The next wave is narrower, not bigger
For a couple of years the AI story was a race for scale: bigger models, longer context, higher benchmark scores. That race is far from over, but the place where real value is being captured has quietly shifted. The next big wave isn't a bigger model. It's a narrower one.
The signals are clear. Anthropic shipped a set of preconfigured agents purpose-built for finance — investment banking, asset management, insurance. Novo Nordisk signed a multi-year deal to embed AI across its entire drug pipeline, from discovery through to manufacturing. The momentum is moving away from general-purpose chat and toward vertical agents that know one job deeply.
Why the brief is changing for builders
When everyone has access to roughly the same frontier models, model access stops being a differentiator. It becomes table stakes. That changes what "building an AI product" actually means.
The moat is no longer the model. It's everything around it:
- Workflow knowledge — knowing the actual sequence of steps a finance analyst or clinical researcher takes, including the parts nobody writes down
- Proprietary data — the labelled examples, historical decisions, and feedback loops a generalist model has never seen
- Domain edge cases — the compliance rules, clinical protocols, and supply chain logic where being almost right is the same as being wrong
A generalist LLM can write a plausible-sounding answer about KYC requirements or a drug interaction. A well-built vertical agent can encode the specific rules, escalation paths, and exceptions that make that answer trustworthy in a regulated context. That gap is the product.
Turning domain expertise into agent behaviour
The hard, valuable work is translating what your team knows into behaviour the model can't reproduce on its own. In practice that looks like:
Encoding the workflow, not just the prompt
The value isn't a clever system prompt. It's modelling the real decision tree — what to check first, when to ask a human, which sources are authoritative, how to handle the ambiguous middle.
Building the data flywheel
Every correction a domain expert makes is training signal. The teams pulling ahead capture that feedback systematically and feed it back into evals and retrieval, so the agent gets sharper at exactly the cases that matter to their vertical.
Designing for the edge cases first
Generalist tools optimise for the common case. Vertical winners obsess over the long tail of exceptions, because that's where domain expertise lives and where trust is won or lost.
This is precisely where having a partner who builds, not just advises, pays off. We're here to help founders and teams design and build digital products that are built to scale with you, not slow you down. If you're looking to build something, get in contact with us today.
The takeaway
The teams winning the next 12 months won't be the ones with the smartest model — everyone will have access to that. They'll be the ones who turned deep domain expertise into agent behaviour a generalist could never reproduce. So if you're building an agent today, the question that matters is simple: what does your team know that a general-purpose LLM never will? Whatever the honest answer is, that's the layer worth shipping.