# Has AI reached its saturation point?
When we first asked whether Artificial Intelligence had reached its saturation point in 2024, generative AI was dominating the technology landscape. ChatGPT had brought Large Language Models into the mainstream, image and video generation were advancing rapidly, investment was pouring into AI startups, and seemingly every technology company was adding an AI capability to its products.
Two years later, the question is arguably more interesting.
AI has not stopped advancing. Models have become more capable, multimodal and significantly cheaper to operate. Reasoning models can spend additional computation on difficult problems, smaller models can achieve performance that once required vastly larger systems, and AI is increasingly moving beyond generating content towards planning and executing tasks.
Yet something else has changed. The novelty has begun to wear off.
The question is therefore no longer simply whether AI itself is reaching saturation. Instead, we need to distinguish between technological saturation, market saturation and organisational saturation. They are not the same thing.
The AI market may be crowded. AI itself is not.
There is certainly evidence of saturation in parts of the AI market.
The first wave of generative AI produced thousands of products built around similar capabilities: writing assistants, document summarisation, image generation, coding assistants and conversational interfaces. Many were effectively thin application layers around the same underlying foundation models.
That market is now maturing.
At the same time, competition between the underlying models has intensified. Stanford's 2025 AI Index found that the performance gap between leading closed and open-weight models had narrowed dramatically, while differences between the highest-performing frontier models were also becoming smaller.
This creates an interesting paradox. Individual models are becoming harder to differentiate at precisely the same time that AI as a technology is becoming more useful.
Cost is part of that story. Stanford found that the inference cost of achieving performance comparable with GPT-3.5 fell by more than 280 times between November 2022 and October 2024. Smaller models have also become substantially more capable.
The implication is important: AI capability is becoming commoditised faster than AI itself is becoming saturated.
For organisations adopting AI, the competitive question increasingly becomes less about *which model is best?* and more about *what can we actually do with it?*
From chatbots to reasoning systems
One of the most important changes since 2024 has been the emergence of reasoning models.
Traditional Large Language Models largely generated responses directly from a prompt. Newer approaches allow models to expend additional computation while solving a problem, breaking difficult tasks into intermediate stages before producing an answer.
The improvement can be significant. Stanford's AI Index highlighted the dramatic performance gains produced by test-time computation on mathematical reasoning tasks.
But these advances also expose why claims that AI has somehow been "solved" are premature.
Reasoning models can be slower and considerably more computationally expensive. More importantly, impressive benchmark performance does not automatically translate into reliable performance in complex real-world environments.
Older benchmarks are themselves becoming saturated. Models now perform extremely well on tests that were considered challenging only a few years ago, forcing researchers to create substantially harder evaluations.
On some of these newer benchmarks, performance remains far below human levels.
We may therefore be approaching saturation on particular benchmarks, rather than saturation in intelligence.
The rise of agentic AI
Perhaps the biggest shift is from AI that answers questions towards AI that can take actions.
Agentic AI systems combine language models with tools, memory, planning and external systems, allowing them to perform sequences of actions rather than simply generate a response.
An AI system might analyse an inbox, identify an action, retrieve supporting information, update another system and prepare a response. In software development, agents can inspect code, make changes, run tests and iterate on failures. Similar patterns are emerging in research, customer service, cyber security and business operations.
This represents a fundamentally different proposition from the chatbot model that characterised the early generative AI boom.
Interest is already substantial. McKinsey's 2025 State of AI survey found that 62% of respondents reported that their organisations were at least experimenting with AI agents.
But experimentation and dependable autonomy are very different things.
Agents inherit the weaknesses of the models beneath them while introducing a new problem: an incorrect answer becomes an incorrect action.
That dramatically raises the importance of permissions, monitoring, validation, cyber security and human oversight.
The challenge for the next generation of AI may therefore be less about making systems capable of doing things and more about determining when we can safely allow them to do them.
The real bottleneck is increasingly organisational
Perhaps the strongest evidence against technological saturation is that organisations have still captured only a fraction of AI's potential value.
McKinsey reported in late 2025 that 88% of respondents said their organisations were regularly using AI in at least one business function. Yet nearly two-thirds had still not begun scaling AI across the enterprise.
Only 39% reported an impact on EBIT at enterprise level.
This is an important distinction.
AI adoption is widespread. AI transformation is not.
The barrier is increasingly not access to a sufficiently powerful model. It is the much less glamorous work required to integrate AI into real organisations.
That means connecting fragmented data, redesigning processes, establishing governance, integrating legacy systems, managing security, measuring performance and deciding where humans should remain responsible for decisions.
Deloitte has reported a similar pattern: organisations are achieving returns from some advanced generative AI deployments, but scaling experiments into production remains difficult.
The next stage of AI adoption therefore looks considerably less like a model race and considerably more like organisational transformation.
Reliability remains an unsolved problem
Despite remarkable improvements in capability, many of the problems identified in 2024 have not disappeared.
AI systems can still hallucinate. They can behave unpredictably when confronted with unfamiliar situations. They can confidently generate plausible but incorrect information.
The 2025 International AI Safety Report identified hallucination, failures of common-sense reasoning and failures to incorporate relevant or current context as continuing reliability problems for general-purpose AI.
These limitations become considerably more important as AI moves into consequential environments.
A hallucinated paragraph in a brainstorming exercise may be irritating. A hallucinated legal precedent, incorrect medical recommendation or autonomous change to a production system can have much more serious consequences.
The problem therefore changes as AI becomes more capable.
We increasingly need to ask not simply "How intelligent is this system?" but "How much confidence should we place in this particular output, in this particular context?"
That requires better evaluation, provenance, explainability, assurance and human oversight.
Multimodality is becoming normal
In 2024, multimodal AI still appeared to be a distinct research direction. Today it increasingly looks like the default architecture for general-purpose AI systems.
Models can work across combinations of text, images, audio and video rather than treating each medium as an isolated problem.
This matters because the real world is inherently multimodal.
People do not interact exclusively through text. Organisations contain documents, conversations, diagrams, photographs, sensor data, spreadsheets, video and databases. Systems capable of reasoning across these different information types have access to a much richer representation of the problems they are being asked to solve.
The important development is therefore not simply that AI can generate images or understand speech. It is that previously separate AI capabilities are converging into increasingly general systems.
AI is becoming cheaper, smaller and more accessible
Another reason to question the saturation argument is economics.
Frontier models remain extraordinarily expensive to train, and the computational demands of the largest systems continue to grow. However, the cost of *using* capable AI has fallen rapidly.
At the same time, smaller models have improved substantially.
This creates opportunities that were difficult to envisage during the first generative AI boom. Organisations increasingly have choices between enormous cloud-hosted frontier models and smaller models capable of running in private infrastructure, at the edge, or potentially directly on devices.
That matters particularly in environments where privacy, latency, connectivity or security make sending information to a large external model undesirable.
Rather than AI converging on a handful of enormous universal models, we may instead see a much more heterogeneous ecosystem combining frontier models, specialised models and smaller locally deployed systems.
The governance problem is getting harder
Greater capability also brings greater responsibility.
Bias, privacy, intellectual property, transparency and misuse remain significant concerns, while agentic systems introduce questions about autonomy and accountability that were less urgent when models primarily generated content for human review.
Reported AI incidents are also increasing. Stanford's AI Index recorded 233 AI-related incidents in 2024, a 56.4% increase over the previous year.
Regulation and assurance are consequently becoming part of AI engineering rather than activities performed after development.
This may prove particularly important in regulated sectors such as healthcare, defence, financial services, government and critical infrastructure, where a system being technically impressive is insufficient.
It must also be demonstrably safe, secure, auditable and appropriate for its intended use.
So, has AI reached saturation?
No. But parts of the AI ecosystem probably have.
We are arguably reaching saturation in generic AI products. We are seeing saturation in some established benchmarks. The novelty of attaching a chatbot to an existing product is disappearing, and simply claiming to "use AI" is no longer much of a differentiator.
That is healthy.
The first phase of the generative AI boom was dominated by demonstrating what these systems *could* do.
The next phase is about proving what they can do reliably, safely and economically in the real world.
The biggest opportunities may consequently be less visible than another spectacular model demonstration. They lie in integrating AI into complex workflows, combining it with trusted organisational data, designing effective human oversight, deploying smaller specialised models and building systems capable of operating reliably within constrained domains.
There is also an important economic shift underway. As foundation models become cheaper and increasingly interchangeable for many tasks, value moves upwards in the technology stack. Proprietary data, domain expertise, integration, workflow design and trust become more important.
The organisations that benefit most from AI may therefore not be those with access to the most powerful model. Increasingly, almost everyone will have access to powerful models.
The differentiator will be what they build around them.
AI has not reached its saturation point.
But the era in which simply using AI was innovative probably has.