#impactai

By late 2024, Impact had a network, a hiring tool, and an on-set communication platform. The studio deals were coming in. The next question was whether AI could surface the right people faster than search could. This is what that looked like to build.

Showing The World What Impact Could Do For Production Studios

By late 2024, Impact had a functioning network, hiring system, and on-set tools. The ecosystem worked — but it was a closed environment. You had to already be inside it to understand why it was valuable. Acquisition was the bottleneck.

Instead of asking people to sign up first, we built impactAI — a conversational search tool powered by an LLM connected directly to the Impact database, available to anyone without logging in. The wall appeared only at the moment of intent: connect, message, or view a full profile triggered a sign-up prompt. Value demonstrated before the ask.

Business Goal

Test AI as a growth mechanism. Make the Impact network legible to outsiders — specifically to coordinators and production companies at major studios — without requiring account creation upfront.

Where it lived

Two surfaces: embedded on the Impact homepage, and as a standalone public link — no login required

From MVP to Product

What We Stated With

Engineering shipped a working prototype called "Listmaker." The LLM connected to the database, queries returned results, candidates appeared. But it communicated nothing about why you should trust it, how to use it, or what it was for. The name said it all — a utility, not a product.

From MVP to Product

Educating The User & Building Trust In The Results

FRAMING

Key To Hollywood" → "Who Do You Need?"

"The Key To Hollywood" overpromised relative to what the tool actually did. I reframed it narrower — "Who Do You Need?" — and designed it to double as onboarding. Users immediately understood what input the system needed from them.

ONBOARDING

Empty input → Suggested prompts

An empty search box communicates nothing. Suggested prompts modeled three things simultaneously: the range of query types, the specificity the system needed to work well, and the questions Hollywood professionals actually ask when hiring. A tutorial disguised as UI — borrowed from Perplexity, adapted for a higher-stakes professional context.

TRUST

Based On" tags + Edit Filters + Report an Issue

After parsing a query, we surfaced exactly what the AI had extracted as editable tags — role, location, experience level, genre, production type. Users could see the reasoning and correct it. "Report an Issue" was built into results — partly a trust signal, partly a real feedback loop into the model. In an industry where a bad hire damages your reputation, transparency about AI reasoning wasn't optional. It was the product.

AVATAR

Not a wizard. Not a database. Something in between.

Three territories explored: wizard/magic (intelligent but opaque), spiral/abstract (AI-coded but cold), generic profile silhouette (familiar but misleading). The final direction was neutral without being clinical — visually focal, matching the Impact brand without feeling robotic.

Two Experiences

Unauthenticated showcases. Authenticated brings value.
 

The unauthenticated experience was open enough to demonstrate real value without a sign-up. The authenticated experience was meaningfully richer: mutual connections surfaced and flagged visually, search history informing current results, and ranking drawn from your full professional relationship map — not a generic database query.

The gap was intentional. Unauthenticated created desire. Authenticated made the value of an account immediately visible the moment you signed in.

Shipping It

Manual testing, studio feedback,
and what happened next

How We Tested Reliability

The gap between "the model works in general" and "the model works for queries real users actually type" is wide. We closed it manually — me, the PM, and the CX coordinator working through a Google Doc list of test queries for hours before each release.
The ML engineer and I iterated on the UI together — exit paths, filter logic, reasoning states. Design and the model evolved in parallel, not independently.

The Specificity Problem

"Producers in LA with four years of experience in horror" — accurate results. "Producers in LA with some experience in horror" — results appeared but weren't reliably relevant. "Some experience" gave the model no signal to score against. Finding this in testing, not through user complaints, meant we could design around it before launch.

What It Achieved 

impactAI became the most effective sales tool the team had — studio contacts could search the network and see real results in under two minutes. No pitch deck could do that.