Overview

From a one-line hypothesis to
a tool sales teams actually use

Tenetic walked in with one sentence: "AI should be able to write our customers' pitch decks." That was the brief. No flows, no feature list, no clarity on which parts of deck-building the AI should touch and which parts the user should still own.

That kind of brief is normal in AI work right now. Someone in the room is sure the technology works, and nobody is sure what the product looks like. Our job was to figure that out without slowing the build.

Every design decision in Tenetic ladders back to one of these three.

What We Were Aiming For

The Decision Filter

Five insights that became
our scoping weapon

We locked these phrases in the first week of synthesis and used them as the test for every feature proposal that followed.

Insight 3 ended up shaping the entire interaction model. Once we accepted that nobody would treat the first generation as their final draft, every screen in Implementation became a question of how to make refinement cheap.

The Market Gap

Five tools audited.
None of them spoke media-sales fluently.

Every competitor was a general-purpose deck builder. The wedge for Tenetic was domain depth. We weren't chasing Gamma on "make any deck prettier." We were chasing PowerPoint on "make this deck, the one a CBS local rep ships on a Tuesday morning, with Nielsen numbers and dayparts and a Q2 narrative."

Designing for that meant making the model fluent in something general-purpose tools never had to handle: media-sales schema. Nielsen Scarborough demographics, Comscore reach, MRI psychographics, dayparts, genre splits, target audience. Most of the AI work in Implementation came down to constraining model output to that schema instead of letting it free-form anything.

Primary Persona

Meet Ali

Ali is skeptical of AI, not afraid of it. She'll use Tenetic if it makes her faster without making her look stupid in front of a client. She'll abandon it the first time it hallucinates a Nielsen number. Every screen we shipped is for her.

The Journey, Before and After

The wait gap that
kills mornings

The current state has a wait gap that kills mornings. Ali starts a deck, hits the data wall, sends a request to her research analyst Lauren, waits, gets numbers that are close but not quite right, and goes back to Lauren. The new state lets her draft in parallel and enriches the deck when Lauren's data lands.

The Product in One Sentence

Every word in this sentence was argued over. After it landed, we used it as the tiebreaker for every scope debate in the next eight weeks.

The Hero Flow

Six steps. Each one
a commit point.

Each step is a place where the user can back up or move forward. That pattern showed up again and again in the insights.

What Discovery actually shipped: a locked persona, the five-insight filter, the six-step flow, and two features explicitly cut from v1. The client wanted Client Management (a CRM inside the deck builder) and Loyalty & Gamification (streaks, badges, points). We argued both violated Ease of Use and would have pushed Implementation into a second engagement. The insights won.

Style Guide

A semantic palette built
for AI system states

AI products throw more system states at the user than typical SaaS. Generating, regenerating, data-pending, partially complete, error-recoverable. The communication palette had to handle all of them at a glance, because reps work fast and don't read state copy carefully.

Sprint 01

The shell. Where Ali comes back to her decks, finds last quarter's work, and starts something new.

The decision worth flagging: the client wanted "Create with AI" prompts in every empty state of the dashboard. We argued against it. Empty-state AI nags train users to treat AI as a chore, and the dashboard is supposed to feel like a calm landing pad after a meeting. We shipped a single prominent Create button at the top right and dropped the AI prompts. By the end of Sprint 01 the client agreed it felt cleaner.

Sprint 02

This sprint defined every AI interaction pattern downstream. The client wanted a single text box — prompt in, deck out. We pushed back. In media sales, a prompt like "Q2 local news pitch for a Chicago auto dealer targeting women 25–54 during morning drive" has five embedded variables. Every misread costs Ali twenty minutes. With four to six decks a week, misreads compound.

We shipped two modes in the same flow: conversational for exploration, structured for repeatable high-stakes work.

Conversational refinement. Ali types her prompt. The AI reads the variables back as a structured summary. She confirms or corrects, nothing generates until she commits. Prompt engineering, exposed as UI.

Storybuilder accordion. Six optional sections, collapsed by default. Each maps to a schema-bound input — audiences, dayparts, genres, KPIs — instead of free-text guessing. Schema-binding was the single biggest hallucination reducer we shipped. First version was a long open form; reps bounced on sight. Collapsing by default and making every section optional doubled completion.

One more reason for the confirm step: output anchoring. Once an AI artifact exists, users get attached and won't redo it even when they should. Moving the anchor point upstream onto a summary the user owns makes redirecting cheap.

Sprint 03

The client wanted a spinner. Ten seconds, rotating icon, "Generating your presentation," done. We shipped a transparent four-phase loading state instead: Structuring narrative, Pulling supporting data, Drafting slides, Applying template.

The design move here is perceived performance, which I'd argue is one of the most underrated patterns in AI UX. Real generation latency for a deck of this complexity sits in the 30 to 90 second range. That's long enough that a spinner feels punishing. Breaking the wait into visible phases doesn't shorten it; it makes it legible. Users tolerate longer waits when they can see progress and infer where they are in the work. As a bonus, each phase doubles as a debug surface. When generation fails, the error can name the phase instead of throwing a generic "something went wrong."

Two more details baked into this screen. Cancel exists on the loading state because aborting a bad generation early is cheaper than waiting for it to finish and starting over. Cost matters here, because every regeneration burns API budget the client has to model into their pricing. And the editor itself was built to look like Slides or Keynote on purpose. The novel interactions in Tenetic are about how the deck got there. The editing surface should feel invisible, not innovative. AI tools that try to redesign the slide editor lose users on the first edit.

Sprint 04

This sprint closed the loop on Refine Output and built in two patterns I now treat as defaults for any AI generation product.

Outline review as a cheap commit point. The AI's first artifact isn't a deck — it's an outline. Slide title, one-line summary, data source tag. Ali reviews, reorders, regenerates individual lines, or sends it back with new instructions. Structure locks before any full slide generates. Direction failures cost twenty minutes; execution failures are recoverable in seconds. Negotiate structure first.

Slide states as the AI failure surface. Standard SaaS has two states: loading and loaded. AI products need at least six. Ours: empty, generated, edited, regenerating, data-pending, error-recoverable. Each has a distinct visual treatment so Ali always knows what's her work, what's model-generated, and what needs attention.

Data-pending and HOTL/HITL granularity. Ali drafts before Lauren's research lands. When data arrives, pending slides auto-enrich. The model proposes; Ali accepts, rejects, or edits. Human on the loop continuously, human in the loop at the commit gate.

Inferred vs. user-provided values, marked. Where the model filled in a variable Ali didn't supply, the slide carries a subtle marker. Hover to see what was inferred, override in one click. Without this, reps either over-trust and ship a bad number, or under-trust and redo everything. Data grounding was the single thing that made legal approve the tool for client-facing use.

The Patterns We Kept Reaching For

A short list of moves that
became defaults

A short list of moves that became defaults whenever the model touched the interface. Most of them are about lowering the cost of being wrong.

How the Sprint Actually Ran

Three designers. Parallel lanes.
Embedded engineering.

A typical sprint week looked like this. The three designers scoped together on Monday with the BAs to align on user stories and edge cases. Mid-week we ran internal critiques with the Creative Director. Engineering joined the second-half reviews to flag feasibility before pixels got pushed too far. The client took two passes at the end of every sprint, one UX review focused on flow and decision points, one UI review focused on visual fidelity and brand. Anything unresolved carried into the next sprint as design debt.

Engineering caught data-state edge cases we would have missed on our own. BAs caught business rules that changed how the AI should be allowed to behave. Client review surfaced enterprise constraints (compliance, data governance) we couldn't have known to ask about up front. The AI design patterns we shipped came partly from a research-backed point of view and partly from those review cycles.

Conclusion

Two parts. One AI workflow
that earns trust.

Discovery turned a one-sentence client hypothesis into a defendable spec, anchored in user research the design team ran ourselves. Implementation translated that spec into an interface where the model behaves like a collaborator with clear gates.

Credits

Photon Interactive

Designed at Photon Interactive, June to September 2025. Three UX designers, one Creative Director, an embedded engineering team, the Tenetic BAs, and a client who learned to trust the design process by the end of Sprint 02.