AI for Data Entry in Government: Keying, Re-keying and the Registers That Have to Be Right
An officer has a scanned application open on the left of the screen and the case system on the right. She reads a line, types it, reads the next, types that. At the end she opens the register in a different system and keys several of the same values again, because the register is what the public and the courts rely on and nothing feeds it. Then she notes the fee in a spreadsheet finance collects at month end.
Nothing in that hour is a judgement. All of it has to be right. This is about that layer, and what an AI system can honestly carry of it.
How a submission becomes a record
Intake is rarely one channel. A form arrives folded in the post. Another arrives as a scanned attachment, sometimes a photograph taken on a phone at an angle. A portal submission looks modern and still lands as a document rather than as data, the integration having been descoped. A telephone applicant talks while an officer types into a screen the applicant cannot see. And a shared inbox became an intake channel nobody designed, where a service request, a chased application and a complaint look identical until somebody opens them.
An agent is good at the first move: reading whatever arrived, classifying what service and stage it belongs to, and proposing a structured record with every value linked back to the place on the page it came from. That link matters more than the extraction: an officer who can see the source line checks in seconds.
Volumes are also uneven. A scheme opens, a deadline falls, a round of renewals arrives in one week, and peaks are where keying quality drops. The shortcuts taken under them stay in the record.
One fact, several systems, several field names
A date of application can exist in four places under four names, with four validations, one of which silently truncates. The case system wants a reference in its own format. The finance system wants a customer record that does not exist yet. The register wants a subset of the fields in a stricter form, and the spreadsheet wants whatever the team decided years ago.
Making those agree is not reading documents. It is mapping fields and respecting each system's rules, including the undocumented ones. An agent can hold that map, prepare the write for each destination, and show the officer one screen of what is about to go where. It can also compare the systems continuously and report where they have already drifted apart.
When the list itself carries legal weight
A wrong case record is a problem for one applicant. A wrong register is a different category of problem. Licences, permits, consents and registrations are relied on by people who were never party to the application: a buyer checking a property, an inspector acting on what the register says, another authority treating your list as the authoritative answer.
So the register deserves its own controls rather than being treated as one more downstream copy. Repeat entries are a particular hazard, because a second entry under a slightly different trading name is how one holder appears twice, as "Riverside Stores" and "Riverside Stores 2", with the lapsed entry the one the public finds. An agent can propose candidate matches with evidence attached, and create, amend or remove nothing on its own authority.
Addresses, and the property that was split into flats
Almost every public service record points at a property, and the authority keeps its own address list the record has to tie to. Residents do not type addresses the way the list holds them. Someone writes "Flat 2, 14 Mill Lane" where the list holds "14A Mill Lane". A house converted into flats still exists as one entry. A new build is occupied months before it appears anywhere. The same front door is recorded three ways across three services.
An agent can propose the matching entry, show the alternatives it rejected, and say plainly when there is no credible match rather than take the nearest. The useful output is the queue of properties needing adding or splitting, which belongs to the team that owns the list.
Whether the record is about a person, a property or an organisation
Keying names, dates of birth and identifiers fails in familiar ways: a name in a different order, a hyphen in one document and not the next, a day and month transposed. Behind that sits a question officers answer without noticing. Is this record about a person, about the property, or about an organisation that happens to have one director? Entitlement and what happens on a change of occupancy differ. An agent should have to state which it thinks it is reading, because when that is wrong the record is about the wrong thing.
Checking that what was attached is what was asked for
Applications arrive with evidence: proof of residence, income documents, photographs of a site, certificates, a plan. Before anyone assesses anything, somebody opens each attachment and confirms the right document is there, for the right party, covering the right period, and legible.
This is a completeness check and nothing more. It is not a judgement about whether a document is genuine, and should never be described as one. An agent can identify what each attachment appears to be, match it against the required list for that service, and draft the request for what is missing. Doing that at intake rather than at assessment is most of the value, because an application that waits a fortnight to be told it is incomplete has spent the clock twice.
Fees that have to reconcile with the finance ledger
A fee is taken in one system and has to land in another that was never designed to agree with it. Part payments, refunds on withdrawal, waivers applied as a local decision, a payment made against the wrong reference. Reconciliation then happens by eye and late. An agent can match transactions to cases continuously, group the unmatched by cause rather than by date, and propose the correcting entry. Posting anything that moves public money stays a human action with a named approver.
The officer who is the integration layer
In most authorities at least one system has no interface worth the name. A green screen reached through an emulator. A portal belonging to another body with no bulk upload, so records go in one at a time. An export reshaped and uploaded somewhere else every week.
Where an interface exists, use it. Where none does, an agent can prepare the file, check it against the destination's rules before anyone uploads it, and read the destination back to confirm what was saved. Reading back is the step teams skip and the one that catches silent rejections.
Returns that have to tie back to the record
Statutory and management returns are usually assembled by one person extracting from operational systems and retyping into a template, with adjustments carried forward because no system produces the required shape.
The requirement is not a plausible number. It is a number traceable to records. An agent can assemble the extract, reconcile each figure to the records behind it, and mark which fields are derived and which have no system source at all. Signing the return stays with a person who can explain it.
A correction has to travel as far as the error did
Correction is what separates public sector keying from everything else. A fee has been charged, a licence issued, a decision notified, another authority has consumed the data. You cannot quietly overwrite the field.
So every automated write needs the dull things recorded at the time: the source document, the region the value came from, the confidence, the destination field, and who approved it. With that, a correction is a traceable event with a before, an after and a reason. Without it, somebody months later reconstructs from memory why a value changed. And no agent overwrites a value a person has corrected by hand.
Where the officer's name has to be on it
Set the bar per field, by consequence rather than by one confidence number. Descriptive and contact fields can be written automatically once measured. Anything that creates or amends a register entry, moves money, changes a date determining entitlement or liability, or alters the identity a record attaches to waits for a person, permanently, not only in a first phase.
A record may be prepared automatically, but a decision about someone's entitlement, status or liability is made by an officer who can explain it to that person, to an elected member and to an auditor without referring to a model.
Where AI fails here
Matching defeats it where it matters most. Two residents with the same name at one address, a trading name that changed last year, a property mid-conversion. The model returns an answer with no sign it was a guess.
Confidence does not track legibility either: a value read from a creased, photographed form comes back as cleanly as one from a typed document.
Contradictions are not solvable by reading more carefully. When the form and the attached evidence disagree about a date, somebody has to ask the applicant, and a system that picks the likelier value has invented a fact.
Local convention blocks this more often than model quality does. If the rules for which system wins and when a register entry may be amended live in two long serving officers' heads, writing them down is most of the project.
And silent wrong writing is worse than a visible queue. A queue gets worked. A confidently wrong register entry is found by whoever relied on it.
One form, one register, one quarter
Pick one service, one intake channel and one register. Start where the agent writes nothing: a drift report between the case system and the register, an address match queue, an attachment completeness check that only drafts the letter asking for what is missing. Then let it prefill records an officer confirms before saving.
You do not need a pilot, because the evidence exists already. Take applications your team keyed and later corrected, run the agent over the original documents, and compare its output against both the first keying and the correction. Look hardest at the cases that were argued over.
Measure the baseline first: how many records are keyed twice, how long an application waits before anyone opens the attachments, and how many register corrections were raised last quarter. To talk through which part of this an agent should carry and which part it must not touch, write to hello@adoptaipartners.com.