Most AI predictions from 2023 were wrong. Not directionally wrong — the technology did arrive — but wrong about where it would stick. At Nuclear Marmalade, we've spent two years building AI into real products for real businesses. The gap between what the demos promise and what ships to production is still wider than most people admit. Here's what we've learned.
What's actually working with AI right now?
AI earns its keep on high-volume, low-stakes repetition. Customer triage, document summarisation, first-draft generation, log parsing — tasks where "good enough" genuinely is good enough. Telehance was partly built around this: route the easy stuff to automation, surface the hard stuff to humans. Client phone handling dropped from 4 hours a day to under 12 minutes. That's not a rounding error. That's someone's entire afternoon back.
The pattern holds across every project we've touched. AI doesn't replace judgment. It clears the road so judgment can move faster.
What does AI still get badly wrong?
Confidence without competence. That's the core failure mode.
The model sounds certain when it's guessing. It hallucinates citations, invents API methods, mis-remembers data it was handed two paragraphs ago. We watched it tell a client, with complete conviction, that their contract said something it absolutely did not say. That's not a UX quirk. That's a liability.
Anything where being wrong has real consequences still needs a human in the loop. Medical, legal, financial, compliance — these aren't spaces where you ship and iterate. The companies winning right now aren't the ones who automated everything. They're the ones who were honest about where the floor is.
Why do so many AI projects fail before they ship?
They confuse prototypes with products.
Getting GPT-4 to answer a question in a Jupyter notebook takes 20 minutes. Making that same answer reliable, safe, auditable, and fast inside a production system takes months. I've seen this play out more times than I'd like. The demo wows the boardroom. Developers get greenlit. Six months later the thing still isn't live because nobody scoped the edge cases, the latency requirements, or the data pipeline underneath it.
We now build a "failure budget" into every AI feature we scope — an explicit, agreed tolerance for how wrong the model is allowed to be before we add a guardrail or a human checkpoint. That conversation alone saves weeks.
How has AI changed what clients actually ask us to build?
The requests have gotten more specific. More honest.
In 2023, everyone wanted "an AI chatbot." In 2026, clients arrive with actual workflow problems. We have 800 inbound messages a week and two people answering them. Our team spends 3 hours a day categorising support tickets. That's a productive conversation. It means there's a measurable before and after.
Forge and Nuclear Directories are both examples of what happens when the problem is concrete — the AI component slots in cleanly because there's a real job for it to do. Vague ambition still produces vague software. That hasn't changed.
Is AI going to replace developers and designers?
No — but it's already replacing some of what they do, and pretending otherwise is bad advice.
Boilerplate code, UI scaffolding, first-pass copy, unit test generation — AI handles this faster than any junior can. What it can't do is understand why something needs to exist, or make the call when two valid approaches conflict, or know that the client's stated requirement isn't actually what they need. My background is in both design and technical development — more on that at /founder — and the honest read is that the mid-tier of both fields is under real pressure.
The people who'll be fine are the ones treating AI as a multiplier. The ones who won't be fine are the ones who think the baseline skill floor just got lower. It didn't. The bar for what passes as "good" went up.
What's actually coming in the next 12 months?
Three things worth watching.
First, agentic workflows going from demos to defaults — AI that doesn't just answer questions but takes sequences of actions across tools. We're already building early versions of this into client work. The productivity gains are real. So are the failure modes.
Second, memory and context getting dramatically better. Most models are effectively stateless between sessions right now. That's changing fast, and it changes what's possible for personalisation.
Third — and this is the one nobody's talking about loudly enough — the commoditisation of the model layer. When GPT-5, Claude 4, and Gemini Ultra are all roughly equivalent on most tasks, competitive advantage shifts entirely to implementation quality, product thinking, and the data you own. That's good news if you build things. It's bad news if your whole strategy is "we use AI."
How should a business actually start with AI today?
Start with a problem, not a technology.
Find one process that's high-volume, low-variability, and currently eating someone's time. That's your pilot. Don't build a platform. Don't hire an "AI team." Don't commission a strategy deck. Run one experiment, measure it honestly, and decide from there.
If you want to think through what that looks like for your situation, reach out to us at Nuclear Marmalade — we do this scoping work and we'll tell you straight if AI is the right tool or if you're solving the wrong problem. The worst outcome is six months building something that wasn't worth building. The second worst is not starting because the whole thing feels too big.
Key Takeaways
- AI earns its keep on repetition and volume. High-frequency tasks where "good enough" is genuinely good enough — those fit. If being wrong has serious consequences, keep a human close.
- The demo-to-production gap is still enormous. If your AI project hasn't shipped yet, the problem probably isn't the model — it's the edge cases and infrastructure underneath it.
- Clients are asking better questions now. "We have 800 messages a week and two people" is a solvable problem. "Give us AI" is not a project.
- The model layer is commoditising fast. In 12 months, your edge won't come from which AI you use — it'll come from how well you've built around it and what data you bring to it.
- Start with one boring, specific, time-consuming problem. Solve that. Then decide what's next.