AI is More Than a Chatbot

In the first week of July, Jacob Andreou, the executive Microsoft put in charge of Copilot in March, sent a memo to the 11,000 people who build it. The Information reported it. Copilot has to earn the right to exist.

The rest of the memo is even more revealing. Andreou wrote that the sprawl of Copilot features had confused users and become an embarrassing problem internally. He wrote that enterprise buyers were now scrutinising the return on their AI spend, and that the bar had risen across enterprise software. And he told the team to focus on real work and outcomes rather than chase intelligence for its own sake.

Microsoft is paying the first-mover tax. It moved early to diffuse this technology into real enterprise products, and got the diffusion model wrong. Bolting a chat interface onto an existing product was never the answer. A sidebar that summarises the thing you are already looking at is not a productivity revolution. It is a feature. A small one. And it lives or dies on whether the user remembers to open it. A chatbot answers. It does not act.

Despite millions of dollars in advertisement spend. By early 2026, fewer than 4.5% of Microsoft 365's paid commercial base had taken the Copilot add-on, the $30 a seat licence that lets the assistant into your mail, your files and your meetings. Other estimates put it as low as 3.3%, and of those, only a fifth to a third open it in any given week. Real weekly use sits close to 1% of the customer base.

Real diffusion means integration into the workflow, not adjacent to it. It means automating the work that costs the most, not the work that is easiest to reach, and most of that work needs no interface at all. It needs headless agents: no screen, no prompt, triggered by an event and running to completion. 

Where a person does need to see something, it means moving past static screens to interfaces that assemble themselves around the job, what is known as Generative UI. 

And it means systems that act rather than systems that answer. A bid lands. The system reads the specification, the pricing schedule and the supplier's compliance history, applies the rules the organisation actually runs on including the exceptions nobody ever wrote down, does the work, and stops for a human-in-the-loop only where the decision is a real judgement call.

On top of this, the models are commoditising. Moonshot released the weights for Kimi K3 on Monday, a 2.8 trillion parameter open model that ranks third on Artificial Analysis's intelligence index behind only Claude Fable and GPT-5.6 Sol Max. The gap between the top foundation models narrows every quarter. The gap between how they are implemented inside the enterprise widens. The moat is the business itself, and the winners will not be the ones with the best model access — they will be the ones willing to rebuild the work around what the technology can now do.

For every other company building or buying enterprise AI in 2026, the lesson is simple. Stop bolting. Start rebuilding. Treat AI as a feature on top of how you already work and you will spend the next three years explaining your adoption numbers. Treat it as a reason to redesign the work itself and you get the productivity everybody has spent three years promising.

This is the work we do at Elsewhen. We build agentic AI systems that sit inside the workflows that run a business, grounded in your own data and rules, with people on the judgement calls. We take one workflow if that is where the cost sits, or the whole operation if you are ready for it. You own what we build, IP included, and it is model agnostic, so the next model plugs in instead of forcing a rebuild. Intelligence you own, not software you rent.

But most importantly, a lot more than just another chatbot.