Podcast

8 min read

Live and hope: what’s bugging us about AI halfway through 2026

We went a bit grumpy before the holidays. Here’s what’s frustrating us about AI adoption right now, and what we’re actually looking forward to next.

Key ideas in this episode

Every software company is bolting AI onto its product, and it’s making things harder for buyers, not easier. The debate has become so emotionally charged that it’s getting in the way of sensible decisions. Businesses don’t need to keep pace with the models. They need a plan. And the shift worth watching next isn’t a new model; it’s AI that works alongside you rather than running off on its own.

Bugbear one: AI bolted onto everything

There’s a question we get asked a lot at the moment. A client points at a platform they’re already paying for and says: this has got AI in it, couldn’t we use that one?

It’s a completely reasonable question. It’s also usually the wrong one, and we understand exactly how we got here.

Nearly every software vendor now feels obliged to have an AI element, whether it makes sense for their product or not. It’s the industry trend, and if you’re not seen to be doing it, you look like you’re going backwards. Call it AI FOMO. The problem is what it does to the buyer. When we run a landscape assessment with a client, we deliberately put the technology to one side first and look at where the pain points actually are in the business. Then we ask how technology as a whole might help. Sometimes AI is part of the answer. Often, technology isn’t the answer at all; it’s a process or a system change. But when every tool on the shortlist has an AI badge on it, that conversation gets much harder to have.

You end up comparing a general-purpose model against an assistant embedded inside a project management tool, as if those were the same category of thing. They’re not. The AI in those platforms is applied within a very specific context, and it can be genuinely useful there while being no help at all across a broader set of use cases.

We want to be clear that we’re not criticising the vendors doing this well. HubSpot is the example I keep coming back to, because as an actual user I’ve found their support agent remarkable. It knows where I am in the system, it understands what I’m trying to do, and it steers me there. Things that would once have meant an hour trawling through documentation, or a conversation with a support agent, now take a couple of minutes. I’ve not yet had it fail to get me out of a tricky spot.

But look at what HubSpot is claiming. They’re saying: here’s an agent to help you use our platform. They’re not saying this is the AI tool you deploy across your whole business to replace everything. Where companies are upfront about that, and focus their energy inward on their own product, it works really well. Where the claims get bolder, and you pull the curtain back and think, well, it sort of does that, that’s aspiration rather than delivery, and it just adds confusion to the mix.

The related risk is AI being used as a shield: deployed to paper over deficiencies elsewhere in a system, or treated as a sort of god tool where every other improvement gets parked because AI is coming to solve it all. It usually isn’t the right architecture for the job.

Bugbear two: nobody can talk about this calmly

Our second frustration is the bigger one.

It’s become almost impossible to discuss AI in a moderate way. You’re either boosting it, or you hate it. There’s very little room for the sensible position, which is that the technology is transformative, has real problems, and needs proper thought.

We were comparing it to previous waves. Cloud felt significant. Mobile felt significant. The first wave of the internet felt magical. None of them carried this much emotion, and part of the reason is that they offered a much clearer choice. With cloud, you either kept the boxes in your building, or you moved them to somebody else’s. Black and white. You had your early adopters saying it’s brilliant, the cost, the scaling, and you had people worried about security, and then you had the classic adoption curve. These days it’s unusual to find anyone running their own kit outside a few niche cases.

AI doesn’t work like that. There’s a use-it-or-don’t choice, but that’s not the real question. The real question is how you use it effectively, and that’s a much harder thing to answer than lift and shift.

Underneath all this, there’s a sense that the technology is happening to us. The spaceship is descending, and we’ve no control over it, so all that’s left is to cheer or to complain. That’s the bit worth pushing back on, because businesses have far more agency than the noise suggests. There are genuine challenges to work through, around cost, environment, copyright and the anxiety about jobs. But treating the whole thing as weather rather than a set of choices makes it much harder to deploy it well. And the emotion in the room is itself slowing adoption down.

Where we think this lands: be more Apple

Apple has taken a grown-up approach here. They didn’t go mad building data centres or racing at the model layer, and plenty of people criticised them for it. Their bet was to let others own that part of the stack and focus on diffusion instead, on the application layer where people actually use the thing.

We’d like to see more of that thinking from businesses. Our honest view is that the models have been good enough for our clients for the last twelve to eighteen months. The new ones are great, and they add real capability in more sophisticated areas like code generation. But for what most SMEs actually need doing, a model from this time last year will drive it perfectly well.

The model announcements are close to background noise now. The interesting question sits one level up: how do you get the best out of what’s already there, and where in your business is it worth deploying?

Which means treating this like any other change programme. Work out where the challenges are in your business, look at the best technologies available currently to overcome them, and then execute on that plan in a considered way. We say currently deliberately, because these things move fast, not just the models but the services built on top of them. So build it so you can change course. You want to be able to ride the wave, and if that one diminishes and another comes along, switch. Get the immediate benefit without locking yourself in.

And yes, we’re aware that’s fundamentally a description of what we do for a living.

What we’re looking forward to

The labs coming up the stack. Anthropic and OpenAI have worked out that the value is in providing solutions, not just the models. We’re seeing it already with things like Claude Design and Cowork, where model capability gets wrapped into a service that solves a particular problem. Compare where those capabilities were this time last year to now. We’d expect a lot more of that in the next six months, and it will be interesting to see how the established software companies respond as the labs move onto their territory.

Collaboration agents rather than automation agents. Almost all the conversation so far has been about the fully automated agent, where you point it at a task and let it loose, and then everyone acts surprised when it breaks out of the sandbox. We’ve yet to see one deliver a clear benefit. What’s more interesting is the halfway house between a chatbot and a fully automated agent: something that takes on a task but pushes back, asks questions and seeks permission before it acts. Claude Code works roughly this way. Better on cost, better on security, and much closer to what most businesses actually need. That’s a richer hunting ground for anyone trying to create real value.

A proper conversation about cost. The pricing is shifting. I logged in this morning to find Anthropic’s newest models moving to pay-as-you-go rather than sitting inside the standard bundle. That’s effectively them acknowledging that the older models are good enough for most work, and if you want the top one, you pay for it. We expect more considered use as a result: different models for different jobs, and more interest in model routing, using AI to pick the right tool automatically. At the moment, that’s something an engineer largely builds, but we would expect this to become more accessible to ordinary users. Nobody quite knows which way the cost structure will go yet, and it may not be a comfortable picture, but a budgeting conversation is a sign of an industry growing up.

Theo shared an article a while back comparing all this to the rollout of electricity as a utility, and it stuck with me. When I was a kid, you got cheaper electricity overnight, so you’d time things to run then. Same idea as routing: infrastructure arrives, and we find intelligent ways to work around the cost. But the deeper point is that businesses had to change how they operated to get the value out, and that took time. That’s exactly where we are. The models are racing ahead, and organisations are trying to keep up.

Our advice is: don’t. Let them race off. Get on with the work.

What should you do?

  1. Ask what problem you’re solving before you ask which tool. The AI in your existing platforms may be excellent inside that platform and useless outside it.
  2. Turn the volume down. Treat this as a technology decision, not a belief system.
  3. Run it as a change programme. Bottlenecks first, technology second, plan third.
  4. Be sceptical about full autonomy. Watch the collaborative agents instead.
  5. Cost it properly. The pricing model is shifting under everyone’s feet.

We’re both off on holiday. When we’re back, we’ll see how much of this actually landed.

One more thing: our 20th episode is coming in August, and we’re using it to answer all the questions we’ve left hanging in previous episodes. We’d like yours too, whether that’s the business side, the technical side, lessons from the past, or hopes and fears for what’s next. Send them through our LinkedIn channels or by email, as addressed in the show notes.

A deep dive into full-trust AI for SMEs.

Below the Surface: full-trust AI

Real AI adoption needs sensitive data, and commercial LLMs can't be trusted with it.

A full-trust AI stack is easier and more affordable than most think.

We compare 6 models, explain 5 layers and outline 4 steps to start.

Read the free guide
Giles Thurston and Theo Paraskevopoulos