The AI Landscape for Executives: What Every Leader Needs to Know

David Turner
3 days ago
5 min read

The AI Landscape for Executives

Boards asking for an AI strategy, often receive a vendor presentation dressed in strategy's clothing.

The vocabulary is being set by people with a commercial interest in your confusion. AI is not that complicated. The industry around it is. What follows is the version of this conversation that the hyperbolic presentations skip. Here's the AI landscape for executives.

A short history that actually explains the present

Machine learning has been running inside your business for longer than the current conversation implies. Spam filters, credit scoring, fraud detection, product recommendations - these are all ML applications, most in production for fifteen to twenty years. A model trains on historical data, learns patterns, and applies them to new inputs. The current generation of tools does the same at greater scale, albeit with one significant threshold crossed in November 2022.

ChatGPT did not invent a new technology. It made a general-purpose language model accessible to anyone with a browser, at a level of usefulness that collapsed the gap between expert and non-expert overnight. Executives heard about it from their teenagers before their IT departments briefed them. That tells you something about the adoption speed you are now managing.

Machine learning became "AI" in the public conversation because generative AI needed a shorter handle, and because the companies involved understood that AI carries cultural weight that ML does not. They are not wrong that something shifted. They are selective about how much is new.

What you are actually buying

The major providers - OpenAI, Anthropic, Google, Microsoft, Meta, Mistral, DeepSeek et al - offer broadly similar capabilities at broadly similar price points. Microsoft's Copilot is OpenAI's models packaged inside Office. DeepSeek, in early 2025, matched the leading American models at a fraction of the reported training cost, repriced AI infrastructure stocks sharply, and forced a more honest conversation about capital expenditure across the sector.

A subscription pays for two things: compute (measured in tokens — roughly three quarters of a word each) and features (memory, file handling, web search, and the ability to take actions on your behalf). A web interface is a product. An API call is your engineers talking directly to the model, paying per token, building something. Different activities, different cost profiles, different governance requirements. Most organisations are running both without a clear policy on either.

Where the older ML still does the serious work

The applications with the highest stakes - medical diagnostics, drug discovery, climate modelling - run on purpose-built ML systems trained on proprietary datasets and validated against external benchmarks. AlphaFold, DeepMind's protein structure prediction system, predicts how molecules fold and is now a standard tool in pharmaceutical research. It does not write emails. These systems do not appear in vendor presentations. They are running in the background of decisions that will affect your sector in five to ten years.

You are navigating a market in its pricing honeymoon

None of the major AI providers is profitable at the model level. OpenAI, Anthropic, and Google are subsidising current pricing with investor capital or profitable adjacent businesses. Broadband, cloud compute, and mobile data all ran below cost during adoption phases. What is different here is the speed of embedding. Businesses are building workflows, retraining staff, and restructuring processes around specific providers while those providers are still working out what the sustainable price is.

This is the lens that belongs underneath every AI procurement decision. Not "what can this do" - every vendor will answer that enthusiastically - but "what does our exposure look like when pricing normalises." Token costs will rise. Subscription tiers will be restructured. The organisations that have mapped their dependency will manage the transition. The ones that have not will find out about it in a budget meeting.

Specific supply chain risks exist here also. Anthropic's Mythos model was deemed too high a risk in cyber-security terms to be released to all but a handful of organisations. The subsequent 'watered-down' version of that model, Fable 5, was released to the public and then retracted for all foreign nationals (i.e. outside America) after the US Government deemed it excessively dangerous. This means your competitors in the USA now hold a capability advantage*

*Since re-released extra-territorially at the end of June 2026

The architecture your engineers are building toward

Your engineering team is talking about agents, MCP, and Claude Code because the technology is moving from answering questions to taking actions.

An agent pursues a goal across multiple steps rather than responding to a single prompt. A subagent handles a discrete task inside a larger automated workflow. MCP - Model Context Protocol - is a standard for connecting agents to external tools and data sources: a common interface that removes custom integration at every connection point. Skills and tools are discrete capabilities given to a model - web search, code execution, file reading, API calls. Projects and Gems are persistent workspaces where a model holds its configuration and context across sessions.

Claude Code reads and writes across an entire codebase at once rather than file by file. For a development team, certain categories of work compress significantly. That is why you are hearing about it.

Getting different companies' agents to reliably exchange context, handle errors, and maintain state is not a solved problem. MCP is a step toward it. Any business considering AI-to-AI workflows across supplier relationships or between divisions running different platforms should treat integration complexity as a live risk, not a roadmap item.

What it costs when this goes wrong

The risks split into two broad categories: recoverable and not.

A poor tool choice, a failed pilot, a wasted implementation - recoverable.

However consider a lock-in at the workflow level. A business that has embedded a specific provider's agent architecture into core operations and then faces say a 40 or 50% rise in token costs or a change in data handling terms has limited options and limited time. The switching cost is not the subscription. It is the retraining, the re-integration, and the disruption.

Deploying a general-purpose model where a purpose-built one is required can equally be disastrous. ADAS applications in vehicles are the clearest case: a system making real-time braking decisions cannot be cloud-dependent, cannot tolerate variable latency, and cannot be optimised for general usefulness rather than deterministic safety. Getting this wrong is not a failed pilot. It is a regulatory and reputational event.

Three things before the next vendor meeting

Map your current exposure. List every AI tool in use, who provides it, what data it touches, and what breaks if the price doubles or the provider changes terms. Many organisations cannot answer this. That is the first problem.

Set an interoperability requirement before committing at infrastructure level. Any system adopted centrally should pass two questions: can we connect it to other systems, and can we leave it. Lock-in at the workflow level is more dangerous than lock-in at the tool level.

Separate the language model conversation from the ML conversation in your governance. Your data science team building predictive models and your operations team deploying a customer-facing assistant are doing different things with different risk profiles. A framework that treats them identically will be wrong for both.

The close

The vendors are not lying about what the technology can do. They are, for understandable commercial reasons, quiet about what it will cost when the subsidised period ends. That silence is not neutral. It is the most important thing missing from every AI strategy deck you have seen this year.

David Turner is the founder of Kói, an independent technology consultancy advising investors, founders, and boards on AI strategy, technology risk, and investment readiness.