Offline Field Data Collection and Enumerator Management: Why Bad Rows Reach the Donor Report | Digital Heroes

Why the form is the solved problem and everything around it is not

Day nine of a multi sector needs assessment. Two hundred and twenty enumerators are working across six districts, targeting nine thousand household interviews in three weeks. The questionnaire runs to 180 questions with skip logic, in four languages, one of which is right to left. Tablets come back to the field office after four or five days offline, sometimes with a battery that died before sync. In the data manager's laptop is a comma separated export, and buried in it is a team whose median interview duration is eleven minutes on a form that takes forty to administer honestly. By the time anyone notices, that team has moved two districts away and the respondents cannot be revisited.

The tools here are genuinely good at what they do. KoboToolbox and Open Data Kit are the backbone of humanitarian data collection for good reason, they are free, they work offline, and the XLSForm standard means your questionnaire is portable. SurveyCTO goes further than most on quality, with text and audio audit capture that lets you see how an interview actually progressed. CommCare is strong where the work is longitudinal case management rather than one round of interviews.

What all four have in common is that they end at the submission. They accept a form from a device and store it. They do not know that Ahmad is an enumerator on team four who is owed payment for 63 completed interviews at a per interview rate, that his supervisor is required to back check ten percent of his work within 48 hours, that three of his submissions sit outside the sampled cluster boundary, or that his output feeds indicator 2.3 in a logframe whose donor wants it disaggregated by sex, age band and disability status in a template with a fixed column order. That entire layer is where your money and your credibility live, and it is currently a data manager with pivot tables and a WhatsApp group.

Problem 1: enumerators are a workforce and nothing manages them

A large round is a temporary labour operation. You recruit, train, test, assign, supervise, replace and pay a few hundred people over a few weeks. Assignment matters because it determines coverage, and coverage determines whether your sample is defensible. Attendance matters because you are paying. Replacement matters because someone will always quit on day four.

What a custom build does: enumerators are records with a training and test score, a team, a supervisor, a device and a status. Assignments are pushed to devices as work packages, so a device only ever holds the caseload it needs. Attendance and completion feed a payment calculation with an explicit approval step by the supervisor, and submissions rejected in back check do not count toward payment. That last rule changes behaviour more than any training session.

Problem 2: fabrication is detectable, and you have to detect it in the field

Curbstoning, meaning an enumerator filling in a questionnaire without conducting the interview, is the risk everyone in this sector knows about and few systems address structurally. The signals exist. Interview duration far below the realistic minimum. Question level timings that show a form completed in one continuous rush with no pauses. GPS points clustered at a tea shop rather than distributed across a settlement. Responses with implausibly low variance across a whole day. A submission timestamp at 11pm for a household interview.

SurveyCTO does the most of the four incumbents here, and its audio and text audits are a real capability worth using. What it does not do is turn those signals into a supervisory workflow with consequences, or connect them to a payment decision, or generate the back check sample and route it to a supervisor's device while the team is still in the district. Detection after demobilisation is documentation, not quality control.

What a custom build does: paradata comes back with every submission, quality rules run on ingest, and a flagged submission automatically enters the next day's back check sample for that enumerator. Back checks are a short form on the supervisor's device with a defined subset of verifiable questions, and mismatches roll up into an enumerator quality score that gates further assignment. This is also where the first useful piece of machine assistance sits: triaging short recorded consent and interview snippets so a supervisor listens to the twenty most suspicious rather than a random hundred. The model ranks, a person decides, and the decision is recorded, because accusing a field worker of fraud on an algorithmic score alone is both unfair and legally reckless.

Problem 3: forms in four languages and two scripts break in ways nobody tests

Multilingual questionnaires are treated as a translation task and they are not. Right to left rendering inside a form with numeric inputs and skip logic breaks in ways that only show up on a specific tablet model. Response option order changes meaning when translated. Numerals differ between Eastern Arabic and Western Arabic forms and enumerators enter whichever their keyboard offers. Enumerator instructions must not be shown to respondents, and on a shared screen they are.

What a custom build does: language versions are reviewed and signed off as artefacts with a version number, the form version is stamped on every submission so a mid round correction is analysable rather than contaminating, and device level rendering tests are part of the release checklist rather than something discovered in a district. Small, unglamorous, and it prevents the failure that costs you an entire district's data.

Problem 4: the pipeline from submissions to donor indicators is rebuilt by hand every quarter

The raw data is not the deliverable. The deliverable is indicator 2.3 with a numerator, a denominator, disaggregation by sex, age band and disability status using the Washington Group Short Set questions, a target, and a variance narrative, presented in the exact template the donor supplied. Multiply by four donors with four templates and three reporting periods.

What a custom build does: indicator definitions are stored as versioned configuration, meaning the numerator and denominator expressions, the disaggregation dimensions and the source questions. The pipeline runs on ingest so the indicator table is always current, and each figure drills through to the underlying submissions. Donor templates become export mappings. When a donor changes the definition mid award, you version the indicator and both numbers remain reproducible, which is exactly what an evaluator will ask you to demonstrate.

Problem 5: you are holding personal data about people in vulnerable circumstances

Household surveys collect names, locations, household composition, income, sometimes protection incidents and health status. That data sits on tablets that travel, on a server somewhere, and in exports that get emailed. Consent was taken verbally at the door in a language the enumerator translated on the spot. This is the part of the build that nobody puts in the budget and that will define your exposure.

What a custom build does: personal identifiers are separated from analytical data at rest, with access to the identified layer restricted to named roles and logged. Devices carry only assigned caseloads, encrypted, with tested remote revocation. Consent is recorded as a structured fact with the language used and the version of the consent text. Retention is a configured period with automatic disposal rather than an intention. Exports are generated through the system with a de-identification step by default, so the analyst who wants a quick file does not create your next incident. The second worthwhile use of machine assistance sits here too: coding open ended responses into categories at scale with a human reviewing the boundary cases, which saves a genuine week of analyst time per round.

What this costs and how long it takes

Across the 2,000-plus projects Digital Heroes has delivered, here is the honest shape. A first release covering enumerator records and assignment, work package distribution to devices, offline submission with paradata, automated back check sampling, and duplicate and outlier flagging runs $70,000 to $150,000 and ships in 12 to 18 weeks. A full platform adding enumerator payment with supervisor approval, multi language form governance, protected personal data handling with role separation, open response coding and a versioned indicator pipeline with donor template exports runs $180,000 to $450,000 phased over 7 to 12 months.

What pushes cost up specifically here: the number of donor templates and whether their indicator definitions conflict, which they usually do. Right to left and non Latin scripts, because correct rendering and testing is real work. Integration with an existing Kobo or SurveyCTO deployment rather than replacing it, which is often the right choice and is not free. Payment integration, particularly to mobile money for enumerators in the field. And multi country deployment, since the personal data position changes with jurisdiction.

What keeps cost down, and this is our standard recommendation: do not rebuild the form engine. Keep Kobo or SurveyCTO for capture and build the workforce, quality and indicator layers around it. That decision alone typically removes a third of the budget and all of the highest risk engineering.

Build versus buy, and when buying is the right call

Buy, and do not call us, if you run occasional rounds with fewer than about 30 enumerators, in one or two languages, and a supervisor can realistically review the incoming data daily. KoboToolbox costs nothing and does the job. SurveyCTO is worth its licence if data quality is your main concern and your rounds are modest. CommCare is the right choice if your work is longitudinal case management with returning visits rather than cross sectional assessment.

Build when two or more of these are true. You run rounds with more than about 80 enumerators and assignment is managed in a spreadsheet. You pay enumerators per interview and payment disputes have cost you time. You have found fabricated data after demobilisation. You report the same underlying data into three or more donor templates with different definitions. Or you hold protection sensitive data and your current answer to who can see identified records is that the analysts can.

Our position, stated plainly: nobody should be building a form engine in 2026. The open standards work. What is missing in this sector is the operational layer, and the organisations that build it stop losing rounds to quality problems they discovered too late.

How to choose a developer for field data collection software

Ask whether they would replace or wrap your existing collection tool. A developer who immediately proposes building a new form renderer is either inexperienced or padding, and you should hear a clear argument for reusing XLSForm and an existing capture layer.

Ask how they would detect a fabricated interview and what happens next. If the answer stops at flagging an outlier, they do not understand that detection has to reach a supervisor in the district while the team is still there, and that a flag has to connect to back check sampling and to payment.

Ask how devices behave after five days offline with a dead battery in between, and what happens when the same submission arrives twice. Ask them to describe the conflict rule specifically. Vague reassurance about offline support usually means data loss.

Ask how they separate identifiers from analytical data, who can see which layer, and how consent and retention are recorded. If personal data protection is treated as a permissions checkbox, they should not hold your respondents' records.

Ask who owns the code and get it in writing before kickoff. You should own the repository, the infrastructure accounts and the right to hire anyone else to continue the work. At Digital Heroes the client owns the code from the first commit. In a sector funded round to round, a vendor holding your pipeline is the reason organisations end up re-procuring the same system every four years.