Home/ Work/ Meta · Instagram Business
Case Study 02 — AI Research Infrastructure

Instagram Ads Sales Feedback AI Tool

The richest signals about what businesses needed were locked inside sales conversations. The challenge wasn't the data — it was building a system the right people would trust enough to contribute to.

Org-wide
Adoption across Instagram Business
Real-time
Signal access for product & research
Managed
Business sales conversations synthesized
Cross-team
Alignment: sales, product, policy, research
Act I The Hidden Signal Problem

Thousands of sales conversations happened every week. None of it reached the teams who needed it.

Instagram's managed business advertisers were telling account managers exactly what they needed. What was confusing about the ads platform. What features were missing. What competitive pressures they were facing. What would make them spend more. This was some of the richest behavioral and strategic intelligence available — and it was evaporating into Zoom recordings, Salesforce notes, and the departing institutional memory of individual reps.

Product teams were making decisions without it. Research teams were running studies to answer questions that sales conversations had already answered. Policy teams were responding to problems weeks after the sales floor had spotted them. The information existed. The infrastructure to route it did not.

The structural problem had three layers: data access (conversations weren't captured in any usable form), data quality (when sales reps did document, the format was inconsistent and unsearchable), and data trust (reps worried that contribution would be used to evaluate their performance rather than improve the product).

"The bottleneck wasn't gathering data. It was getting the right people to trust the system enough to feed it — and designing the system to be worthy of that trust."
Where insights were dying Sales calls ~1000s/week Rep notes inconsistent insight graveyard no routing · no synthesis · no access Product decisions Research blind spots Policy reactive all making decisions without the same signal
The trust problem
Sales reps worried that contributing insights would be used to evaluate their performance, not improve the product. The system had to be designed to make contribution feel safe before it could ask for contribution.
Act II Getting the Right People to the Table
Knowledge contributor map High signal · hard to reach Senior sales reps trust concerns Account managers time constraints SME bottleneck Who validates the synthesis? Who owns quality? political · scarce · contested What the AI had to do Low-friction input → pattern synthesis surface to right team · no eval risk for reps trust by design

The hardest research problem was organizational, not behavioral.

Getting the right stakeholders involved required navigating competing incentives at every level. Senior sales reps held the most valuable knowledge but had the most to lose if contributions were misused. Account managers had time constraints that made structured data entry feel like overhead on top of an already heavy workload. SMEs who could validate the AI's synthesis were scarce, politically contested, and difficult to commit to a system they hadn't designed.

My research focused on the contribution system before the AI. What made reps willing to share? What formats felt low-stakes enough to complete in real time? Whose endorsement made the system feel legitimate? These were social and organizational design questions as much as UX questions.

The insight that unlocked the design: reps were willing to contribute when the act of contributing felt like helping the product team, not reporting to management. The framing of the interface mattered as much as the functionality. Input fields that felt like suggestions were completed. Input fields that felt like audits were abandoned.

Getting SMEs to participate required a different approach entirely. Rather than asking them to review everything, the AI pre-clustered insights by theme and surfaced only the ones with high uncertainty — reducing the cognitive load of review while preserving the quality signal that made the repository trustworthy to downstream teams.

Act III What Changed Org-Wide

The repository became the shared language between sales and product.

Once the system was trusted, it compounded. Reps who contributed early became advocates for the tool because they saw their own observations reflected in product decisions months later. That visible loop — contribution leading to change — was the most powerful adoption mechanism available, and it was only possible because the research identified what motivated contribution in the first place.

The impact for product and research teams was immediate. Teams that had previously been running studies to understand what managed businesses needed could now triangulate their findings against a real-time, high-volume signal from actual sales conversations. Research velocity increased. Study designs became sharper because baseline understanding of the problem space was richer.

Organizational Impact
Org-wide adoption across Instagram Business — product, policy, research, and sales teams using the same knowledge base for the first time
Real-time signal access for product and research teams — previously dependent on quarterly studies, now supplemented by continuous sales intelligence
Trust-by-design contribution system — framing and IA changes that made reps feel like product collaborators rather than data sources, driving sustained engagement
Scalable SME review workflow — AI pre-clustering reduced expert review burden by surfacing only high-uncertainty insights, making quality control sustainable
The systemic lesson
Research infrastructure is a product problem. The AI tool succeeded not because the algorithm was sophisticated but because the contribution system was designed around the psychology of the people being asked to use it. Trust is a design decision, not a communication strategy.
Broader implication
As AI tools proliferate in enterprise contexts, the research question shifts from "does the AI work?" to "will the right people contribute to it, and will they trust what it returns?" These are human behavior questions that no algorithm can answer on its own.

Next Study

Building trust in automation

Optym · RouteMAX →