The AI Arc — four horizons, one surface
AI arrived in waves — classic ML, then retrieval, then generative — and each wave redefined what the product was for. As features multiplied they all competed for the same thing: an agent’s attention, mid-call, with a customer waiting. The risk was fragmenting the experience all over again, one AI feature at a time.
The exploration that inverts the product. If Horizon 3 put the Assistant inside the interface, Horizon 4 asks what happens when the Assistant is the interface — the traditional desktop demoted to a data layer, available when needed. Here that idea is taken to a device: state, sentiment and handling time glanceable, everything else spoken. Not shipped, and deliberately so — explorations like this run ahead of the roadmap to keep pace with what the technology is about to make possible.
Above: the Assistant as it shipped. Rather than each new AI capability inventing its own placement and personality, they were consolidated into one surface that reads the live conversation and proposes a response — with its source one click away, and identity verification handed over as a card the agent can act on rather than a paragraph they have to parse. Design pushed for a side-car treatment; engineering constraints landed the MVP on a floating panel — the compromise is visible here.
years the arc runs, from the first machine-learning feature to AI-native exploration.
technology horizons, each one redefining what the product and design were for.
Assistant that a proliferating set of generative features was consolidated under.
the Assistant and AI Agent rollout that changed how the platform was judged.
-
Re-diagnose at every wave, don’t re-skin
Each wave changed the job, not just the capability. Treating it as a styling problem would have produced four coats of paint over the same misunderstanding — so the question every time was what the agent’s work had actually become.
-
Horizon one — put machine output where the work is
Transcripts and classic ML arriving into the agent’s live surface. The constraint set here held for the next five years: the agent has one attention, mid-call, and every feature is spending it.
-
Horizon two — retrieval, grounded in what the enterprise knows
Answers pulled from the customer’s own knowledge rather than a general model’s guess — which made accuracy a content and configuration problem as much as a model one.
-
Horizon three — one Assistant, not one feature per team
Generative features were multiplying faster than the surface could absorb them. Consolidating them under a single Assistant kept the experience from fragmenting a second time — built provider-agnostic and admin-configurable, so the model choice stayed the customer’s.
-
Horizon four — explore AI-native on purpose
A hackathon used as a structured way to ask a different question: not what AI we add to this product, but what the product would be if it had been designed AI-first.
-
Write the principles down
AI principles committed to paper, so teams facing the same question in different quarters answered it the same way. Governance by shared position rather than per-team guesswork.
- Three technology waves inside five years, each redefining what the product and design’s job were.
- AI features multiplied faster than the agent’s surface could absorb — every one spending the same finite attention.
- Without a written position, each team would have answered the same AI questions differently, and the surface would have fragmented by default.
- The platform was being judged behind cloud-native rivals on AI, so the work carried a perception problem alongside the product one.
The Assistant and AI Agent rollout shifted how the platform was seen — after years of being judged behind cloud-native rivals on AI, the work became the evidence that the gap had closed. The role moved with it, from owning a surface area to leading the product direction. And the team came out fluent in complex AI rather than relearning it from scratch at each wave, which is the part that pays off on the horizon nobody has named yet.
Ask what changed about the job
Every wave invites a re-skin. Insisting on re-diagnosis — what is the agent actually doing now — is what kept four different technologies from producing four disconnected products.
Attention is the scarce resource
Not screen space, not model quality. An agent has one attention with a customer waiting, and every AI feature spends it. Designing against that budget was the constant across all four horizons.
Principles outlast horizons
A written position lets teams decide separately and still decide alike. It was the only artefact from this arc that stayed useful after the technology underneath it changed again.