AI doesn’t deliver returns on its own. Ready people, clean data and clear guardrails do.
I lead technology products and AI enablement for Unilever Prestige’s eight brands, including Tatcha, Hourglass, Dermalogica, Paula’s Choice and K18, and own the central technology budget, applications and vendors.
People
Built the adoption program that took AI from pilot to production: all 17 employee-built tools from our hackathon now run across sales, marketing, e‑commerce, procurement, R&D and data.
Data
Designed agents that prepare product data for Google’s AI shopping: three months of work in 3.5 hours.
Guardrails
Got Claude approved at Unilever, where no protocol existed, by setting its security controls with the AI governance and cybersecurity teams.
Career
- 2026 – Present
Unilever Prestige
Senior Product Manager, Technology Portfolio Ops & AI Enablement
Own the portfolio’s annual technology plan and budget, and lead AI enablement across eight brands, from getting Claude approved at Unilever to agents that now prepare product data for Google’s AI shopping.
- 2022 – 2026
Unilever Prestige
Senior Business Systems Analyst & Technical Program Manager
Led the design and integration of seven brands’ regulatory and formulation systems into one packaging and ingredients platform, connected commerce, ERP and 3PL systems across four countries, and cut software costs by up to 30% with portfolio-wide contracts.
- 2020 – 2022
Shutterfly
Senior Data Analyst
Built a machine learning model (random forest) that flagged orders at risk of shipping late, avoiding $4.5M a year in expedited shipping and refunds. Led the migration of 200+ dashboards from Looker to Power BI and designed the Azure data lake behind them, covering 20TB of historical data.
- 2016 – 2019
Relocation
Moved from Colombia to the United States
- 2014 – 2016
Claro, Colombia
Regulatory Audit & Automation Lead
Governed 650M+ monthly records for national telecom regulatory reporting and automated audit reporting, cutting cycles by 40%.
Education
B.S. Systems Engineering
Universidad Industrial de Santander, Colombia
- 2019 – 2020
Data Science & Machine Learning Fellowship
Lambda School
Selected work
Making product data ready for AI shopping
The problem
Shoppers increasingly ask Google’s AI Mode and Gemini for product advice, and those answers come from product data. Standard product feeds carry titles, prices and identifiers, not the conversational detail that lives on each brand’s product pages: who a product is for, what it pairs with, which variant to choose.
The approach
I designed and deployed agents that read each brand’s storefront, generate conversational product attributes and structure them for Google Merchant Center. Alongside the feeds, we rebuilt first-party tagging and conversion signals, so the data reaching Google and media platforms is complete and consistent.
How the agent works
The agent drafts. People approve before anything reaches Google.
Brand website sitemap, product pages, ingredient lists
Google’s rules conversational attributes spec, re‑checked on every run
- 01Map the catalogFinds every public product on the brand’s site.
- 02ExtractIDs, shades, sizes, ingredients and claims.
- 03Write Q&AAnswers shoppers’ questions in the brand’s own words.
- 04Fact-checkA second agent checks every answer against the source.
- 05ValidateGoogle’s limits, banned words and duplicates.
- 06PackageUpload files and a review workbook for each brand.
Errors found in fact-check go back to step 03 and are rewrittenErrors from 04 go back to 03 and are rewritten
People approve
- Brand lead checks content and IDs
- Legal signs off flagged claims
- Test upload of 5 to 10 SKUs
Rules it follows: the brand’s own words only · no prices or promotions · fragrance and vegan claims checked against the ingredient list · never uploads on its own
What I decided
- Agents instead of manual work or an agency: the same extraction method for every brand, at a fraction of the cost.
- One standard, set centrally. We defined the legal requirements for an approved conversational attribute once, under our AI policy, instead of leaving each brand to decide.
- All six brands at once, in parallel, which moved faster than a brand-by-brand rollout.
- The US market first, because that is where Google is taking AI shopping first.
Tradeoffs
- Speed over a single-brand pilot. We managed the risk with random samples checked against the source, and a deep review of the most critical questions and topics, ranked by importance.
- Other markets wait while the US goes first.
- Brands give up some say over how their products are described, in exchange for answers that are consistent and approved.
This isn’t optional. AI tools will answer customers’ questions about our products either way. The only choice is whether they use our approved information or fill the gaps from third-party sites.
Three months of work, now 3.5 hours.
How it will be measured
- Attribute coverage: the share of products with complete conversational data
- Feed health in Merchant Center: approval rates and attribute errors
- Visibility and traffic from AI shopping surfaces, where Google reports it
- Click-through and conversion for enriched products against the rest of the catalog
Thinking
AI doesn’t deliver returns on its own
In McKinsey’s 2026 global survey, nearly nine in ten respondents say their organizations use AI regularly. Only about 6 percent attribute 5 percent or more of their EBIT to it. The first question I hear about AI is which tool to buy. In my experience the tool is the easy part. Returns come from three things that have to be ready before the model is: the data it reads, the guardrails it runs inside, and the people who use it.
Data first
When we prepared six of our brands for Google’s AI shopping, the agents were not the hard part. The hard part was the product data underneath: the attributes, the variants, the answers to the questions shoppers actually ask. An agent can only work with what the data contains. Once that data was structured, work that used to take three months took 3.5 hours. The research points the same way: the companies McKinsey calls AI high performers are more likely to have a semantic layer or knowledge graph, one structure that connects data from different systems.
Guardrails before scale
Claude wasn’t an approved application at Unilever, and no protocol existed for it. We started with a small pilot group, then worked with Unilever’s AI governance and cybersecurity teams to set the security controls, showed that Prestige already met the same standards as other approved applications, and got Claude approved. The wider rollout to 400 people came after the controls were in place. High performers manage more of these risks: 45 percent are working to mitigate unauthorized or unintended actions by AI, compared with 33 percent of others. Guardrails are not there to slow people down. They are what let a company say yes quickly.
Adoption is a change program
Giving people a license is not adoption. Nearly three-quarters of high performers have fundamentally redesigned workflows because of AI, compared with one-quarter of others. The Prestige Pioneer plan treated AI the same way, as a change in how people work, with training at its core. Our annual AI Hackathon showed what happens next: teams from across the brands built prototypes in two hours, and all 17 of those projects are now in production across the business.
You can’t mandate across independent brands
Each of our eight brands runs as its own company. A central team can’t order them to adopt anything, so the job is to make the case and build a foundation they want to join. With AI shopping, the case made itself: AI tools will answer customers’ questions about our products either way. The only choice is whether they use our approved information or someone else’s.
AI by itself doesn’t do anything. Prepare the people, the data and the guardrails, and the returns follow.
Source: McKinsey & Company, The state of AI in 2026: On the road to ROI, August 2026. Online survey of 1,719 participants in 97 nations, May 4 to June 8, 2026.
How ready is your company? Take the five-minute AI Readiness Scorecard →