The Conversational On-Ramp

The current generative-AI discourse collapses two distinct creative needs into a single product surface: the desire to start quickly and the need to finish precisely. We argue that the conversational, agentic interface is a superb on-ramp — and a poor finishing tool. The right architecture is mixed-initiative: a fast conversational canvas sitting beside deterministic craft surfaces, not above them. We describe how this is implemented in AIM Previz as the relationship between the Infinitum on-ramp and the five craft studios that finish the work.

Pre-visualization has always had two phases. Ideation is divergent: the director wants something to react to, fast, before committing to a direction. Production is convergent: once the direction is set, every frame needs to be reproducible, controllable, and accountable to a continuity record.

Conversational agents excel at the first. They collapse what used to be ten setup decisions into a single sentence. But the qualities that make them fast — interpretive freedom, autonomous planning, opaque routing — are the same qualities that disqualify them from the second. A production pipeline cannot tolerate "the agent decided to vary the lens that take."

Products that put the agent at the centre of the workflow tend to fail in the same way: they impress at the demo and frustrate at the deliverable. A creator can generate a striking first shot, then spend hours trying to coax the system into matching that shot's continuity, lighting, or framing for the next nine. The agent's helpfulness becomes the bottleneck. Each clarification turn re-rolls the dice on parameters the creator wanted held constant.

This is not a failure of model quality. It is a failure of locus of control. The agent owns decisions the professional needs to own.

The architecture that works in practice is mixed-initiative: agentic for the on-ramp, deterministic for the finish, with a one-click handoff that preserves continuity references. The agent does what agents are good at — interpreting a sentence, planning a few steps, fanning out variations. The craft studios do what craft tools are good at — exposing every parameter a professional needs, holding them constant on demand, and producing reproducible output.

"Spark in the agent. Finish in the studio that owns the craft."

In AIM Previz, the on-ramp is a workspace called Previz Infinite, where a creative-director agent — Infinitum — sits on an infinite canvas. The creator talks, points, and reacts. The agent plans a few steps per turn: generate images, fan out coverage, draft a character bible, sketch a sequence, or assemble a mood board.

None of this work has to live in Infinitum. Any asset can be handed in one click to Keyframe (cinema-grade frames with Continuity Lock and camera profiles), Director (multi-engine motion previz with multi-take grids), CineForge (staged 3D with PBR materials and a path-traced render pass), Storyboard (deterministic shot coverage), or Script (screenplay structure and breakdown). The continuity references travel with the asset.

Crucially, the craft studios do not expose an agent at their centre. They expose controls — camera body, lens, lighting Kelvin, look profile, take count — that a cinematographer can name and a producer can sign off on.

A common temptation is to position the agent as the universal interface — every action routes through a chat surface, and the craft tools become "advanced settings." We reject this framing. The agent and the craft surfaces are peers. Most creative work moves back and forth between them: a quick conversational pass to find a direction, a long deterministic pass to render the deliverable, another short conversational pass to vary it, and so on.

Placing the agent above the craft studios reintroduces the failure mode above: every precise action becomes a negotiation with an interpreter. Placing it beside them keeps the agent's strengths (speed, ideation) without surrendering the craft layer's strengths (control, reproducibility).

Once the on-ramp is conceptually separated from the finish, the pricing architecture follows. The on-ramp can be made broadly accessible, and PREVIZ takes that to its end: every plan, the free one included, opens every surface — Cinematic render tiers, multi-take video, sequence intent and the craft studios alike. What a plan buys is the credits the deliverable costs, not which door it may use. Users self-select by how much work they are doing, never by which layer they are allowed to reach.

The conversational surface is not the future of professional creative tools — it is the front door of them. The future of professional work remains deterministic, reproducible, and controllable. The agent's job is to get the creator into the building. The craft studios' job is to let them finish the film.