Trust in the new era of AI window shopping
AI is now smart enough for most business admin, and the frontier labs are no longer the only option. Cheaper, open and self-hosted models are changing how companies should buy AI, and raising the stakes on how they control it.
Frontier AI labs often grab headlines by racing to be the most intelligent. For most companies, however, that race stopped mattering a while ago. Productivity in the workplace is often best leveraged from the automation of routine office work.[1] Having a model that can (allegedly) solve Millennium Problems[2] performing those tasks is overkill. Having a model that can perform those tasks to a high degree of efficiency and accuracy in a cost-effective manner sounds like the much more attractive option.
Enterprise leaders asking about costs and more diverse questions around deployment, data sovereignty and personalisation are doing so at the right time. A growing field of open-weight, regional and specialist models are catching up in intelligence for a fraction of the cost,[3][4] can be run on infrastructure you control and post-trained to be personalised for your company. Companies like Revolut are getting the most from AI by breaking out of the single-vendor habit and picking the right model for each job, and even in-housing their own.[4][5]
But more models mean more doors to guard, and a new generation of agentic attacks shows what happens when they’re left open. Making the most of AI choice now depends on being able to trust the tools that you are onboarding.
The playing field has opened up
For years, the honest answer to “can AI do this job?” was “parts of it, badly”. Within half a decade of ChatGPT’s first release,[6] that has changed, with many jobs now looking completely transformed by the use of AI. Most productivity gains from AI so far have tended to be from automating basic tasks. Reading documents, extracting data, checking it against rules, drafting code and updating data between systems. For these tasks there is a minimum level of intelligence needed, and the leading models cleared it some time ago.[1] Past that line, extra capability adds less than lower cost, more control or a better fit for the job.
Enterprise leaders that are starting to reckon with that and other questions around AI deployment, sovereignty and personalisation have chosen to do so at a good time. Frontier labs are no longer the only route to intelligence, with competitive models now being released from a dozen labs across the world.[3] Many are open-weight, meaning that you can download them, run them on your own infrastructure and post-train them to adapt them to your own business. Of course, this means investing in infrastructure and upkeep, but instead of getting a monthly bill from your AI vendor, now those AI tokens only cost the electricity used to run your own machines.
Many enterprises have noticed. A September 2026 Enterprise Technology Research survey found open-weight models now handle 34% of enterprise AI token usage, up from 23% a year earlier.[7] UK and Europe’s more digitally advanced neo-banks are a great example of when it works. Bunq says its assistant now runs on open-weight models at 60% of the cost per answer of leading closed models,[4] while Starling and Revolut have trained and are hosting their own banking models on UK infrastructure.[4][5]
Breaking the single vendor habit
It isn’t just the models themselves that are a source of competitive advantage. Harnesses sitting on top of foundation models have defined a category of purpose-built tools that are really good at doing a certain type of job. Cognition’s Devin provides an agentic engineering team that sits upon a set of models, both their own and others, that are dynamically routed to when the time is right.[8] Harvey and Legora have built their platforms specifically for legal functions.[9] BNY have built Eliza, a model-agnostic platform for their employees to use across client services and internal operations.[10]
Taking advantage of enough of the right tools means changing the mindset behind how most companies buy AI. Today, the default is still to go with one vendor with one enterprise agreement, chosen once and applied to everything. That made sense when only a handful of models were good enough for the job. It makes a lot less now.
Truthfully, most companies won’t build their own harnesses and even fewer will play around with self-hosting their own AI models. But the principle travels. AI models and tools should be seen as a portfolio. Use frontier models where their extra reasoning earns its price, cheaper or open models for high-volume routine work, and your own adapted models where your data and / or operating model is your competitive advantage.
More choice, more doors to guard
Every new model is also a new door into your business. Each brings its own terms, data handling, authentication and failure modes. And once those models become agents that read your company’s sensitive information and have access to your applications, the stakes rise sharply. This year showed what that looks like.
In July, Hugging Face disclosed an intrusion that turned out to come from OpenAI’s own pre-release models.[11] In May, the Mini Shai-Hulud worm came from the other direction, compromising hundreds of npm and PyPI packages, and writing itself into AI coding assistants’ settings to replicate those every time a developer opened a project.[12][13]
Agents can go beyond their brief, and they can be turned against you, both at machine speed and with whatever access they hold. Multiply that by every model in your portfolio and the risk is no longer one vendor’s problem. It’s yours.
Trust turns a gamble into a decision
The companies that benefit most from AI choice will be the ones that can govern it. That means being able to answer four questions at any moment, across every model and agent: what is running, what can each reach, what has each done, and how do we stop any of them?
If you can answer those, adding a new model is a decision. If you can’t, every new model is a gamble, and the safe choice looks like staying with one vendor and missing out.
For those leading technology, operations or risk, I challenge you to list every model and AI tool in use across your organisation today, and who chose it. Then ask which of the four questions you could answer for each. Most teams find their portfolio is bigger than they thought and their answers thinner.
This is why we built Mantle. Mantle is a governance layer that sits between your organisation’s data and your AI tools, providing you with real-time visibility and control over the agents in your system. Set your policies once: which models are approved, what data they may see, what each agent can do and what needs a human to sign off. Mantle applies those rules in real time across frontier, open and self-hosted models, keeps a full record of what every agent did, and lets you switch any of them off. Built and run in the UK for organisations that want full sovereignty over where their AI runs.
If you’re interested to see how this would work for your organisation, feel free to reach out directly or book a conversation at elvarsystems.co.uk. Next in this series: where does your company sit on the AI adoption frontier, and what is holding it back?