In May, two researchers at the University of Southern California published a paper describing a tool I have wanted for fifteen years: a conversational AI agent that pulls thousands of fragmented federal and foundation grant portals into one place, reads your draft proposal, and surfaces the opportunities that fit.
The authors report it cutting grant discovery from the thirty minutes of manual portal clicking down to ten. In the past, I would have read it, admired it, and filed it under “things someone with a budget should build someday.” Instead I described what I wanted to Claude Cowork and by the end of the afternoon I had a working version.
Funder Finder: takes your focus areas or generic proposal PDF and it queries a unified index of roughly 12,000 federal and foundation opportunities, falling back to live web search only for the last few days, with every result linked to the real agency or funder page.
A New Way to Create Software
A non-engineer turning a peer-reviewed systems paper into a running web app between lunch and dinner would have been a fantasy two years ago. It is ordinary now, and the development sector needs to accept this new reality.
The reflex in our sector is still that real software requires grant funding and a procurement cycle. You convince leadership, write a proposal, wait for a win, then hire a consultant, write an RFP, evaluate vendors, and a year and a quarter-million dollars later you have a platform.
That reflex built an entire category of GenAI grant-writing tools to find opportunities and draft compliant proposals. The reflex made sense when building software meant employing dozens of software engineers. It no longer describes how a growing share of useful software gets made.
Sign Up Now for more software engineering for good insights
Built by User
My friend Alexis Bonnell calls this approach: Built by User. A domain expert who knows the pain, knows what success looks like, and is accountable for the result builds the software themselves, in days, for the price of an LLM subscription.
This is not low-code. The previous generation of “democratize software” tools, Bubble, Airtable, Zapier, gave you pre-built blocks to snap together, and the result was always constrained by what the vendor imagined. They remain powerful for standard workflows, which is why Gartner projects the combined low-code and no-code market will reach $44.5 billion in 2026. But when your problem is specific enough that no vendor has imagined it, low-code hits a wall.
AI coding agents work differently.
Claude Code and OpenAI Codex read a codebase, write and edit files, run tests, and produce standard code you own and can deploy anywhere. The output is bounded by your understanding of the problem, which for a domain expert is often deeper than any vendor’s.
You also inherit the maintenance, the debugging, and the security risk that comes with real code, a point I will return to because it is where most of the sector’s optimism is misplaced.
Built by User: Me
The evidence for Built by User is embarrassingly literal for me. I had never opened GitHub before May 2026. Matt Berg and Richard Stanley can confirm my lack of software engineering experience.
A month in, I now have 88 contributions across ten public repositories, most of them tools for NGO fundraising, grant compliance, and digital health built with Cowork and Codex. For example:
- Funder Finder, which I explained above
- Claude Skills Creator to improve Claude for NGOs
- Grant Fit Auditor that tells an NGO when not to apply,
- NIH Subaward Checker built for a single 2026 NIH notice,
- Payer-contract simulator to show how AI optimized contract goals
- College Fit Finder for my teens to find a university
- Senate Testimony Strategy for people testifying
- First Signs Health to democratize dementia screening
Commit count is a crude measure and most of those commits are small. The shape is the point: from zero to a working portfolio in a month, by someone who could not have coded his way out of a paper bag in 2025.
The Tool I Care About Most
In 2023, researchers at USC’s Brain Health Observatory analyzed Medicare data covering 40 million Americans over 65 and found that only about 8% of expected mild cognitive impairment cases had been diagnosed, leaving an estimated 7.4 million cases undetected. A companion study found that primary care clinicians underdiagnose MCI in roughly 99% of cases.
The reasons are structural: fifteen-minute visits, weak reimbursement, no standard protocol for what happens after a positive screen. Meanwhile the two FDA-approved Alzheimer’s drugs, lecanemab and donanemab, only help patients still in the early window. Miss the window and the patient is no longer eligible.
This matters to me because my mom has dementia. She was diagnosed too late for either drug.
No product on the market lets a family caregiver run a validated screening protocol at home and walk into a doctor’s office with a clinician-ready packet specific enough to trigger a referral.
So I built one.
First Signs Health combines four validated cognitive instruments in a three-gate browser protocol. It does not diagnose. It produces a handoff document a physician can act on. I built it in days, for about $100 a month in tools, as a non-engineer who became a functional expert in dementia because my family forced me to.
Built by User in Software Development
The private sector market has already reacted to Built by User. In late January 2026, Anthropic released industry-specific plugins for its Claude Cowork agent. By February:
- Reuters reported Thomson Reuters falling nearly 18%, its largest single-day drop on record.
- CNN put the broader software selloff at the center of investor fears that the seat-based SaaS model is at risk.
- Fortune described the cumulative rout as a trillion-dollar event once OpenAI launched its enterprise Frontier platform.
Attributing stock moves to single product launches is unreliable, and these figures span advertising, consulting, and financial data well beyond SaaS. The direction, though, is not in doubt.
SaaS is not dying. It is splitting.
Single-purpose tools whose value is organizing, searching, or synthesizing information are the most exposed. Tools whose value is maintained data infrastructure, workflow orchestration across complex systems, or regulatory compliance have far more runway.
No vendor was ever going to build First Signs Health – there is no profitable market for it today – and a tool that handles HIPAA-compliant storage, integrates with an electronic medical record, and maintains audit trails still requires enterprise engineering.
Built by User does not threaten complex SaaS. It may make it more necessary, because it puts more half-built software into the world that eventually needs to connect to real infrastructure, and real infrastructure is still somebody’s full-time job.
We Need More Software Engineers
The optimism I most distrust is the idea that Cowork and Codex displaces engineers. It will not. It will change what they do.
The Bureau of Labor Statistics projects 15% employment growth for software developers, QA analysts, and testers from 2024 to 2034, much faster than average, and CNN reported in April 2026 that developer openings are rising precisely because non-engineers can now prototype, increasing demand for people who can finish the job safely.
The reason is failure modes.
Veracode’s 2025 GenAI Code Security Report tested more than 100 large language models across 80 tasks and found that AI-generated code introduced security vulnerabilities in 45% of cases, with no improvement from newer or larger models. A printed screening checklist fails visibly. Software fails invisibly.
My dementia tool uses published scoring protocols, but a poorly-built screener could miscalculate a score and reassure someone who should be seeing a neurologist. The engineering judgment that catches that is becoming more valuable, not less.
Built by User Future in ICT4D
1. Please stop treating a $250,000 custom platform as the default.
A foundation about to fund one should first ask whether a program officer, working with a coding agent for a week, could prototype the core functionality. That experiment will reveal which 20% of the build actually requires professional engineering, and it will expose how much the cost structure of social-sector technology has already changed.
2. Create a governed Built by User process.
Let domain experts build and deploy the simple things, no sensitive data, no authentication, no production integration, with explicit guardrails, and pay engineers to handle everything that touches real data or real decisions.
Funders that finance digital tools should ask grantees a blunt question they cannot currently answer: who maintains this, and who reviewed the code. I have argued before that more proposals written with AI is not the same as more good proposals, and the same caution applies here.
3. More software is not the same as more working software.
The dementia screening gap will not be closed by a fifteen-minute primary care visit, and it will not be closed by a SaaS product built for the largest addressable market. It will be closed by people like me who see the problem up close and now have the tools to act.
Whether those tools are safe enough to trust depends on whether we pair domain expertise with engineering judgment, instead of making one wait in line behind the other. The software is now the cheap part. The judgment is not.
