Missing-Docs Agent helps a practice see what a client has sent, what is still missing, and what to ask for next. We have built the local prototype. The post-fix evaluation is complete. Dashboard captures and the demo recording are still pending.
A client sends some receipts, a few invoices and most of their bank statements. Someone still has to match them against the job, spot the gaps and write another email.
We separate the reading from the rules: AI extracts the fields, a deterministic checklist finds missing or unclear items, and a human checks the draft.
This is a demonstration of the workflow, not evidence of saved staff hours or improved client response rates. Those need to be measured in a pilot.
The local pipeline reads a synthetic client folder. Providers classify documents and extract type, date, total and counterparty. The live LLM route uses Gemini, a backup Gemini key and Groq for text fallback.
Boundary: this is folder-based input, not a deployed client upload portal. The offline baseline is regex, not AI.
Rules for sole traders, limited-company year ends and VAT packs compare received documents with the client profile. Results distinguish complete, missing and unclear, including missing statement months.
Checks: separate validation flags wrong year, wrong client and non-business documents. Rule sources cite public gov.uk guidance; an accountant must confirm the checklist.
Deterministic templates turn missing or unclear items into a polite chase email. The Streamlit dashboard brings the client checklist and draft into one review screen.
Boundary: chase copy is template-based, not LLM-written. The dashboard uses synthetic records.
A named reviewer approves the draft. The prototype writes a local outbox entry and approval audit trail, then calculates reminder dates at day 3, 7 and 14.
Boundary: no email is sent. Reminder logic can re-check remaining items; a live scheduler, mail integration and database-backed audit trail are not verified.
Integration status: Companies House lookup code can supply accounting dates when configured. It has not been confirmed live in this review. The n8n workflow is a scaffold and the database connection still needs verification. These are not advertised as deployed services.
Intake → extraction → missing-item flags → chase draft → human approval. The final capture will use labelled synthetic data throughout.
This space is ready for the checked 90-second demo. Approval in that demo records a local outbox entry, not an email send.
Actual dashboard capture showing complete, missing and unclear items. No simulated UI is presented as product evidence.
Actual capture of the chase draft, named reviewer and local approval record. Synthetic client details only.
The post-fix evaluation covered all 53 labelled synthetic documents on 10 October 2026. These are test-set results, not a real-client study or a guarantee of production accuracy.
Document type / date / total / counterparty.
All 53 documents were evaluated, including 8 images. The OCR-noise text category had all fields correct on 4 of 7 documents (57%); the other six categories scored 100%.
evalset, prompt v2, commit 37f198a. Primary Gemini answered all 53 documents.
The current harness targets 53 labelled synthetic documents: clean text, British date formats, OCR damage, photo-like images, wrong-year and wrong-client documents, and junk.
The evaluation covered all 53 labelled synthetic documents. This is not a real-client study or a guarantee of production accuracy.
Type and date must match the labelled answer. Totals must match within half a penny. Counterparty matching normalises case, punctuation and certain company suffixes.
The harness excludes fields with no labelled value. Skips and request failures must be reported alongside scores, not hidden behind a percentage.
A possible pilot shape, to agree with your practice. No rollout, pricing or real-data processing is committed by this page.
Hosting, access, retention, deletion, provider use and international transfers must be agreed first. No GDPR compliance or ICO registration claim is made here.
Tell us which documents your team spends time chasing. We can discuss a small pilot, what it needs to prove, and where human review stays in the loop.
Email to arrange a time. A booking calendar is not connected yet.
Mojtaba Javeed Siddiqui
mojtaba.javeed@gmail.com
WhatsApp · +91 9699651506
Hyderabad, India