Benefits navigation and the potential of AI

Introduction: AI’s potential in government benefits navigation#introduction-ais-potential-in-government-benefits-navigation

Dario Amodei, the CEO of Anthropic, published an essay in October of 2024 in which he articulated a more concrete vision of the positive potential of AI on society at large. In that essay, he wrote:

There is also a clear opportunity for AI to be used to help provision government services—such as health benefits or social services—that are in principle available to everyone but in practice often severely lacking…

Having a very thoughtful and informed AI whose job is to give you everything you’re legally entitled to by the government in a way you can understand—and who also helps you comply with often confusing government rules—would be a big deal…

[T]hese are somewhat vague ideas, and as I said at the beginning of this section, I am not nearly as confident in their feasibility as I am in advances [in other domains]...

At Propel, this problem — the problem of "benefits navigation" — is one we are actively working on and applying AI to.

In this essay, I aim to accomplish something analogous to Dario’s goal, but more narrowly for those interested in the problem of “benefits navigation,” proposing answers to:

• What is the benefits navigation problem, in more specific terms?
• Given that more specific problem definition, in what specific ways do new AI capabilities present opportunities?

What is the benefits navigation problem?#what-is-the-benefits-navigation-problem

The simplest framing goal often stated around government benefits (including in Dario’s essay) is: everyone should get every benefit they are entitled to.

The fact that this is not already true points to the need to identify the constraints or problems behind this. I believe there are two core challenges underneath navigation:

1. Information problems are ubiquitous in the process of accessing benefits
2. There is a tension between individual goals and system goals that drives how things work today

The information problems come in a number of forms:

• Information asymmetries (gaps in knowledge between the person and administrative staff)
• Information retrieval costs (getting access to the rules and policy documents that govern the benefit)
• Information processing costs (pulling out the relevant rules from a policy document, reasoning through rules for a particular situation, making sense of complexity)

The goal alignment problem is that government benefits programs have far more goals that govern their design and administration beyond the individual's own goals and interests. Administrators must constantly balance competing mandates when implementing programs.

For example, beyond delivering benefits to eligible people, an administrator also must consider goals such as minimizing administrative costs, collecting documentary evidence of eligibility to meet payment integrity goals, and compliance with requirements dictating specific processes or inclusion of verbatim legal language.

System goals overlap with — but do not directly align to — the goals and interests of individuals navigating benefits: most commonly getting, maximizing, and keeping benefits with minimal hassle and effort. [1]

These are the two main problems that create a need for “benefits navigation” support in the first place. They also help us get to a firmer definition of what “benefits navigation” means:

Benefits navigation is the process of helping a person with benefits in alignment with their goals, largely by solving the information problems that stand in the way of achieving those goals.

This definition also gestures toward where much more powerful AI capabilities presents opportunity: the cost of many information operations falling to near zero.

Information costs in navigating benefits#information-costs-in-navigating-benefits

AI presents significant potential for this problem precisely because so much of the work of successfully navigating benefits takes the form of costly information operations. For example:

• reasoning through long policy documents
• extracting core meaning from complicated text
• applying general policy to specific situations, etc.

The potential of this new generation of AI is that the cost of many of these information operations appears to be approaching zero. To the extent these capabilities advance and costs fall, more benefits navigation supports become feasible, at a lower cost.

Here is how AI can help address the three main information problems mentioned at the beginning:

1. Information asymmetries

Modern AI can process and use the information contained in thousands of pages of policy manuals, court decisions, and federal guidance. Most applicants would never be able to access or process these, and so this can create parity of information access with the administrative staff they interact with.

• SNAP example: A person can be told up front that, based on federal rules, they can ask the agency for help if they cannot provide requested documentary proof, and the agency must help them

2. Information retrieval

Modern AI can continuously monitor policy changes across jurisdictions and incorporate updates; extract the relevant subset of policy for a decision; identify more options available.

• SNAP example: AI can find and extract the relevant procedures for a given person’s state from that state’s policy manual

3. Information processing

Modern AI can translate complex legal notices and other complicated information sources into plain language explanations of what matters for them and why.

• SNAP example: Extracting the core relevant information from complicated legal notices, and assessing needed action based on the individual’s goals

Some other concrete examples of where new AI capabilities enable more navigation:

• Software can reduce the burden of certain tasks, such as filling out forms or labeling documents. If AI code generation can make such burden-reducing software cheaper to build, maintain, and customize, then many more personalized tools may become feasible
• AI can extract core information from notices and letters, as well as identify and add for context relevant policies that may not be mentioned on the letter, but are relevant to the person’s interests.
• AI can enable customization across varied jurisdictional contexts, processing information across 50 states where previously only a handful of large states may have been feasible to support.
• AI can address concerns about intimidating language in long letters by helping the person understand if it actually applies to their situation (e.g. if a work requirement letter requires immediate attention or is instead a more general notice sent to everyone, but only some people are subject to)

These capabilities can be bundled into specific supports for people navigating benefits, where the information problems and costs create barriers today.

A taxonomy of benefits navigation supports#a-taxonomy-of-benefits-navigation-supports

The specific types of help that “benefits navigation” entails can be described in a few discrete types of help. Each category addresses different aspects of the information problems described above:

A. Decision support
Helping a person make a choice among options presented

At many points through the process of accessing benefits, people face decisions and given a set of options to choose from.

For example, in SNAP, someone might have to decide whether or not they need to report getting new income if it is a small amount. Navigation assistance can provide information that lets them make the right decision for them (e.g. is it required under the rules? when does someone need to report it by?)

Often someone might remember that they have a requirement to report getting a new job, but practical details may not be addressed in the letter they have about those requirements.

Instead of having to call their agency, the ability to immediately pull up what more detailed policy documents say about that practical question — e.g. report within X days after your new job has paid you, not before — provides meaningful support. And “extracting relevant policies” is a capability AI appears to be making significantly cheaper.

B. Expanding available options
Surfacing "hidden paths" not presented by default

This is an important intervention in benefits navigation that is often overlooked. It is particularly relevant for the goal alignment problem described previously.

A person accessing benefits often has more options available to them than what is presented by default, without navigation assistance.

For example, in our recent SNAP notice experiment, one legal aid expert who used the tool with a “notice of missed interview” in their state noted that one of the most valuable pieces of information we could provide someone receiving that notice was nowhere on the notice itself. So our tool’s initial focus on “translating” what the notice said was incomplete for the need.

Specifically, when they were helping clients with this, they made sure to tell people that they have a right (per policy) to request an in-person interview — not simply to call for a telephone interview, as the notice suggested.

This was an option not made known to the person by default, but was perhaps the most helpful option to know about if, for example, their state is facing major call center backlogs.

I refer to these options as “hidden paths.” They exist, but a person does not find out about them when navigating on their own.

Providing people all the paths available to them is one of the highest value aspects of benefits navigation, and one of the key ways in which external navigation support provides differentiated help.

C. Reducing the burden (cost) of a part of the process
Making underutilized choices less costly

This is where software technology has most frequently been applied in benefits navigation previously.

For example, taking a complicated paper form or PDF and creating a simple version, fillable on a smartphone, including immediate feedback on any validation errors that could cause delays later.

In a way, this is less strictly navigation, because it is less related to wayfinding through choices in the process.

But the ability to make some choices less costly to take — e.g. generating template language for someone who wishes to appeal a decision — can increase how many people make use of that choice.

D. Psychological support
Assuaging fear, anxiety, or stigma

Technology has generally played less of a role in this aspect of benefits navigation, given the outsized impact a person-to-person relationship can play here. But it is a component of many benefits navigation services.

For example, a person may need reassurance about intimidating language or processes they encounter to continue.

Concretely in SNAP, the recourse to appeal a decision is a “fair hearing.” To some people, just this phrase is intimidating.

Providing extra, contextualizing information — that it can be by phone instead of in person, that you can bring someone with you, that often the agency will fix a problem before a hearing happens — can help reduce the psychological costs of making a choice.

The decision leverage gradient: why some moments matter more#the-decision-leverage-gradient-why-some-moments-matter-more

Not all navigation moments are created equal. There exists what we might call a "decision leverage gradient" across the benefits journey, where certain decision points disproportionately impact outcomes.

High-leverage points share certain characteristics:

• They often involve binary outcomes with lasting consequences (approval vs. denial)
• They occur at peaks of information asymmetry (where the system provides least guidance)
• They occur when cognitive load or stress is highest (reading a notice to repay thousands of dollars)
• Their effects compound over time (providing fewer proof of deductible costs, leading to a lower benefit amount for every month going forward)

Some concrete examples of higher-leverage points that we have been experimenting on in our work applying AI to benefits navigation in SNAP:

• Missing an EBT deposit (and the paths to restoring benefits)
• Being denied benefits (and the opportunity to appeal)
• Identifying whether benefit amounts are maximized (by assessing unused deductions)
• Understanding the appeals and fair hearings processes

The point is that benefits navigation may provide benefits in general like reducing the burden in cognitive load or time for people. But targeting navigation support to the narrower, highest impact set of decision points people face has the highest return on the work.

“Unwritten rules” and the value of navigation at scale#unwritten-rules-and-the-value-of-navigation-at-scale

It’s worth noting that an important aspect of benefits navigation as provided by on-the-ground organizations such as legal aid or community-based groups is learning what actually works (or doesn’t) through interaction with the benefits system — at a scale beyond one person.

Policy documents what should happen, but is often incomplete for determining what does happen. To be clear, I am not describing this as a moral failing, but rather the simple reality that all benefits programs operate under constraints, and while policy sets a direction, the process people actually navigate often deviates from policy due to these constraints.

In SNAP, as in other benefits programs, a common driver of deviation from policy is understaffing. If there simply are not enough people to process workload, the existence of a policy requirement for something like “timely processing” (30 days for an application) does not overcome that constraint. This is a reality that people navigating the process must work within.

Here, too, however there is potential from AI in changing the cost to provide benefits navigation. If AI can increase the scale and scope of benefits navigation services available, then it can also increase the aggregation of practical knowledge from people’s actual experiences (what works, what doesn’t, how) and redistribute that knowledge and advice back to others navigating.

Implications for others building in navigation technology#implications-for-others-building-in-navigation-technology

At a practical level, this way of conceptualizing benefits navigation (as information and goal alignment problems) and how AI fits in specifically (reducing the cost of information operations) can provide some direction for those looking to apply this new technology to other forms of navigational technology, in the benefits domain or beyond. This is by:

1. Identifying high leverage decision points where more navigational help yields outsized outcomes
2. Looking for “hidden paths” where people have options not presented to them by default when navigating on their own
3. Applying AI/LLMs as a new capability set (and fall in cost) for information operations — a framework for deployment that looks quite different than simply “better chatbots”
4. Building compounding systems that learn from collective experience

Rather than digitizing existing processes or creating basic informational tools, what we are seeing rapidly developing in AI can enable more ambitious forms of benefits navigation technology than what has been feasible with software up to now.

My hope is that this essay has provided a more concrete picture of what that could look like, and how we might achieve it.

Thanks to Alan Williams, Jacob Solomon, and Claude 3.7 Sonnet for discussing drafts of this essay with me.

[1] AI also offers significant potential for improving the administrative machinery of government itself. I consider such improvements distinct from “navigation” supports by definition, because of the goal alignment mismatch problem described. Such improvement on the administrative end will improve the capacity of these programs to achieve system goals, but not necessarily in direct alignment to the interests and goals of the individual.