Why Now? The Breakthroughs Powering the NavTech Opportunity

Consider a mother of two in the week after the warehouse closes, taking her job of six years with it. She scrolls through an endless application for assistance that asks her to prove her residence after she’s already proven it three different times. She sits through mind-numbing hold music just to be told there isn’t an answer to her question.

At best, she gets a cookie-cutter solution — the same generic help that might not fit her unique situation, and might not make her much better off. At worst, she falls through the cracks and slips further down the economic ladder.

For generations, we’ve been forced to compromise: systems could serve millions, but only by offering generic, one-size-fits-all solutions. Meanwhile, true, personalized support — the kind we know actually changes outcomes — remained painfully scarce.

There have never been enough caseworkers, counselors, and personal advisors, and never enough money for more. Meanwhile, people with resources increasingly pay for expert support — a financial advisor, a lawyer on retainer, a college counselor, a doctor who picks up the phone — only deepening a persistent inequality gap.

But we have a new opportunity to start reversing this trend.

The AI capabilities we have now could unlock decades-old bottlenecks all at once: the steep cost of expert guidance, the constrained supply of expertise, the challenge of connecting databases and platforms built decades apart, and the maze of doors and obstacles people have had to move through. The result is a new opportunity for what we call Navigation Technology, or NavTech — tech-enabled tools that provide specific support at scale, built with and for the people they’ll serve.

And these tools are here as life grows more complex and barriers to prosperity grow taller. We live in an era when 90 million people live in households struggling to make ends meet, and nearly everyone has felt overwhelmed and stretched thin. All Americans deserve to benefit from the rapid advancements in AI capabilities, not just those closest to the technology itself.

For that to happen, we at NextLadder Ventures believe that we must build a robust and sustainable market for these types of solutions, with industry-leading standards for quality and safety and category-level infrastructure that no single builder can create alone. We also need benchmarks and evaluations, rooted in the right ground truth, to test these tools — which will be valuable to everyone, but hard for anyone to build alone. That’s why we’ve brought together expertise across venture capital, philanthropy, user insights, quality and safety, measurement and impact, and partnerships to seize those opportunities. In my role, I spend a lot of time thinking about the technical side of these products: the underlying innovation that powers them, the systems we’ll need to create truly people-centered technology, and the potential and limitations.

As AI technology has advanced, the NavTech moment is being unlocked by breakthroughs along five vectors: comprehension, memory, connection, action, and prediction. In this piece, we’ll trace each advancement, explain how it unlocks groundbreaking capabilities, and explore how, compounded together, they enable powerful new systems of action.

Vector #1: Comprehension

The arduous work of navigation — learning about, locating, and locking down the support you need through life’s toughest challenges — often requires wading through thousand-page medical histories, dense court filings, fragmented housing listings, or reams of rules buried on a government website. Before LLMs, no system or person could digest all of that information — and the first wave of LLMs couldn’t take in the vast volume of complex input.

In 2020, the context window for a frontier model was roughly 2,000 tokens — just a few pages of text. Today, the baseline is over a million and counting (opens in a new tab). That’s enough to hold a family’s entire file and a full policy manual at once.

Meanwhile, retrieval methods — like knowledge graphs and vector embeddings — can now map how pieces of information relate, capturing the relationships that inform outcomes (when one rule supersedes another, or what a full financial picture adds up to), where accuracy is paramount.

Picture a career planning tool that can take in your full work history — every job, certification, and gap — and give you guidance that’s tailored to you. Not a list of job postings, but an actual path to follow, and the first step on it. Or a benefit filing tool that can safely process thousands of pages of medical history against a complex set of requirements to help someone document the right case.

While frontier models will continue to advance in reasoning, we still see reasoning degrade across extremely large contexts today. Purpose-built tools that use supplemental context layers to deliver the right context for the task (for example, a state’s rules or a set of eligibility thresholds), continue to be particularly valuable for NavTech applications that aren’t well-served by generic models.

Vector #2: Memory

A long-term understanding of someone’s situation is one of the most valuable parts of providing the right support. That information might live in a caseworker’s head or a file that can’t leave the building. Most often, it stays with the individual themselves — as they navigate between systems and people in their journey, forced to re-explain themselves each time. Early AI models were similar, treating every session as a blank slate. That was serviceable for quick prompts, but entirely useless for complex, long-term tasks like a four-month housing search.

Now, recent advancements have gone beyond models to the harness architecture that surrounds the base model with software, memory systems, and tools. Systems can now retain context with every use and store it securely — but also know when it’s useful to bring it back in future exchanges. Persistent memory turns navigation into a more empowering experience. Instead of starting from zero every time — retelling the story, re-uploading the documents — a person picks up where they left off, with a tool that already knows their situation.

Importantly, we believe the data that builds this context should be owned by the user in a portable, durable personal context layer — a secure, consent-driven, user-owned personal record that lets them share stored memory across platforms where they’re asked repeatedly to provide the same information (like income, work status, household details). We’ll have more to say on that in the months to come.

Vector #3: Connection

For too long, it was hard for systems to talk to each other. Integration costs scaled with pairs: every tool, times every system, meant its own custom build with brittle, often expensive custom connectors that had to be maintained. Every database had its own schema. And bridging any two was a drain on an organization’s money, time, and attention.

Two recent developments changed that math.

First, models can reliably translate messy, fragmented inputs into the exact format another system expects (target schema), with built-in checks (validation layers) that ensure accuracy. That means a model can assemble a personal record based on data extracted from pay stubs, leases, benefit notices, and court filings — and match it to the format needed. That’s work that could otherwise take a caseworker their full afternoon.

Second, new standardized protocols like the Model Context Protocol (MCP) let a system owner stand up one well-built server that any application can use. In other words, once a state government creates a single connector for its benefits system, any tool can plug into it. No custom integration for each new tool, no waiting years for each build. Just quicker, more reliable results, at a fraction of the cost. And with MCP owned as a community standard, it’s a great example of the type of shared market infrastructure that makes NavTech possible.

Vector #4: Action

Knowing what to do is one thing. Getting it done is another.

Moving from text generation to agentic orchestration allows models to execute complex, multi-step workflows on someone’s behalf. Combining specific tool calls and reasoning loops with larger context windows enables autonomous action — a major leap for NavTech that turns passive advice into real execution. Filling out forms is the smallest example: imagine how much easier it could be to file for bankruptcy or unemployment — whether you’re one family or a professional helping twenty.

And in the process, agents are changing how software gets built. In the software era, bridging fragmented government systems meant millions of dollars in modernization contracts and years of work. In the agentic era, that equation changes: whether through well-built connectors that can do what once required replacing the system outright or using agentic tools to build a new modular system with greater controls. Eventually, families should get a seamless experience, no matter how convoluted the system.

We still have a lot to learn about how we can safely deploy agents in settings where a family’s livelihood is on the line. But the evolution of evaluation techniques is giving us new ways to measure and continuously monitor whether systems work as intended end-to-end. Reliability still compounds across steps — 95 percent per step is roughly 60 percent over ten — so today’s most trustworthy agents are narrowly scoped, with transparent verification loops, extraction tested against ground truth, and logic validated before any action. Evaluated performance is what will allow institutions and individuals to adopt tools with confidence. And as accuracy and reliability increase, it also opens up opportunities for personal agents to help people safely navigate, though these tools are still on the jagged frontier and will require thoughtful testing, policies, and protocols.

Vector #5: Prediction

Providers juggling enormous caseloads don’t have bandwidth for proactive outreach. And until recently, applying machine learning (ML) models was usually limited by the thin, often lagging indicators available for public services: a missed rent payment, an eligibility lapse.

But modern long-range ML can run continuous inference across messy longitudinal data, including case notes, call transcripts, uploaded documents and more — meaning improved pattern recognition and the ability to spot outliers even in smaller datasets. For example, a caseworker stretched thin could have a tool that flags which five clients out of their hundred are most in need of high-touch support this week. Importantly, this triages scarce attention, and it doesn’t score people for approvals, an approach with a long and cautionary history.

For the person on the other end, this could mean that help finds you — not one more system to find.

How the Vectors Converge

As these five vectors converge, the economics of providing navigation support fundamentally shifts.

The costs of building (infrastructure, time, and the cost of software) have fallen, and costs of ongoing model use continue to decrease. And importantly, the marginal cost of serving one more person has dropped dramatically, relative to the impossible benchmark of staffing every family with a housing navigator, career coach, and benefits caseworker. So now, NavTech can make providing expert-quality help for every family a reality.

And NavTech specifically — more so than general-purpose AI. Generic models train on the internet’s averages. That means, when you’re in need of financial advice, it defaults to the internet’s idea of financial advice and helps you pick your next stock. But a purpose-built tool knows the right debt management plan exists for you and helps you get signed up. These types of tools are grounded in a specific domain’s rules, evaluated against particular outcomes, and shaped by the expertise of human navigators — who haven’t just learned the facts, but helped people actually act on them.

And to be clear: This technology should augment the work of navigators, giving them back their hours: less time wasted on paperwork, more time spent on people.

Now go back to that mother of two we talked about before, the week the warehouse closed. What has typically followed is a torturous maze: an unemployment claim on one portal, SNAP on another, a job search across a dozen boards. Each with its own slog through bureaucracy, each starting from square one.

But NavTech tools could transform her experience. They could read her situation holistically and steer her toward what she qualifies for and what to do first (comprehension). They could recall her week to week (memory) and assemble records scattered across agencies into one picture (connection). They could file the claim, book the appointment, and move the paperwork (action). When she’s about to miss a recertification deadline, they could flag her caseworker before the lapse (prediction). And these tools can meet people where they already are — like in the messaging apps they use every day — and whenever they need help, which is often outside of office hours (opens in a new tab).

Not long ago, every step was prohibitively hard or prohibitively expensive. Today, tools to address these very problems are already being built. Like CareerVillage’s Coach (opens in a new tab): an AI-powered counselor that can answer questions related to 30 different career navigation challenges and has already supported 13 million people. Or SaverLife (opens in a new tab): which integrates AI into financial planning, expanding access to high-quality, personalized guidance that was once reserved for well-off families.

At NextLadder Ventures, we invest in tools like these that are co-designed with families and frontline navigators, led by builders proximate to the problems, with incentives pointed at outcomes — like dollars in pockets (opens in a new tab) unlocked — rather than reach.

Not a Moment Too Soon

While game-changing, these technologies are not magic wands that will erase decades-old inequalities. They’re tools that can be part of making a person-centered future possible. But together, we can make it inevitable. We’re at an inflection point, with the default settings and foundational architecture for AI systems being written. The next couple of years will be critical and give us a unique, generational opportunity to build systems that actually work for low-income families — and prove that AI can deliver real value.

The families we aim to empower do not get help by default. It’s up to motivated builders and investors to determine whether our AI future is one that lends a hand or leaves people behind. So join us in building tools that work for the people who need them most — and for as many of those people as possible.