Page 32 - FCW, March/April 2020

Artificial Intelligence
Executive Viewpoint
A Conversation with
KEN THOMSON
KEN THOMSON
BuySmarter Solution Architect, Department of Health and Human Services
A BuySmarter leader discusses how HHS is using artificial intelligence as part of a data-driven initiative to leverage the department’s collective purchasing power
How is HHS using artificial in- telligence to improve acquisition through its BuySmarter initiative? The Department of Health and Human Services has over 40 different contracting shops across the 20-plus operating and staff divisions, each buying similar and frequently the same goods and services at disparate prices. Due to siloed purchasing, HHS agencies have minimal buying power, and vendors thus have the advantage in contract negotiations.
To shift this power equation in favor of the government, HHS is using AI to centralize its contract information into a single data layer, converting contract and attachment data into metadata (plain text), structuring it by category and sub-categories, and building a universal taxonomy for how to understand the information in a common way. The AI identifies where we are buying “like-to-like” goods and services across our divisions utilizing data algorithms derived from the variables developed by HHS contracting and mission subject-matter experts (SMEs).
The algorithms list every variable that could drive pricing (e.g., a registered
nurse working in New York vs. Iowa). Our Category Collaboratives — the groups of SMEs and contracting professionals in charge of a category — will have the full suite of information on every vendor’s pricing, delivery, logistics, quality, support and more for a group purchase of a product or service. The new AI tool will empower HHS divisions with new levels of negotiating power.
What new developments are in the works for BuySmarter?
HHS is in the final phase of developing the AI capability described above, known as the Full Contract Scan Tool, for the medical
category. It has already ingested 1.37 million documents comprising 4.6 billion words with the taxonomy built from the primary category (product, equipment and service) down to the sixth sub-category level.
Using natural language processing, the team developed an inference protocol that allows the AI tool to “read” in a human way and make logical inferences when matching words and phrases. For example, when searching for “neonatal nurse,” the AI tool makes the logical leap that “NICU” belongs alongside “neonatal” in other search results.
What lessons have you learned along the way?
In my first experience with an AI vendor for our proof of concept, I learned to adapt my approach to a more hands-on engagement with the vendor where we built the tool together. The phrase “AI learns” holds true, but you have to “teach the teacher.” This translates to the need for experts in that domain to sit with the AI development team and teach them what the tool needs to do, how the data should be structured, what the outputs need to be and more.
We have all heard of agile as a quicker way to conduct development in two-week sprints. When it comes to building AI tools, the process is even quicker. I call it “full-contact development.” We meet daily with our AI vendor to give guidance and clarity on what the tool should do. Then we start bringing
in waves of SMEs to study the tool and help teach the teachers to improve the capabilities. This full-contact development approach is truly what it takes to build a robust AI tool.
This interview continues at Carah.io/Thomson-HHS.
S-32 SPONSORED CONTENT