
SaaS contains almost a half a century of business acumen, and I’m sorry, but you’re not going to just displace a half a century of experience, you’re just not. Every company in the world runs on systems of record and has overlaid systems of engagement, so nobody’s going to rip them out….When you’re an established enterprise, you have to work within some level of tradition.
He concludes:
The key narrative – and this is a narrative which I think the world is going to get, but it has been slow in getting it – if you believe that the end product is a deterministic action, the universe is deterministic actions live in the systems of record and systems of engagement. They don’t live in AI. OpenAI has no library of determinist actions. SAP does, Oracle does, Hubspot does, Salesforce does, ServiceNow does.
As for Altman’s view:
Sam’s a young man.
My take
As ever, an insightful conversation that could have covered a lot more areas. I felt coming into 2026 that this would be a year of not so much correction as adjustment of expectations around AI and agentic AI in particular, and that this was going to take some modified messaging on the part of vendors to set user expectations accordingly, without at the same time putting them off of the tech’s potential.
My summation of that was it was to be a year of, ‘Look, it’s not quite as simple as the evangelical fervor of last year might have suggested, but we’ve learned a lot and here’s what we now know about making this work’. Moore’s point about metaphors being a battleground is an interesting and chimes with my own worldview here.
I’ve never bought into the ‘AI as silver bullet’ meme – in 35 years of being around the enterprise tech sector, I’ve dodged far too many of those fired at me from vendors left, right, and center – and diginomica has always taken the stance that AI is a powerful enabling tool and complementary enhancer to human intelligence. Finding a way to communicate that should be the narrative of 2026 rather than indulging ill-advised PR flurries around so-called SaaSpocalypes! (Did Klarna’s CEO make a fool of himself in vain? It seems so!)
The concept of focusing on trapped value, even pockets of trapped value, within organizations rather than trying to boil the ocean, is also a highly useful one. We’ll return to that topic in the very near future. Ways to make that work in practice are worth further exploration.
Meanwhile, onwards – mind your step and don’t look down as you cross that chasm! (Sorry, Geoffrey, it’s just too irresistible.)