Insights

Savannah, Reef, Invasion or Farm? Four futures for the AI industry (Part 1 of 2)

By Will Hardman · · 7 min read

Lately, I’ve been wondering what the AI industry might look like as it matures and what kind of economy it will shape for everyone else. The range of views among the commentariat is striking, opinions range from “AI will barely show up in long-term growth statistics” to “we’re at the beginning of an exponential that will reshape everything”.

Such diversity of opinion poses a challenge to business leaders: how should they invest in AI - and plan for the future - when said future is so uncertain? In this, two part, article I will present four plausible scenarios (with a rough three-year time horizon) and use them to craft a simple list of “no regret” actions, actions that every leadership team should be executing on today. In brief, they are:

  1. Watch for and study your “AI native” challengers
  2. Ready your data and knowledge to build an “AI moat”
  3. Plan your investments to avoid AI vendor lock-in

The question has many facets: economic, competitive, technological and regulatory. To bring clarity, I’ve attempted to simplify it. Sifting through the analyses of many well-credentialed commentators, I’ve lifted out two principal axes along which opinion ranges:

  • Substitution vs complementarity: will AI agents fully substitute and replace (a significant fraction of) human knowledge workers or merely empower them to achieve more?
  • Concentration vs diffusion: who generates the tokens that get used for economic work - and at what price?

These axes take the form of questions which will resolve, to a greater or lesser extent, in one of two ways: will it be this way or the other way? I chose the axes according to three criteria:

  1. independence: how one resolves does not affect how the other will resolve
  2. decisiveness: the answers will materially affect the future
  3. uncertainty: it’s genuinely unclear how it will play out

Before arriving at the axes, I considered and discarded several others which failed to meet the three criteria (for instance, “Will model capability plateau?”, which is tangled up with the Concentration axis or “Will energy and power constraints bite?”, which could influence both axes but will be decisive in neither). What follows is the strongest pair that is independent, decisive and unknown.

The first axis: substitution vs complementarity

AI models have limitations. Among these are an absence of memory and personalisation mechanisms, a range of cognitive biases and a propensity for task hyperfocus that misses the bigger picture. At present, deficiencies like these prevent widespread substitution of humans for knowledge work. But will this change?

The case for substitution rests on two claims:

  1. Most firms have not yet learned how to apply the technology effectively. Many observed deficiencies of current models can be addressed by engineering systems (“harnesses”) around them. The craft of harness design is in its infancy and there is no reason to believe that, as best practices emerge, agents will not become greatly more reliable at specific tasks. In fact, we should assume the opposite.
  2. There is no ceiling in sight that might curtail further improvements in models. Every release pushes progress on the benchmarks - and this progress seems to be accelerating. Indeed, the AI labs are chasing a capability called “recursive self-improvement” (RSI), whereby frontier AI models will drive research into their own replacements. If realised, model gains will happen even faster than they have to date.

In case that sounds fanciful, in autumn 2026 METR - an independent AI evaluation group - was given access to Anthropic’s internal reviews of model progress. They concluded that Anthropic’s models were improving at twice the rate seen in the period between March 2024 and October 2025, with most of the acceleration down to Anthropic’s own use of AI in its research.

The case is compelling. However, there is an equally strong case for complementarity - which could hold even if models continue to improve - and its proponents lock on to one of two rebuttals:

The “lump of labour” fallacy

Substitution assumes there is a fixed amount of cognitive work to go round - the “lump of labour” fallacy. But surely, organisations’ appetite for cognitive labour is effectively unbounded. Models and humans are cognitively very different and complement one another well; ergo, agents will augment - rather than replace - their human counterparts.

Jagged progress

Models’ improvements have concentrated almost overwhelmingly in so-called “verifiable” domains. A verifiable domain is one in which tasks have definitive and easily checked answers. Think factual recall, mathematics or programming (where software tests can check that code performs as required). Verifiable domains lend themselves very well to large-scale, automated model-training programmes.

On the flip side, “non-verifiable” tasks make up much of the work that we actually ask models to perform: design briefs, expert advice and broad research are examples. Models have been notably slower to improve at these tasks. A case in point: many practitioners believe that the quality of creative writing produced by LLMs has actually decreased since the advent of reasoning models - a direct result (for technical reasons) of over-training on verifiable tasks.

The implication of this “jagged” rate of improvement is that there is a plausible argument that models will not reach expert human levels of capability across the full spectrum of cognitive tasks for a very long time. If true, it will constrain the domains in which fully autonomous agents can be deployed without human mediation.

The second axis: concentration vs diffusion of supply

At present, a small handful of frontier (US-based) labs - chased by a larger handful of (mostly Chinese) fast-followers - supply most of the tokens used for work. Beneath them sits a broad ecosystem of smaller suppliers, many developing or fine-tuning their own models, each betting that specialised AI will win in niche use cases or industry verticals.

The labs’ business model is this: they supply tokens and consumers (companies) apply them to do work. Any economic benefits therefore accrue to the consumer. But could this change? In July, Anthropic announced it had set up an in-house, pre-clinical drug discovery programme. Drug discovery is a use case in the crosshairs of the labs for two reasons: firstly, it is a market with enormous financial potential and secondly, there is a tight and verifiable feedback loop between molecular design and experimental assays.

One way to look at this announcement is to realise that selling tokens to a pharmaceutical business allows the majority of the value of a drug patent to be captured by the customer, not by Anthropic. Owning the research capability and the patent would flip the script. If this works within life sciences, where else could it work?

The answer to this question depends on two factors that are very difficult to forecast, at the time of writing.

Do specialised models have a place?

As the leading labs continue to train larger and more effective, general-purpose models, a swarm of smaller firms are building specialised, smaller ones. A recent example is the Jev decision model from TypeSafe AI. Jev took a specific use case (structured decisions from natural language inputs) and built a cheap, effective product. Almost overnight, they captured a chunk of work that had previously been routed to LLMs from frontier labs.

Specialised model providers are betting that a sizable fraction of enterprise workloads can be delegated to faster, cheaper models better at specific tasks. If they are right, then firms will build systems on these specialised models and retain a competitive edge within their domains.

Where are moats unbridgeable?

Most industries have barriers to entry - barriers that cannot be overcome by simply applying AI compute. Consider the following:

  • proprietary knowledge and know-how
  • long-term customer and supplier relationships
  • manufacturing and distribution capabilities
  • licences and regulatory approvals
  • privileged market and channel access
  • access to suitable capital and credit lines

To compete with incumbents, AI labs and “AI native” firms must bridge these moats - or buy their way in. Their ability to do so is, as yet, unclear.

Next up

In part 2, I will take these two axes and use them to build four scenarios for the AI industry of 2029. I will draw out the implications for non-AI businesses and make three recommendations for action today.

Ready to move from pilot to production?

Book a free 30-minute consultation. You'll leave with a straightforward answer on where agentic AI can work for your organisation, and the first steps to get there.

Book a free 30-min consultation →