
This is the Built By User reality today: A community health officer tells a program manager on a WhatsApp call that an input field in a digital health tool is wrong. The manager starts Codex on the fix within minutes of hanging up. The health worker wake up the next morning and the corrected version is waiting.
The total cost: a phone call and a handful of credits.
In June I argued that $250,000 NGO software procurements are dead and offered my own portfolio as proof. The obvious objection came from Benedict Evans, who took apart the idea that cheap tooling makes everyone a builder.
- Most people never see that their work could be automated.
- Of those who do, most are not the right person to design the fix.
- And of those who could, most cannot get anyone else to adopt the result.
He is right on all three, and the overnight loop above answers only the second one. That still changes something the sector has never had.
Sign Up Now for more software developmnet insights
Feedback Only Works When Acting on It Is Free
Human-centered design has been the sector’s stated method since the Principles for Digital Development put “Design With the User” first. A narrative review in Global Health: Science and Practice reports that USAID has tracked sharply rising HCD use in global health since 2016.
So the method is not new. What has always been broken is the latency.
Three weeks of interviews, a design workshop, a requirements document, a twelve to eighteen month build, then a pilot, then feedback arriving after the budget is closed and the code is frozen. By the time the health officer says the input field is wrong, changing it costs a contract amendment.
The loop exists on the workplan and is ceremonial in practice, which is a large part of why co-designed tools still fail on contextual fit, a pattern visible in global digital health chatbot pilots and in the research on optimizing for community usage context.
Compress that loop to overnight and the health officer stops being a research subject.
They become the person deciding what the software does, without writing code, attending a workshop, or spending anything but a phone call. Two decades of ICT4D capacity building assumed the endpoint of local ownership was a health worker learning to maintain a codebase in their spare hours.
That was never going to happen. Overnight iteration gets you the authorship without the fantasy.
The Built By User reality does not build alone and ship blind. Automated software agents like Codex write test scripts against the code, and Claude can check the results, the health officer tests before use, and errors fold into the next night’s build.
That is more review than most sector software receives.
It also has a limit worth naming: the code writer, the test writer, and the reviewer share a model family and therefore share blind spots, and every stage verifies behavior against intent rather than hunting for what nobody thought to ask about.
Veracode tested more than 100 models across 80 tasks and found AI-generated code introduced security vulnerabilities in 45% of cases with no improvement from newer or larger models.
Acceptance and Adoption Are Hard Tests
Vulnerable dependencies, injection paths, exposure through logs: these are mostly mechanical, and tooling handles them whether or not the tooling is AI. The class that resists automation is authorization logic, because whether one user can reach another user’s record is a claim about intended rules that appears nowhere in the code.
Everything above is still the easy part.
- Will the health officer use the tool next week?
- Will their colleague in the next district?
- Will the supervisor accept its output?
Those questions were the bottleneck before AI and they remain the bottleneck, exactly as Evans argues, which is his broader point that AI is a new flavor on dynamics that are thirty years old.
The ratio is what changed. When development cost falls toward zero and adoption cost holds constant, adoption becomes nearly the entire cost of the project whether or not anyone budgeted for it.
Two Questions, Not One
Reach and risk are separate axes, and collapsing them produces the cell that hurts people.
| Low risk: no persistent state others depend on | High risk: holds data others depend on, or feeds a decision | |
| Low reach: one user, one context | Disposable. Ship overnight. Rebuild it in a weekend or let it die. Death is completion, not failure. | The trap. Small enough to feel casual, serious enough to harm someone. Architect the data out, or do not build it. |
| High reach: a workflow many people run | Convert. Worth ministry engagement, a vendor, a contract, a maintenance line. | Not a Built by User target. The $250,000 procurement with engineering, procurement, government ownership. |
Cells are not destiny.
My dementia screening tool, First Signs Health, sits in the top left only because I designed it to produce a handoff document instead of storing patient data. The same tool writing to a record sits in the top right. Stripping persistence is the single most transferable move in my portfolio.
The top right is where my own method is most dangerous. One health officer, one WhatsApp call, an overnight ship, patient-level data. Every incentive in fast iteration points at that cell because it feels small.
Reach is a prediction, not an observation.
My test is whether the ministry has already named the problem as national in its digital health strategy, HMIS gap assessment, or sector priorities. Building for one officer’s problem is fine and cheap. Building for a country-wide problem that happens to be personified in one officer is the only thing worth converting.
Deciding which is which requires knowing the national portfolio, which is the hardest analytical work in the chain and the part that does not generalize to anyone with a subscription.
Cheapness Makes Absorption Harder
A ministry can procure a $250,000 platform.
There is a vendor, a contract, an invoice, a support agreement, a line item. There is no procurement mechanism for the thing an American built over three weekends for a hundred dollars in tokens. No public financial management system has a category for it.
The economics that made the tool possible strip it of every institutional handle a ministry uses to take ownership. Absorption was already the wall, and why digital health pilots never scale and why we need to design for donor exit. Cheap software raises it.
So funders should stop financing builds and start financing conversions.
Take a working prototype and pay for what makes it procurable: a vendor of record, a contract, a support agreement, a maintenance line the ministry can carry. That is a fraction of what the old build cost, and it does the job the old build never did.
The New Budget Rule
I advocate that we reject any digital tool proposal where the development budget exceeds the adoption budget.
I searched for a published benchmark ratio of development to adoption spending in digital health and global health implementation. I could not find one. The literature says adoption is expensive and chronically underfunded and never quantifies it against build cost.
So treat my threshold as a proposal, not a finding, and produce better numbers if you have them.
Program officers did not write $200,000 for development and $20,000 for change management because they thought adoption was easy. They wrote it because donors fund capital rather than recurrent costs, development produces a milestone you can invoice against, and adoption is three years of training, supervision, integration, and someone answering the phone, which disappears into an indirect rate.
The ratio was a financing structure, not a judgment.
Define adoption before anyone games it. Training, supervision, integration work, ministry engagement, and conversion budget count. Anything that produces code does not.
Every digital health organization can apply this rule to the next proposal that crosses their desk. If the applicant cannot explain why building costs more than getting people to use it, they have not thought about the second problem at all.


There is a saying that “good judgement comes from experience….and experience comes from bad judgement”.
Sure, the general public can create magnificent, incredibly complex castles of complex, multi-tiered software in the clouds, that could be written in any one or many Tower of Babel of languages (Python, Java, Flutter…)
But … a chain is only as weak as it’s weakest link. What will establish the trust necessary to run “mission-critical” services (banking, retail, healthcare, etc) on such things? What entity will actually ensure these castles are tested thoroughly, and have rollback and hardened, audit-ready logs for accountability?
In a world where everyone can build airplanes, establishing the public’s trust to step aboard a shiny new one built by AI and take off won’t come for free.