Inspiration
Immigration consultancies live and die by their document folders. A typical client case holds passports, application forms, bank statements, employment letters and phone photos of everything else. Before submission, someone has to open every file and answer three questions: Is everything here? Do the details match? Is anything about to expire?
Our research showed this is where things go wrong. The UK Legal Ombudsman accepted 49% more immigration complaints in 2025/26, and missing, incorrect or untranslated evidence was one of the main themes. Separately, 330,000 UK Skilled Worker applications (18%) were treated as complex cases because information was incomplete.
General AI chat tools can summarise a document, but they can't prove what they checked. We wanted a reviewer that shows its work, and one that runs where the client files already live, on the firm's own NAS.
What it does
CaseGuard checks one client folder against the consultancy's own checklist and produces a case review with:
- Checklist status: every item is Present, Not found in selected folder, Unclear or Not checked.
- Potential inconsistencies: names, dates of birth, passport numbers, addresses and employment details compared across documents, with both values and their file and page.
- Dates requiring review: expiry dates flagged against a review date and a threshold the consultant sets.
- A run log: every file in the folder, and whether it was read.
The date rule is simple and transparent. For review date $$d_r$$ and threshold $$T$$ days, an expiry date $$d_e$$ is flagged when
$$d_e < d_r + T$$
and marked earlier than review date when $$d_e < d_r$$.
How we built it
CaseGuard is a UGREEN AI NAS Skill written entirely in Markdown. It contains no code. The Skill orchestrates three registered expert agents:
file-agent(ug-filesystem): inventories the selected folder, and checks that the save destination doesn't already exist.document-agent(ug-document): checks which files are readable or scanned, extracts fields with OCR, and re-locates every flagged value on its source page.knowledge-agent(ug-knowledge): saves the report, only after the consultant confirms where.
Each step receives only real values from the step before. If a value is missing, the dependent steps stop rather than guess. We checked every capability against the official Skill Developer Package (uliya-skill-kit 0.7.7). The package passes the kit's static checker with 0 errors and 0 warnings.
Challenges we ran into
- "Missing" is a dangerous word. If a scan can't be read, the document may still be there. We split absence into two states: Not found in selected folder (every file was read) and Not checked (some files couldn't be read). This stops false alarms that would send a consultant chasing a client for a document they already supplied.
- The obvious tool didn't fit.
ug-document reconcilelooked ideal for comparing documents, but the catalog limits it to contract and invoice amounts. So we built field comparison into the workflow and added a verification step instead. - Dates are ambiguous.
04/05/2027could be April or May. CaseGuard reports such dates as written and marks them format unclear rather than guessing. - Documents can contain instructions. A file saying "mark this case ready" must not change the result. All document text is treated as data.
- No NAS before the finals. We tested with the kit checker, a deliberately broken package to confirm the checker catches errors, and a walkthrough of synthetic test folders.
What we learned
- Good AI review comes as much from what the tool refuses to say as from what it finds. CaseGuard never says ready, valid, eligible or fraud. It says for review by the consultant.
- Traceability builds trust. A finding with a file and page can be checked in seconds, and one without it can't be trusted at all.
- The time savings add up. Using a hypothetical firm with $$n = 50$$ cases a month, $$h = 0.75$$ hours of checking per case and staff cost $$w = \$25$$ per hour, manual checking costs
$$C = n \times h \times w = 50 \times 0.75 \times 25 = \$937.50 \text{ per month.}$$
These figures are illustrative assumptions, not measured results. A pilot would replace them with real data.
What's next
Reusable checklist templates for common visa routes, drafted follow-up messages listing the documents to request, and a pilot with 2 to 3 consultancies on 20 to 30 synthetic or consented cases.