The case for Groundwork

As artificial intelligence capabilities rapidly improve, it wouldn't be surprising to see the use cases expand out of the technology space and into more human-centric fields. New technology is often kept locked behind gilded doors, not easily making its way to the people who could benefit most from it. I find myself increasingly optimistic that, as a society, we'll be able to bridge this gap.

I come from the childcare subsidy space. I am a former case worker at Childcare Choices of Boston. It's a field that is immensely rewarding in a lot of cases, and incredibly frustrating in others. It's a complex program, spread across 5,000+ lines of regulations, policy guides, and advisory memos. Caseworkers spend years absorbing this as institutional knowledge, but families and childcare providers — the people the system is built to serve — have almost no way to access it on their own.

One of the main shortcomings of the childcare subsidy system is that clients who already have high literacy or that know the system well are the easiest for the system to help, while the clients that speak little to no English, or have more complicated subsidy qualification situations, are often the ones that are left behind. As an example, a Haitian Creole immigrant family I worked with had a tremendously difficult time obtaining and renewing a childcare subsidy, as some official documentation was only available in English, Haitian-fluent staff were few and far between, and both qualification and immigration status was complex. These cases can take a long time to process, stretching beyond the subsidy's deadline, putting children's care, and parent's jobs, at risk, and leaving the childcare provider in a state of limbo. On the contrary, a family who knows the system well, who have been through this and similar before, or that has access to assistance, is able to submit their complete documentation with ease. Those cases are often processed quickly.

There is an effort-resource mismatch here. The system lacks the resources and manpower to effectively help all clients equally, so effort gets distributed across the system and finds the path of least resistance first. There are not enough resources to give a complicated case the requisite care that it deserves. This is natural, and not an accusation or an indictment.

And so, to me, artificial intelligence presents a massive opportunity for equality, fairness, and a level playing field in a space where the intended design is exactly that. Here lies an AI agent that is trained specifically on EEC's documentation, and can respond to any question in that field in great detail, or respond simply, in the tone, language, and at the reading level of the person asking. It is an agent that can answer at all hours of the day or night, and do so without prejudice, without stress, without getting tired. It is also an agent that knows its own limits, and is not meant to answer all questions to their final point. The agent references out to the existing infrastructure (Mass211, CCR&Rs, DTA and DCF) as often as possible. It's meant as a safety net that catches those in need that would otherwise fall out of the system. It's designed to lower the threshold for families that lead busy and often stressful lives, and who don't have the time or energy to dedicate to learning the thousands of words of process and logic that are codified in official documentation. It's designed as a relief valve for case workers and providers who field tens or even hundreds of questions per day.

So you may wonder — is this meant to compete with humans? Or even replace us? Unequivocally, no. I'm confident that a tool like this will increase the number of clients that come through the metaphorical doors of the childcare system. Information is a resource that is often available to those only who have the time, money, or connections. This is a bridge across that. It fills a gap that has always existed but never been addressed. And in an increasingly AI-powered world, humans will be more important than ever in a field like this. The personal, human-to-human interaction, the care and empathy we offer each other, is irreplaceable.

Static website ←→ AI chat agent. Two interfaces built upon one source of information.

Groundwork is two interfaces drawn from one source of information. The website lets a user browse and search at their own pace. The chat agent lets a user describe a situation and get routed to the right answer. Both surfaces draw from the same 39 wiki pages — there is no separate AI training required. Think of the website as a library, and the agent as a librarian.

To make the contrast concrete: here is the same rule — Service Need, or what counts as an "approved activity" for a parent receiving subsidized care — rendered two ways. On the left, the regulation as published. On the right, the Groundwork wiki page covering the same rule.

Before · Official EEC Policy Documentation

Click to enlarge

After · Groundwork wiki

Click to enlarge

And the same Groundwork page, rendered in four languages — the user picks once per session, and the chat agent answers in the same language the page is rendered in.

English

Español

Kreyòl Ayisyen

简体中文

Here is the chat agent itself — first the introductory screen where the user picks a role and a language, then a typical exchange.

Structured lookup, not search. The agent doesn't guess at relevance. It routes through a small index that names every page and what it covers, then fetches that page directly.

Practitioner expertise lives in a separate layer. Policy as written and policy as practiced are often different at the ground level. Groundwork keeps them in separate files — the wiki says what the regulations say; a parallel annotation layer captures what a seasoned caseworker would tell you the regulations don't.

A hard handoff at the boundary. Groundwork never determines eligibility, submits a form, or gives legal advice. Every conversation ends with a referral upward.

Groundwork collects simple, anonymous telemetry data. The chat interface captures session-level signals — turn count, satisfaction (scored from the transcript by the LLM at session end), which human resources the agent referred users toward, which topics came up, language, device, and the queries that returned no useful answer.

It is completely anonymous by design — no IPs, no cookies, no cross-session identifiers. Conversations are summarized into a single sentence at session end, with explicit guardrails against capturing names, locations, family composition, or specific ages.

The instrumentation serves two purposes. First, operational: keep the system honest, catch content drift, see what's missing. Second, the data itself is policy-relevant. If a quarter of weekly sessions cluster around reauthorization deadlines, that's a signal the underlying forms are confusing — and a signal worth surfacing to the agencies that own those forms.

One million conversations would cost under $50,000 to operate.

• Not legal advice. Groundwork explains how the program works. It does not tell anyone whether they qualify, what to file, or how to appeal. Every conversation ends with a referral to a human authority.
• Not a replacement for caseworkers. Groundwork is a triage and orientation layer. The decisions, the paperwork, and the human relationship still belong to CCR&Rs, DTA, DCF, and the rest of the existing infrastructure.
• Not trained on private case data. The agent is grounded in published EEC regulations, policy guides, and advisory memos. No client records, no case notes, no personally identifiable information.
• Not a search engine. It does not guess at relevance. Every answer is retrieved from a specific, named wiki page that anyone can open and read.
• Not a black box. Every retrieval is traceable. Every wiki claim links back to a source document. The architecture is inspectable from end to end.
• Not finished. Provider-side practitioner expertise, additional languages, and adjacent benefit programs are ever-evolving.

I believe that the process used to create and operate Groundwork is applicable to a wide variety of government (and non-government) programs that have information as a primary gateway to access. Put more simply: any program where the complexity of information is a bar that needs to be crossed would benefit from a tool like Groundwork. I believe that, aside from being a financially accessible tool (it doesn't cost very much to operate), making information simpler and easier to access is vital to ensuring our communities have equal access to the benefits that are available to them. We're seeing the vision, the use case develop in real time as word spreads around the community, so I hope you find this article encouraging as well.

Feedback of any kind is always greatly appreciated, and can be emailed directly to us at groundworkma.admin@gmail.com. As mentioned earlier, this process is ongoing, and may always be ongoing, so we'll take your feedback and suggestions to heart.

Warren Zhao
May 26, 2026