Ben VanOuse | Product Management, Business Systems and AI Adoption

Component sourcing was slow, manually intensive, and dependent on supplier familiarity. The purchasing team was also relatively new to AI, so an open-ended chat interface added friction instead of making the work easier. I built a lightweight web portal connected to Google Gemini. Users could enter a product and quantity, refine the criteria that mattered, receive ten close matches, and generate a tailored inquiry email to the contact listed for each result asking about current cost and lead time.

Proof pointDesign target · 10+ ranked alternatives per searchRecorded design target

The critical reframe

“The opportunity was not simply better AI search. It was giving the team a familiar, controlled interface that embedded AI inside the sourcing workflow they already understood.”

I bring

Mapped the sourcing workflow, designed the user experience, defined the search controls, and determined what supplier communication should contain.

AI contributes

Expanded the product search, organized close matches, and drafted supplier-specific outreach from the user’s criteria and quantity.

Together

I made AI approachable and operationally relevant; Gemini supplied the search and drafting capacity behind the interface.

How I approached it

Translated the sourcing process into a guided web interface
Allowed users to refine searches around the criteria relevant to the purchase
Used Gemini to return ten close matching products with supplier context
Generated a reviewable outreach draft for each result using the requested quantity, product, cost, and lead-time questions

Where judgment mattered

I designed for the purchasing team’s actual comfort level and kept search criteria and supplier outreach under user control rather than forcing adoption through an unfamiliar conversational interface.

What changed

A functioning sourcing portal turned AI-assisted research and initial supplier outreach into a guided workflow the purchasing team could use without needing to become expert AI prompt writers.

A frequently requested storage concept began as a loosely defined opportunity. It still had to survive customer validation, structural testing, manufacturability decisions, packaging, installation, system setup, positioning, and launch before it could become a dependable commercial product. Over 17 months, I led the work from requirements and early prototypes through customer installs, production readiness, three launch SKUs, packaging validation, documentation, ecommerce setup, and an August 2025 launch. The first design changed substantially as testing exposed binding, latch, installation, structural, and packaging problems.

Proof pointRecorded · $146,665 launch-year revenue · 102% above forecastRecorded 2025 result · 514 units and 17.3% attach rate by March 2026

The critical reframe

“The job was not to protect the first design. It was to build a disciplined loop for finding what was wrong, making the right tradeoffs, and carrying those decisions all the way through the customer and operating experience.”

I bring

Owned the product-management continuity, made the commercial and customer tradeoffs, coordinated the cross-functional work, and remained accountable for physical testing and launch readiness.

AI contributes

Supported later component sourcing, structured meeting decisions and actions, and preserved continuity across the 2026 adaptation and cost-reduction work.

Together

I led the product and business decisions; AI accelerated research and program continuity without replacing engineering validation, supplier confirmation, or launch judgment.

How I approached it

Turned an ambiguous opportunity into testable customer, product, and commercial decisions
Used prototypes and customer installs to challenge assumptions and prioritize changes
Coordinated design, manufacturing, quality, purchasing, finance, marketing, ecommerce, suppliers, and customers
Built the BOMs, assembly process, packaging, guides, item setup, inventory, positioning, and launch readiness needed to commercialize reliably

Where judgment mattered

I repeatedly chose customer reliability and manufacturability over attractive features, rejecting ideas such as folding construction, dampened slides, flat-pack assembly, and broader factory changes when the evidence did not justify their cost or risk.

What changed

The product generated $146,665 in 2025 revenue, finishing about 102% above its revenue forecast. It reached 514 units within roughly seven months, achieved a 17.3% attach rate, and established the identity for a broader storage collection. A later redesign modeled an $80.10 per-unit cost reduction and gross-margin improvement from about 15% to 32%.

Marketing and Customer Service regularly needed product answers about cost, margin, inventory, and likely stockout timing. The source data existed in Salesforce and Shopify, but neither system gave these teams a simple, complete answer, so the questions kept coming back to me. I used Coefficient to bring live Salesforce and Shopify data into a structured spreadsheet that acted as the engine under the hood. I organized the fields and calculations around the questions people actually asked, then built a focused Gemini interface so Marketing and Customer Service could get current answers in plain language without learning the underlying systems.

Proof pointEstimated · 30 to 40 hours/year recoveredEstimated from 30 to 45 minutes of repeated questions avoided each week

The critical reframe

“The system closest to a transaction is not automatically the best place for someone to get an answer. The source systems needed a dependable logic layer and a simpler doorway for the people using the information.”

I bring

Defined the useful questions, selected the authoritative sources, designed the data model and calculations, and established what a trustworthy answer needed to contain.

AI contributes

Provided the approachable question-and-answer layer over the connected, continuously updated product data.

Together

I built and governed the information system; Gemini made it easy for other teams to use without depending on me.

How I approached it

Started with the recurring questions Marketing and Customer Service needed answered
Connected live Salesforce and Shopify data through Coefficient
Built a controlled spreadsheet layer for cost, margin, inventory, and stockout logic
Created a focused Gemini interface that made the resulting answers approachable and immediately available

Where judgment mattered

I determined which sources were authoritative, how the data needed to be joined, which calculations could be trusted, and what the answer needed to show before Gemini ever presented it.

What changed

Marketing and Customer Service could get product answers in real time instead of waiting for me to investigate, and I recovered an estimated 30 to 40 hours per year for higher-value product work.

Inventory, rolling sales, unshipped orders, open purchase orders, vendor lead times, ETAs, and product configuration data lived in separate systems. There was no single view that answered the operational question that mattered: where is inventory headed, and what needs attention next? I built the planning engine beneath my later real-time product-answer experience. A data connector refreshed Shopify inventory, Salesforce sales, unshipped orders, vendor and purchase-order data, and product configuration into one Excel model. The system normalized demand, incorporated inbound supply, projected inventory four months forward, and automatically surfaced products expected to reach zero.

Proof pointRecorded · 75% less management time · $400K lower year-end inventoryFirst-year result · year-end inventory finished at approximately 50% of the prior year

The critical reframe

“The goal was not another spreadsheet. It was a connected decision system that started with the exceptions, preserved product-specific logic, and made future inventory risk visible before it became an operating problem.”

I bring

Defined the decisions the system needed to support, selected the dependable sources, designed the demand and component logic, and remained accountable for purchasing judgment.

AI contributes

The planning engine did not depend on AI. AI later became the approachable interface over this governed data, translating current product information into plain-language answers for other teams.

Together

I built the trustworthy operational foundation; AI later made that foundation easier for colleagues to access without weakening the underlying controls.

How I approached it

Worked backward from the purchasing decision instead of starting with available reports
Connected refreshed inventory, sales, order, vendor, lead-time, PO, ETA, and configuration data in one model
Converted rolling sales into expected monthly demand and projected inventory one through four months forward
Built kit-to-component logic so configurable-system sales created demand against the underlying components
Surfaced zero-inventory projections and product-specific exceptions as the purchasing focus

Where judgment mattered

I determined which sources were authoritative, how demand should be normalized, where configurable products concealed component consumption, which exceptions belonged in the logic, and what threshold should trigger action.

What changed

In the first year after implementation, inventory-management time fell 75%, year-end inventory was $400,000 lower, and total year-end inventory finished at approximately half of the prior year. Instead of repeatedly reconciling systems, I could begin with the exceptions and focus on purchasing decisions and inventory risk.

If this were your organization

I would connect the sources already carrying your operational truth, encode the product and exception logic that currently lives in people’s heads, and turn the result into a forward-looking view that tells owners where attention is needed next.

Want another way into the story?

Evaluation completed · follow-up report-filter tests passed under administrator access

Before expanding AI access, leadership needed to know what each tool could actually read and change in the CRM. A connector listing or a confident model answer was not enough evidence. I used a common 13-prompt protocol across three AI platforms, covering standard records, custom objects, relationships, freshness, saved reports, and record changes. I checked important outputs against the CRM. When a report failed through one connection but worked through the API under the same login, I isolated an integration gap and configured a separate reports connector. Follow-up tests verified ad-hoc report filtering without changing saved reports, using my System Administrator account.

Proof pointEvaluation scope · 13 prompts across 3 AI platformsSeptember 2026 evaluation · findings limited to tested configurations and dates

The critical reframe

“Test the same real work, verify the source results, and distinguish product limits from configuration gaps.”

I bring

Defined the evaluation, checked source results, investigated failures, configured the connection, and communicated the adoption tradeoffs.

AI contributes

Executed test prompts, queried permitted records, helped reconstruct reports, and accelerated diagnosis and documentation.

Together

I set the standard of proof; AI made it faster to test that standard across systems and tools.

How I approached it

Defined common tests around real operational questions
Compared important answers against the source system
Distinguished record queries from saved-report execution
Diagnosed the failed path against the underlying API
Configured report access and documented what still needed testing

Where judgment mattered

I defined what counted as a pass and kept setup requirements in the adoption decision. A workaround or a tool’s description of itself did not prove capability.

What changed

A documented evaluation and configured reports connection gave leadership a clearer basis for tool selection and integration ownership. Follow-up report-filter tests passed in the tested administrator configuration. Employee-level access and broader deployment require separate validation.

A cross-section of how I work