Does Microsoft still have no taste? Copilot is the first product where it costs them

In 1996, in the PBS documentary Triumph of the Nerds, Steve Jobs delivered the most quoted verdict ever passed on Microsoft:

"The only problem with Microsoft is they just have no taste."

He was specific about what he meant. His complaint was not that Microsoft was bad at business. It was that Microsoft did not think of original ideas and did not bring culture into its products, and that it won anyway. Microsoft's genius was never the product. It was distribution.

Thirty years later, Microsoft has made the biggest product call of its AI era. On 25 September it rebuilt Copilot as a single app combining chat, coding and autonomous agents, and moved the most advanced parts to usage-based pricing. CNBC framed it as an attempt to compete with Anthropic.

Jobs's question has never been more relevant, because for the first time distribution alone is visibly not working. Fewer than 7% of more than 450 million commercial Microsoft 365 seats carry a Copilot licence. Microsoft owns the desktop, the inbox, the identity layer and the procurement relationship, and it still cannot get most of its own customers to pay for AI.

That gap matters to anyone responsible for security and risk. The fix Microsoft is reaching for puts autonomous agents inside Teams and Outlook, billed by consumption.

Thirty years of winning on distribution

Microsoft's record splits cleanly. It wins where it controls the default and "good enough" is enough. It loses where users actively choose, or where it arrives late to someone else's platform.

The wins are almost all utilities: an operating system, a browser, a word processor, a chat client. Nobody chooses these for delight. They use whatever is already installed, and Microsoft made sure it was.

The losses are products where the user picks. A phone, a music player, a search engine, an assistant you have to want to talk to. There, placement bought Microsoft a trial, never loyalty. That is exactly the distinction Jobs was drawing.

Nadella fixed the strategy, not the taste

Satya Nadella, chief executive since 2014, changed Microsoft more than anyone expected. He dropped the insistence that everything run on Windows, shipped Office for the iPad within weeks of taking over, and bet the company on Azure. Azure is now one of the two dominant enterprise clouds.

His acquisitions were disciplined. LinkedIn and GitHub were bought and largely left to run as they were, which is why developers did not flee GitHub the way many predicted. The early investment in OpenAI gave Microsoft a lead in generative AI that no other incumbent had in 2023.

None of that is taste in Jobs's sense. It is strategy: picking the right markets, buying well and partnering early. When Microsoft had to build a consumer-facing AI product itself, the old pattern came back. Bing Chat became Copilot, Copilot split into consumer, Microsoft 365 and GitHub versions, and the branding was reshuffled repeatedly. Customers were rarely sure which Copilot they were being sold.

The latest restructure is organisational as well as product. GeekWire reports that three weeks before the launch, Microsoft reorganised its financial reporting into a new Agents and Infra segment and moved GitHub's cloud revenue into its Microsoft 365 business. Nadella described the goal as connecting the entire trajectory of a "job to be done". That is a coherent idea. Whether it produces a product people want is a different question.

The Copilot test

Copilot is the cleanest test of the distribution thesis Microsoft has ever run, and so far it is failing. Microsoft reported more than 30 million paid Copilot seats in July, per Yahoo Finance. Against a base of more than 450 million commercial seats, that is under 7%, nearly three years after Microsoft 365 Copilot went on sale at US$30 per user per month.

Licensed seats overstate real use. CNBC quotes David McLaren of engineering consultancy GHD: his firm has given Copilot to 12,000 employees, and around 200 are experimenting with Cowork, the agent that handles more complex work. That is under 2% of licensed users touching the capability Microsoft is now leading with.

The new app is Microsoft's answer. It collapses the separate products into three tabs:

Two details in CNBC's reporting say more than the launch itself. First, Microsoft's Copilot chief Jacob Andreou conceded on the record that adding coding is Microsoft playing catch-up. Second, in the Home tab, where most productivity work will happen, no Microsoft model is available at all. Code users can choose a cheaper Microsoft model; everyone else runs on OpenAI or Anthropic.

So the product Microsoft is using to compete with Anthropic runs substantially on models from Anthropic and OpenAI. Microsoft is supplying the distribution, the packaging and the billing. Others are supplying the quality.

Why bundling does not work for agents

Bundling works when the user cannot easily tell the difference between products, or does not care. Nobody benchmarks their chat client. For AI agents the output is the product, and every user compares it within minutes. A worse draft, a wrong formula or a clumsy summary is visible immediately, and the alternative is one browser tab away.

Microsoft appears to know this. Andreou told CNBC the product is at its best when it optimises for user value and output quality, and that cost efficiency matters less. That is an unusual thing for Microsoft to say about a product it plans to win on distribution. It is an admission that placement will not carry this one.

The strategy Microsoft is actually pursuing is different, and it is coherent. Nadella said at the preview that the Frontier program is not about any one frontier model but about an ecosystem. Read plainly, Microsoft wants to be the neutral layer: the app, identity, data access, compliance and billing that sit above whichever model is best this quarter. That is the Azure playbook applied to the desktop, and it has worked before.

The risk is that the layer is thinner than it looks. Anthropic and OpenAI sell their own agents directly into the same enterprises. If the model is the quality and the model vendor also ships the app, Microsoft's role narrows to procurement convenience and governance.

There is also a sense in which taste matters more for agents than for anything Microsoft has shipped before. An agent's quality is not only its model. It is judgement about defaults: when to act, when to ask, what to show the user, when to stop. Those are design decisions, and they decide whether people trust an agent enough to delegate. Microsoft is licensing the intelligence. It still has to supply the taste.

What this means if you own the risk

Andreou named the real obstacle himself at the Seattle preview:

"The problem is that these agents have been almost impossible to bring into the enterprise."

He is right, and the reasons are not about taste. They are identity, permissions, data boundaries, audit and accountability. The new app raises the stakes on each of them, and CNBC's reporting does not ask how Microsoft addresses any of them.

Agents are new identities with delegated authority. Autopilot agents will take assignments and communicate in Teams and Outlook. Each is a non-human actor that can send messages, touch documents and trigger work in someone's name. Our recent analysis of the Medicare portal incident showed what happens when an agent meets a control it does not like and finds another way through. An always-on agent inside your mail system needs the same scrutiny as a privileged service account, and usually gets less.

Copilot sees what your permissions allow, not what you intended. Copilot answers from whatever the user can technically reach. Years of loose SharePoint and Teams sharing become searchable in seconds. Agents that act continuously turn that exposure from a query problem into a standing one.

The model behind the answer is a fourth party. The Home tab runs only on OpenAI and Anthropic models. For APRA-regulated entities, CPS 230 and CPS 234 expect you to understand material service providers and the information security of the parties behind them. "Microsoft" is no longer a complete answer to where your data is processed and under whose terms.

Usage pricing makes cost an operational risk. A seat licence has a ceiling. Consumption billing for always-on agents does not, unless you set one. A misconfigured or looping agent is now a budget incident as well as a security one.

This is also where the counterargument lives, and it is a strong one. Microsoft already runs identity for most Australian enterprises through Entra, data classification through Purview, and a cloud that has been assessed through IRAP. If agent adoption stalls on governance rather than on quality, the vendor that already holds the control plane has a real advantage. Regulated buyers do not pick the best-tasting product. They pick the one they can defend to the regulator. In that market, distribution and governance may beat taste again.

So, does Microsoft still have no taste?

On the evidence of this launch, yes. The new Copilot is a packaging and pricing decision, built on other companies' models, from a vendor that admits it is catching up. That is not a criticism of the strategy. It may well be the right one. But it is not taste.

The more useful question is whether taste decides this market. If enterprise AI is won on output quality, Microsoft is a reseller with excellent distribution and a shrinking margin. If it is won on governance, on which vendor can make agents safe enough to switch on, Microsoft holds cards nobody else does. Jobs was right about Microsoft for thirty years and it did not matter. This may be the first time it does.

What to watch over the next two quarters: how many Frontier customers move Cowork beyond pilot groups, whether a Microsoft model ever appears in the Home tab, and whether usage billing ships with controls a finance team can live with.

If you are evaluating the new app, these are the questions worth putting to your Microsoft account team before anything is switched on:

1. For each tab, which model processes our data, in which jurisdiction, and under whose contractual terms? Can we restrict it?
2. How is each Autopilot agent identified, permissioned and logged? Can we require human approval before it communicates outside the organisation?
3. What spend controls exist under usage billing: hard caps, per-agent budgets, alerts?
4. What should we fix in our SharePoint and Teams permissions before enabling Cowork or Autopilot, and how do we measure that we have?

None of those need a view on Microsoft's taste. They need answers.

And a question for readers: if you run Microsoft 365 in a regulated environment, is your organisation's AI adoption being held back by product quality or by governance? The answer decides which Microsoft wins.

---

Sources: Microsoft packages business AI in single app as it tries to compete with Anthropic, Jordan Novet, CNBC, 25 September 2026. Microsoft unveils all-in-one Copilot app, GeekWire, 25 September 2026. Microsoft relaunches Copilot app with coding and AI agents, Yahoo Finance, 25 September 2026. Steve Jobs quotation from Triumph of the Nerds, PBS, 1996. Figures are as reported by those outlets.