Microsoft wants to tune AI on your data

On 2 June 2026, Microsoft AI launched seven in-house models and a new idea called Frontier Tuning. Mustafa Suleyman framed it as the next phase of AI. The models matter less than the pitch behind them. And the pitch deserves a careful read before any UK SME gets excited.

Three giants, three different games

For a while the contest looked simple. Whose model scores highest. That framing is now out of date.

Watch what each company is actually doing. OpenAI and Anthropic are moving outward, into your workspace. Cowork and similar tools let AI act across your documents and apps. Claude and ChatGPT now work inside Excel, Word and PowerPoint.

Microsoft is moving the other way. Inward. Frontier Tuning trains a model on your own work, inside your environment, and the model stays yours. Microsoft describes the valuable data as the trace of real work your people complete each day.

So the choice for an SME is no longer which model is best. It is which bet you want to make. Expand what AI can reach, or teach a model your specific way of working. Both are reasonable. Neither is free of consequences.

The promise underneath Frontier Tuning

The idea is genuinely interesting. A model that learns your processes could capture knowledge that usually lives in people\u2019s heads.

Microsoft offers proof points. A tuned model for Excel matched GPT 5.4 at up to 10 times the efficiency. Tuned to a market-leading organisation\u2019s standards, it achieved the highest win rate of any model tested.

Read that second claim again. It is the most revealing line in the whole announcement.

The quiet assumption

Frontier Tuning works by learning from your data. That only helps if your data is worth learning from.

Here is the gap. Most SMEs do not have clean, current, well-organised data. They have old files, duplicate documents, half-abandoned shared drives, and knowledge stuck in a few key people. Tuning a model on that does not capture how you work well. It captures the mess, and gives it confidence.

Now look again at Microsoft\u2019s headline customer. A market-leading organisation with exacting enterprise standards. These are firms whose actual business is disciplined process and immaculate documentation. Of course tuning works there. Their data was already in order.

That tells a 30-person firm very little. The proof point reveals the assumption. Frontier Tuning shines when your data is already excellent. For most SMEs, it is not.

The honest sequence

This is not an argument against Microsoft. It is an argument about order.

AI adoption is 80% people and culture, 20% technology. The technology question, which model, comes late. The first questions are about your foundations. What do you actually know. Where does it live. Can anyone find it.

At gecco we call these the Five Core Documents. Your business overview, brand, audiences, services, and search positioning. Getting them right is unglamorous work. It is also the work that makes any AI useful, tuned or not.

Tune a model before that work is done, and you automate your confusion. Do the work first, and the model has something good to learn from.

A more flexible path

There is a second reason to be cautious about tuning too early. Once knowledge is baked into a models weights, you cannot easily inspect or correct it.

A general model offers a different shape. Pair a capable model with two things. Skills that encode your good process in plain language. Connectors that give the model live access to your real systems.

This keeps the knowledge visible. When your data is wrong, you fix the source, not the model. When a process changes, you edit a Skill, not a training run. For a business with imperfect data, that openness is a feature, not a compromise.

What to do this week

You do not need to pick a side today. You need to know where you stand.

• Audit your data honestly. Is it current, organised and findable, or scattered and stale.
• Write down one important process that lives only in someone\u2019s head.
• Pick one repetitive task and test a general model with a Skill, before considering anything bespoke.
• Resist the urge to tune. Earn it by sorting your foundations first.

Where this leaves you

Microsoft has built something serious, and the ownership idea is sound. The catch is the precondition. Tuning rewards organisations that already have their house in order.

gecco helps UK SMEs get there. We sort your foundations, build Assistants on the platform that suits you, and connect them to your real systems through Automations. We stay platform-agnostic, so the advice serves you, not a vendor.

The giants are fighting a loud battle over models and frontiers. Most SMEs will win or lose on something quieter. Whether their own data is worth learning from.