What's after the screen

1. The afternoon the typewriter rhythm stopped

Mid-2024, I noticed something quiet on our engineering floor. One of our senior engineers was running point on a thorny Apollo GraphQL migration, and when I walked past his desk I realised he wasn't typing. He was talking to the chat pane on his second monitor, watching the model assemble a resolver, accepting some of it, refusing the rest, asking again. The IDE was open but mostly receiving. The work was happening in the conversation.

I watched the same shape repeat across the team over the next six months. By the end of 2024, the engineers shipping the most production code in our 40+ engineering organisation were the ones who had stopped using their editors the way the editors were designed to be used. The screen wasn't dead — it was still where the diff landed. But the work had moved one layer up.

That observation generalised.

Once the most click-loyal workforce in the building had stopped clicking, the only question was which knowledge-work category would notice it next.

We started rebuilding our HR platform around the assumption that the user would never see our screens at all. Eighteen months later, we shipped what seven Indian press outlets and Yahoo Finance called the world's first Headless HRMS. Forrester's analyst-relations team mapped the architecture to their published research on agentic AI and delegated execution. We launched a seven-agent contractor automation pipeline for a global energy & engineering workforce firm that now runs in 45+ countries. We launched a payroll engine that turned a 4-hour cycle into 40 minutes at a global security services group's Indonesia operation. The same client recently asked us to apply the same shape to two more domains they had nothing to do with our first build for. That is the moment a feature stops being a feature and starts being a meta-pattern.

Here is the argument behind what comes next.

2. What “after the screen” actually means

The phrase gets misread. Most people hear “after the screen” and assume the alternative is voice — that we are about to replace clicking with talking. That is not what is happening. Voice is a side effect, not the thesis.

What is actually happening is that the execution layer— the engine that takes a stated intent and turns it into a sequence of validated API calls against a backend — is becoming a first-class product surface. The interface that sits in front of the execution layer can be a screen, a chat pane, a voice assistant, an MCP-compatible AI agent owned by the customer, or no interface at all. The execution layer doesn't care. It just executes.

For three decades, a SaaS company was identified by its screens. Its “look and feel,” its UX, the workflows it baked into a series of clicks. The execution layer was an implementation detail behind the screens. The taxonomy reverses now.

The execution layer is the product. The interface is the implementation detail.

3. Why HR was the early casualty

We didn't pick HR because we were brave. We picked HR because HR is where the user-experience-meets-policy battle is the bloodiest. The same workflow has to handle a corporate director, a warehouse supervisor, and a delivery rider. Every screen is a compromise none of those three users actually likes. The director wants a clean executive dashboard. The warehouse supervisor wants the leave application not to require eleven fields he doesn't understand. The delivery rider would prefer the entire conversation happen in WhatsApp. The screen has to be three different things, and so it ends up being none of them well.

HR was a category where the screen never worked. That made it the cleanest test case for what after-the-screen would actually look like, because there was no entrenched screen-loyalty to fight.

What we found, building Zero UI: workflows that look complicated when you have to design a screen for them become simple when you don't. A leave application is, at its core, three values — who, when, why. The reason it takes six form fields in a traditional HRMS is that the form has to disambiguate every possible variant of who and when up front, because the screen has no way to ask follow-up questions. A conversational execution layer can ask the follow-up. The user types “I need to take Monday off” and the system asks “annual leave or sick leave?” — two turns instead of six fields.

The screen made the workflow look complex. Removing the screen made the workflow simple again.

This is the pattern. The categories where the screen was the work — where users wanted complex interfaces because the work itself was complex — survive. The categories where the screen fought the work — where the screen existed only because someone had to build a workflow into something — disappear.

4. Prediction one — data-entry SaaS dies first

HRMS, expense management, procurement, contract administration, internal helpdesks. Categories where the user's job is to feed the system rather than interpret what the system says. The user opens the screen, fills in fields the user already knew the values of, and clicks submit. The screen is overhead.

These die first because the interface was always the cost, not the value. The product was always the data and the workflow; the UI was the necessary tax.

The companies that win these categories rebuild the backend so the interface can be anything — including no interface at all. The companies that retrofit a chatbot on top of an existing screen pay the reconciliation cost between the chat surface and the legacy data model at every customer, every quarter, forever. They look fast for two years and then fall over on the long tail of customisation.

If your category sells primarily on UX, you are already losing — because UX was the lock-in, and once the AI assistant is the interface, the lock-in disappears.

5. Prediction two — screen-is-work categories survive, for now

Analytics, design, CAD, IDEs. Categories where exploring is the value, not the friction. A dashboard exists because the user needs to see the data in a way that lets them ask the next question. A design tool exists because the user is making something visual. A CAD program exists because the work itself is geometric.

These categories survive longest. But even they have a five-year horizon. Vibe coding has already started to unbundle the IDE. Analytics dashboards are getting their first wave of ask-the-data assistants that bypass the dashboard entirely — “what did revenue do last quarter in APAC, excluding the one-time deal” instead of three filter clicks and a saved view. Design tools have AI-assisted draft generation that operates one layer up from the canvas.

The screens hold longest where the work is intrinsically visual or exploratory. They will not hold forever.

6. Prediction three — the AI assistant becomes the new procurement decision

This is the prediction that catches enterprise procurement off-guard.

Today, procurement asks which CRM, which HRMS, which ERP. The decision is vendor-by-category. Each category has its own RFP, its own integration plan, its own change-management cycle.

In five years, the more important question is which AI assistant the customer's people already use — Claude, ChatGPT, an internal copilot, a customer-built agent. That assistant becomes the surface across every backend system the customer owns. The customer's people don't log into the HRMS to file a leave application; they ask their AI assistant to do it. They don't log into the procurement system to raise a purchase order; they ask their AI assistant. The vendor is invisible in the user's day, present only in the execution layer the assistant calls into.

Vendors that publish a clean, scoped, governed interface to AI assistants — through MCP or whatever standard wins — get carried by whichever assistant the customer prefers. Vendors that don't get bypassed by a thin layer the customer's team builds in-house in a weekend, on top of the vendor's REST API, with the customer's preferred AI assistant on top.

The new lock-in is not the screen the user sees. It is the execution layer the assistant is willing to call into.

This changes how vendors compete. The CRM that wins is not the one with the prettiest dashboard; it is the one whose AI-assistant interface is the easiest to govern, scope, audit, and trust. The HRMS that wins is not the one with the most modules; it is the one whose tool catalog the customer's compliance team is willing to expose to an external AI without rebuilding it from scratch.

7. The honest caveat — procurement lags, and the migration is the moat

None of this means the screen disappears tomorrow. Enterprise procurement is famously backward-looking. Committees buy on integrations, certifications, named system-integrator references, and total-cost-of-ownership models that point to the previous architecture, not the next one. Architecture-led vendors have to do the unglamorous work of building procurement-legible references against forward-looking products. We are spending more time on that than I expected to.

The other caveat is the harder one: the migration is the moat. Anyone can build a parallel AI surface in days. A competitor with a clean sheet can clone the conversational experience over a weekend. What they cannot clone is the year of work it takes to carry an existing customer base of decade-old configurations and tenant-specific workflows onto the new architecture. The build is the demo. The migration is the business.

We are about ten percent of the way through that migration ourselves. The next ten percent is in flight. We expect the work to take another two years even with the platform substrate already built. That is the part nobody wants to talk about, and what actually determines who wins each category over the next five years.

The build is the demo. The migration is the business.

8. The closing line

Ten years ago the question was which SaaS vendor you bought. Five years from now the question will be which AI assistant you let buy on your behalf, and which vendors that assistant trusts to act on your data.

The vendors that survive will be the ones who realised, early, that they were not selling screens. They were selling execution.