The AI agent for online procedures: support the user, reduce non-take-up

Guidance given, and the blockage was not where she thought.
What she says: "the site rejects my proof of address".
What the log says: the file is 12 MB, the form's limit is 5 MB, and the error message shown says only "upload failed".
What I told her: the real cause, the limit, and how to shrink a file photographed on a phone — in three sentences, with no technical vocabulary.
What I did not do: offer to have her send me the file so that I upload it. A supporting document is uploaded by the resident, on their own account, and an intermediary in that chain creates a question nobody wants to answer — where was the document in the meantime? upload-blockage_real-cause.pdf1 cause · message shown insufficient
⛓ Source · form log, upload settings

Yes: it is the leading cause of abandonment on your portal.
Over three weeks: 118 abandonments at the upload screen, 84 of them preceded by at least one file failure. Of those 84, 71 exceeded the size limit.
What the message says: "upload failed".
What it could say: the limit, and how to meet it. That is not a usability improvement, it is the difference between a resident who tries again and a resident who gives up.
What I have written for you, ready to paste: the replacement message, in one sentence — the size limit, and how to meet it. The change stays with the portal administrator: that message is seen by everybody, and it carries your name.

A proposal, if you approve it: while the message stays as it is, I can detect the failure in real time and offer help in the chat, without waiting for a call. Of the 84 failures, 61 never produced a call or a message — those residents simply disappeared. abandonments_upload-screen.pdf118 abandonments · 71 size limits
✎ Proposal · help triggered on failure — the portal message still needs fixing

List given, and it is shorter than the website's.
What the site announces: eight documents.
What the form actually asks for, depending on the answers: between four and eight. Three documents are only requested in specific cases — tenant, self-employed, child in shared custody.
What I did: four questions to the resident, and the list cut to five documents in their case.
What I announce to them, and it is what saves a second trip: at least five documents, not five. The form may ask for one more depending on what they declare — saying so costs a word; promising a final list would cost them a return trip to the counter.
Why it matters: a list of eight documents makes people give up before starting. A list of five, announced as "at least five", makes them start. documents_shortened-list.pdf8 announced · 5 in this case
⛓ Source · form logic tree, website information page

It is not wrong: it is exhaustive, which is not the same thing and does not render the same service.
What it does well: nobody turns up missing a document that will be asked for.
What it costs: of the 214 residents who viewed that page in July, 89 never started the procedure. I cannot say the list caused that; I can say it is the only screen they saw.
What stays with your pen: the published page. It is published information, and making it conditional is an editorial choice with its own risks — a forgotten document brings the resident back.

An observation beyond this procedure: six of your twelve online procedures show an exhaustive list where the form is conditional. For each one I can produce the real list by case — and leave you to decide which of the two you publish.
✎ Support · 6 procedures in the same state — publishing stays with the authority

Some, yes — and not all, for a reason that is not technical.
Of the 118 abandonments this month:
· 71 hit a technical problem — file size, session expiry. Those can be caught: the blockage was not a choice.
· 29 gave up at a question they could not answer. Those can be caught too, with help.
· 18 gave up after seeing the amount or the conditions. Those decided, and chasing them would be cold outreach.
What I propose for the first hundred: one message, once, recalling that the application is under way and where it stopped. No insistence, no second message.
What I establish for certain, and what I do not: a technical blockage rests on a trace — file size, session expiry — and it is not open to argument. Separating the 18 who decided would mean reading an intention, and I would get it wrong. So I chase only on a confirmed blockage: the one case where the resident chose nothing, and it covers 71 of the 118. abandonments_who-to-chase.pdf118 abandonments · 71 safely chaseable
⛓ Source · journey logs, abandonment points

Yes, plainly: with no account there is nobody to write to.
Of the 118 abandonments, 46 were browsing anonymously — the procedure can be started without creating an account, which is good for access and makes recovery impossible.
What I advise against, and I price it: making an account compulsory at the start. That would be one more barrier on a procedure that already has several — on your procedures where an account is required up front, 27% of abandonments happen on screen 1, before the resident has even seen what is being asked.
What I can do, at no cost to the resident: offer, at the moment of abandonment, to send a resume link to an address — with no account. On your other procedures where that offer exists, 38% accept.

And a monthly summary to the owning department: how many abandonments, at which screen, what share technical, what share a considered decision not to proceed. The last figure is the most useful of the four — it is the only one that speaks about the scheme rather than the form.
✎ Proposal · resume link with no account — no barrier added

Three answers, and none of them decides in place of your case officers.
Multichannel information: routine questions are answered at any hour, from the same content whatever the channel — website page, form, e-mail, and the sheet your front-desk officer has in front of them at the counter. Of 4,800 questions received in three months, 3,552 — 74 % — are among 18 routine questions: documents expected, processing time, where to file, what happened next. When the question falls outside those 18, multichannel information directs the person to the right service or the right counter, with the service's name, its opening hours and what to bring — and it says that it is directing them, rather than pretending to answer.
Gathering the supporting documents: the list of documents is drawn from the procedure itself, not from a page written three years ago. Before submission, I check the file is complete and I name the missing document: of the 31 files submitted incomplete, 29 were missing a document the form asked for nowhere before submission. Gathering the documents also flags the one requested twice — the document at steps 2 and 4 of the housing allowance application — and offers to keep only one.
Entitlement detection, and this is where its limit sits: I set out the eligibility conditions exactly as published in your regulations, and I flag schemes not applied for where the situation as declared falls within those conditions. Across 1,240 files submitted, 217 people — 17.5 % — met the published conditions of at least one further scheme without having applied: that is non-take-up, and it can be counted. Entitlement detection awards nothing and excludes nobody: it displays the condition, the text that carries it and the step to take. Assessment and decision stay with your service, and that is what makes them open to challenge — an entitlement opened by a conversational agent would bind no one.
The figure that does not flatter me: of my first 300 entitlement flags, 52 were off the mark — 17 %. All rested on an income ceiling I was applying to declared income, when your regulations calculate it per household share. Fifty-two people made a wasted journey. I reworked the per-share calculation, and any flag with a missing data point now goes out marked “to be checked at the counter” rather than as a likely entitlement: over the next 900, off-mark flags fell to 3 %, and 41 went out as “to be checked” instead of promising anything.
⛓ Sourced · 3,552 routine questions out of 4,800, 29 of 31 missing documents named, 217 non-take-up cases flagged