If you work in national statistics or monitoring and evaluation within the global development space, you know how important digital questionnaires are. If the survey is badly structured, the data quality suffers; if the data quality suffers, so do the decisions that follow.

That makes survey programming a natural place to look for AI support. The work is repetitive in places and full of logic that can in theory be translated into code. But based on my recent experience scripting CAPI surveys, I think it helps to separate AI’s role into two distinct branches: survey scaffold automation and skip/validation programming. In the long run these will probably converge. Right now, they are still different problems.

Survey Scaffold Automation

A lot of questionnaire programming time is spent on work that is important but not intellectually glamorous: entering question text, setting up answer options, creating modules, and building the basic questionnaire skeleton before any real logic is added. On large instruments, that can take days.

AI is already genuinely useful here. It is very good at taking the “paper” questionnaire—usually a Word or Excel draft—and converting it into structured inputs. The catch is that success depends heavily on the CAPI platform and how easily it accepts imported data. AI likes producing structured outputs, so systems that can process those outputs cleanly are the easiest to automate.

For ODK-based platforms such as SurveyCTO, Kobo and MWater, this works well. The questionnaire can be converted into an XLSForm workbook, and that matches the strengths of current AI systems. Six months ago this often meant copying and pasting content from a chatbot into multiple Excel sheets. Now, with agentic tools like Codex or Claude Code, the file can often be produced end-to-end with the correct template and worksheets already in place.

For CSPro, the process is more piecemeal but still promising. CSPro works directly with files on the desktop, which actually makes it easier to automate than many newer web tools. AI agents can be guided to create a `.dcf` dictionary structure and a matching `.qsf` file for question text. It is old-school, but in a way that works in AI’s favour.

For Survey Solutions, this is where things get harder. The core difficulty is that the standard designer does not provide a proper import pathway for questionnaire creation, so the default workflow is manual point-and-click entry. I first tried browser automation. It was possible, but mediocre: low quality, slower than a human in practice, and requiring too much supervision every 10–30 minutes to keep the agent on track. More recently I found that the open-source local version includes admin functions to import questionnaire “backups” as JSON. That is much more promising, because it gives AI a structured target. But it also requires running your own version of the system locally, which is obviously not a route that suits everyone.

Skip Patterns and Validation Logic

I treat logic programming as a separate task because AI often gets confused when it is asked to build questionnaire structure and write complex logic at the same time. The logic layer needs its own step.

This is also where model quality matters much more. Questionnaire logic can get genuinely difficult: nested rosters, flowing skips, cross-field validations, consistency checks. If you want AI help here, use the strongest model you can access and turn the reasoning up.

Again, results vary depending on the survey platform used. For ODK/XLSForm, results are strong. Underneath, the logic is XPath wrapped in the XLSForm standard, and current models seem to understand that reasonably well as long as the prompt is explicit.

For CSPro, AI works well. The language is bespoke, but well documented and old enough that its patterns are clearly present in model training data.

For Survey Solutions, over the last several months of testing I have seen progress but there remains challenges. As I wrote in an earlier blog, generic AI has struggled with its constrained C# environment and platform-specific functions. That is why I built a custom Survey Solutions chatbot to improve results. Newer top tier models such as Gemini 3.1, Opus 4.6 or GPT 5.4,  are getting better without the custom prompt stack, but they still miss some of the important Survey Solutions nuance.

What AI still misses

The biggest limitation across all platforms is not syntax. It is judgement about what is missing. AI mostly works from what is written in the questionnaire documents. But experienced survey programmers know that a large part of the job is spotting what the paper instrument failed to specify: missing validations, inconsistent skips, implied rules that were never written down, and places where the logic simply does not hold together. Even when prompted to “add validations,” AI tends to produce a small set of basic checks rather than the deeper consistency thinking that good programming requires.

So where are we now? AI is already very useful for scaffold generation, increasingly useful for logic programming, and still weak at the questionnaire design review work that experienced programmers do almost instinctively.

That will change. Model improvements are arriving almost weekly, and agentic systems are getting better at multi-step tasks. I do not think it will be long before these branches come together more effectively. But for now, the practical approach is to treat them separately and use AI where it is strongest.

This is exactly the kind of problem we care about at Impact Engines: not AI for its own sake, but practical workflows that help organizations move faster, work to higher quality, and become more self-sufficient in technical tasks they may not previously have had the capacity to handle.

Impact Engines builds practical, responsible AI tools and workflows for development and not-for-profit work. Learn more at impactengines.ai