Imagine LA Benefit Navigator copilot

Nava Labs, part of Nava Public Benefit Corporation, built and piloted a generative AI chatbot on Imagine LA's Benefit Navigator platform. Caseworkers type a client's question during a call. The chatbot answers from approved government documents and attaches direct quotes, so the caseworker can check the answer before relaying it.

How it is used

Benefit Navigator is Imagine LA's digital tool for benefit navigators, staff who help families apply, and for case managers. They use it to connect low-income Los Angeles County families to public benefits. Imagine LA is a 501(c)(3), a tax-exempt nonprofit, with more than eighteen years of direct services. Amplifi handled the platform integration, the technical work of fitting the chatbot into the platform.

The chatbot answers questions about eligibility, application steps, documentation, and how programs interact. It covers roughly forty Los Angeles County benefit programs and tax credits. It answers in several languages, and it can restate an answer in simpler language or in Spanish.

It is assistive. Caseworkers use it while on the phone with a client. They check its answer against the quoted source before relaying it. It does not make eligibility decisions on its own.

Who paid for it

A $3.6 million Gates Foundation grant, announced in December 2024, funded the pilot. Nava's announcement says the grant also covers tools for document processing, referral generation, and case-note summaries.

A $1.5 million Google.org grant in June 2025 funds Nava's next step, with Imagine LA and First 5 Riverside County. It is a move toward AI agents. They would navigate databases and benefit portals, and complete applications under caseworker supervision.

How it was evaluated

Nava and its research partners, Cornell University and Georgetown University's Better Government Lab, ran the evaluation. It had two parts.

The first was a randomized controlled trial, a study that assigns people to groups at random so the groups can be compared. It had 125 caseworkers. It measured how accurate their answers to hypothetical client questions, built from real experiences, were when they were shown the chatbot's answers.

The second was a fourteen-week pilot in real work. It had 61 caseworkers across six organizations in Los Angeles County.

What the evaluation found

Using the chatbot was estimated to improve answer accuracy by about 40 percent on average. The gains were largest on the hardest questions and among newer, less-experienced staff. The evaluation gives that as a direction, not a measured breakdown.

About 65 percent of caseworkers with access used the chatbot, submitting about 14 prompts, or typed requests, each on average. Satisfaction was modest and low-positive: a Net Promoter Score of 11, with 40 percent classified as promoters. That score is a common measure of how likely users are to recommend a product. Use tended to decline over time and varied by site. The evaluation reported that sustained use needs ongoing support from the organizations.

Answers averaged a tenth-to-twelfth-grade reading level. That is easier than the college-level source manuals, but short of a plain-language goal.

Results on administrative burden and time savings were promising but inconclusive. They were not statistically significant, because of sample-size limits and a lower response rate.

Who wrote the evaluation

Nava shared the pilot results at a Demo Day in December 2025. The work was published in March 2026 as a Nava case study. The same team published it as the academic report "Helping the Helpers".

Its authors are Allison Koenecke and Jennah Gosciak of Cornell, Eric Giannella and Zhaowen Guo of Georgetown's Better Government Lab, and Michael Chen and Martelle Esposito of Nava. So the tool's builder co-wrote the headline result with its academic partners. No fully independent replication exists.

Figures that belong to the platform

Some figures often cited nearby belong to the Benefit Navigator platform, not to the chatbot. An earlier platform pilot, from September 2023 to November 2024, reached more than 500 case managers. It assisted over 10,000 beneficiaries and secured on average an additional $10,869 in benefits per household.

What the case file draws from it

The case file argues that the measured benefit and the risk sit in the same place. The chatbot helped most on the hardest questions and for the newest staff. Those are the questions where a wrong answer is hardest to spot, and the staff least able to catch it.

The whole safeguard is one habit: read the quoted source before relaying the answer. The evaluation shows that habit is not settled. Use reached about two-thirds of caseworkers, then declined without sustained engagement. The case file adds that the fluent answers can sound more certain than the policy behind them.

The chatbot flags no one and denies no one. So the case file sees the governance work as keeping a habit alive. It warns that the funded move toward agents that complete applications under supervision would take away the check a caseworker makes on the phone.

How this case differs from the Nava case

The Lab also has a case on Nava's chatbot for benefits caseworkers, drawn from Nava's earlier, exploratory work. This case draws the Imagine LA pilot, which put the same careful design to a randomized test. The test showed where the gains landed and that use faded. The earlier case could not show either.

What this network is drawn from

This network follows the pattern the case file describes. It does not reconstruct the actual chatbot. It shows the chatbot, its search of the approved documents, the documents, the caseworkers, and the evaluation. The families served are outside the network.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. Closing a pathway means mistakes stop passing along it. The work along it goes on.

This case has a budget of 7 units. Each tool costs the same at every target level, except Understand the system, which pays for ongoing study of the deployment. It costs 3 units under Explore (No Targets) and Service Targets Only, and 4 under the two higher levels. While it is on, Escalate checks, Know the tool, Keep prompts neutral, and Vet connections each cost 1 unit less. With its stronger setting, Deep research, which costs 6 units, they cost 2 less, but never less than 1.

Before any tool is used, one failure pathway is open: Quoted answers to caseworkers. The network is self-correcting, meaning it clears mistakes on its own. The caseworkers are keeping up with the work.

Explore (No Targets) sets no targets. Under Service Targets Only, the targets are met before any tool is used. Every set of tools within the budget keeps them met.

Under Service and Safety Targets and All Governance Targets, the targets can be met within the budget. Both levels ask you to close every failure pathway, among other targets. Escalate checks is the only tool on offer that closes Quoted answers to caseworkers. It meets the targets on its own for 2 units. Every set of tools that meets them includes it. Within the budget, 45 different sets of tools meet Service and Safety Targets, and 44 meet All Governance Targets.

The two higher levels also ask you to refill the Privacy gauge, the Lab's reading of how well client data is kept safe. It starts below full. With no pressure added, no pathway here drains it, so it refills as the clock runs.

More is not better. Using every tool, each at its strongest setting, costs 29 units, more than four times the budget. It meets Service Targets Only and Service and Safety Targets, but not All Governance Targets. At that level, the chatbot then adds too little to the work.