The Screen Is Optional Now

Within ten days, two of the biggest business-software vendors said the same thing: the screen is no longer the product. On 15 September Salesforce unveiled AIforce, an interface that AI assembles on request over your governed data. On 25 September Microsoft rebuilt Copilot around three tabs, Home, Code, and Autopilot, so that staff describe the app they need and an agent keeps working after they log off. The decision this forces is not which vendor to back. It is which parts of your work should still have a fixed screen at all, and for most companies the honest answer is fewer than today, but more than the demos suggest.

We build on both platforms and we are a partner to neither, so this is not a pitch for one roadmap over the other. It is also early. Most of what Microsoft announced is in preview or early access, and several of the Salesforce pieces are in beta or pilot, so read everything below as direction rather than delivery. Product facts come from Microsoft's official blog and Salesforce's newsroom, as published on the dates given.

What did Microsoft announce on 25 September?

Per Jared Spataro's post on the Official Microsoft Blog, Copilot now has three parts. Home is the new starting point where Chat and Cowork come together, with Word, Excel, and PowerPoint working inside it. Code lets anyone describe an app, tracker, dashboard, automation, or workflow in plain language and have Copilot build it. The result runs in a sandbox, can be hosted inside your own Microsoft 365 tenant, uses the same underlying technology as GitHub Copilot, and runs on a new Microsoft Copilot Managed Runtime that IT governs. Autopilot is an agent you give a name, a role, and a goal. It is cloud-hosted, keeps working when you are not, and lives in your tenant with its own identity, memory, and workspace, with permissions, audit, and governance behind it. Copilot is also being grounded in Dynamics 365 and Power Platform data, in public preview over the next month. Satya Nadella described the update as building Copilot into "a new OS for work that spans every model, every form factor, and every task."

The detail I would put in front of a CFO is the billing split. A per-user license still covers Copilot in Chat and across Word, Excel, PowerPoint, Outlook, and Teams. Cowork, Code, Autopilot, and frontier models such as Astra and Fable run on usage-based billing. For administrators, that makes cost control and governance a more urgent question this month than the new tabs.

And the honesty note: Home and Code roll out to Microsoft's Frontier early-access program over the coming weeks, and Autopilot expands to private preview at the end of September. None of it is generally available yet.

What did Salesforce announce ten days earlier?

At the Dreamforce keynote on 15 September, Salesforce unveiled AIforce, which it calls "a live interface layer that brings the full power of Salesforce to wherever people and agents work." Salesforce says every request runs on existing permissions and business rules with Zero Data Retention. Underneath sits Koa, a CRM reasoning model Salesforce runs inside its own infrastructure, currently in pilot.

Two earlier pieces complete the picture. Slack Surfaces, announced on 11 September, builds dashboards, reports, decks, calculators, and microsites from a prompt, under the permissions of the person asking. Headless 360, released in April, exposes Salesforce capabilities through APIs and MCP servers without the Salesforce UI. We covered each in detail in our Dreamforce day-one read, our Surfaces explainer, and our Headless 360 piece for architects.

How do the two compare, side by side?

Read the last row twice. Almost nothing on either side is generally available today, and the pricing on both is either new or not yet set. That is normal for announcements this size, and it is exactly why this is the right moment to decide your position rather than your purchase. For the Salesforce side in full, our Dreamforce 2026 recap lists every announcement with its status.

What is the shift underneath both announcements?

Strip the branding and the two companies are making the same four moves:

1. Screens are generated on demand. Code and Surfaces build the view when someone asks for it, and AIforce composes it at request time. The screen becomes an output rather than the product.
2. Work is delegated to agents that outlast the session. Autopilot and Cowork on one side, Agentforce on the other: you hand over a goal instead of a series of clicks, and the work continues after you close the laptop.
3. The platform becomes a permissioned data-and-runtime layer. Microsoft runs your generated code and your agents inside your tenant under IT control. Salesforce puts Data 360, the permission model, and Headless 360 underneath whatever interface gets drawn. What you are buying is less an application and more a governed place where data, identity, and code run.
4. Cost becomes usage. Microsoft's agentic features bill by use, and Agentforce is priced on consumption. The license stops being the ceiling on your spend.

The reason both companies are reorganizing around data and permissions is simple: that is what each of them uniquely holds. Anyone can generate a dashboard. Only Microsoft holds your tenant's documents, mail, and identity, and only Salesforce holds your customer records and the rules about who may see them. The interface was never the real moat. The data and the permission model were, and now both vendors are saying so out loud.

Where should a fixed screen stay?

Where the work is frequent, high-stakes, or regulated. A screen that never changes sounds old-fashioned until you look at what it does well. People who do the same task forty times a day are faster on a screen they have learned than by describing the task each time. A fixed screen shows the same inputs the same way to everyone, which is how you get consistent decisions. And when an auditor or a regulator asks what the approver saw before they approved, a fixed screen with an audit trail has an answer; a view generated on the fly may not.

In practice that means underwriting and approval steps, payment releases, claims and case decisions, safety and compliance checklists, and anything you may one day be asked to replay. I would not generate those screens on demand in 2026, however good the demo looks.

Where do generated interfaces and agents pay off?

Where the work is occasional, analytical, or one-off. The board-prep dashboard someone needs twice a year, the tracker for a three-week project, the one-time pull that answers a single question, and the follow-ups that fall between threads are all work nobody should wait in a queue for. Microsoft's own example for Autopilot, running a supplier review from schedule to follow-ups, is a fair picture of the delegation side: bounded, recurring, and annoying rather than high-risk.

If you run Power Platform today, Code is the question your makers will ask first, and the judgment in our Power Platform vs custom development guide still applies: generate what is disposable, and build properly what the business will depend on.

What goes wrong on each side?

Both directions fail in predictable ways.

• Prompt-built sprawl. Hundreds of one-off apps and dashboards, each slightly different, none owned, some quietly wrong. It is spreadsheet sprawl with a better interface. The fix is not a ban; it is deciding which numbers must come from a governed source, the same question our Surfaces explainer raised about dashboards built from a prompt.
• Agents acting without evals or logs. An agent that keeps working while you sleep can also keep being wrong while you sleep. Before one runs unattended, it needs evals on your real cases, a log of every action it takes, and a clear line for what it may do without a human. Our piece on the agent harness pattern covers how to build that discipline in, and our AI governance policy guide covers who signs off.

Keeping every screen fixed fails too, more slowly: your people keep waiting two weeks for a report that a prompt could have built in a minute, and the shadow tools arrive anyway.

What does a balance audit look like?

Five steps, in this order:

1. Inventory the interfaces people use daily: the screens, reports, and dashboards, plus the spreadsheets that fill the gaps between them.
2. Classify each workflow by frequency, stakes, and regulation. Most land clearly on one side; the few that do not are where the real conversation is.
3. Decide keep, generate, or delegate for each one: keep the fixed screen, let people generate the view, or hand the work to an agent.
4. Set the permission and evals bar before anything generated or delegated touches real data: who may see what, what an agent may do alone, and how you will know when it is wrong.
5. Model the cost under usage billing before switching agents on, at realistic volumes, because usage pricing is cheap to start and expensive to be surprised by.

This is the audit we run with clients before they commit to either roadmap.

The screen is optional now, but optional is not obsolete. The companies that get this right will keep a small number of fixed screens for the work that has to be fast, consistent, and auditable, and let everything else be generated or delegated under permissions they actually understand. Everything above is as published on 15 and 25 September 2026 and worth re-checking before you plan around it. If you want help drawing that line across your own stack, our AI agents and Salesforce development teams do exactly this work, and you can talk to us when you are ready.