01 The state of things
Before there was a company called Marwin, there was one operation: a multi-building commercial property portfolio in London, offices, storefronts and residential units, run the way most operations of that age are run. Paper first, email second, one person holding the rest together in her head.
That person was the office manager. Lease terms, rent schedules, maintenance requests, tenant queries, owner reporting, all of it passed through her at some point. The archive behind her went back 70 years. Some of it was typed. Some of it was handwritten, filed in boxes, never indexed past the label on the outside.
Nothing was broken in the way that makes headlines. Rent got collected. Leases got renewed. But every answer took a search, and every search took her time. The founder built Marwin's first system to fix that, before Marwin was a business anyone could hire.
02 Why paper, not software, was the actual problem
The instinct with an old, paper-heavy operation is to buy a property management platform and migrate everything in. That was tried, in pieces, over the years. It never stuck, because migration assumes the data is already structured. Here it wasn't. A lease from 1994 and a lease from 2019 don't share a template, a filing convention, or in some cases a language.
The real problem wasn't the absence of software. It was that 70 years of documents existed in a form only a human could interpret, and the volume had outgrown the one human doing the interpreting. Any system that didn't start by solving that would just be a nicer-looking dead end.
03 What got built: the digitization pipeline
The pipeline starts with classification, not extraction. Every incoming and archived document gets sorted into type first, lease, invoice, correspondence, insurance, statutory notice, before anything tries to read its contents. Sorting first meant each document type could get handling suited to it, instead of one generic OCR pass forcing every layout through the same net.
Each document type then runs through OCR tuned to that type's layout and vocabulary. A rent statement and a 1970s solicitor's letter don't fail in the same ways, so they aren't handled by the same pass.
Every extraction carries a confidence score. Below a set threshold, the document routes to a human review queue instead of getting filed on a guess. That queue is small by design. It catches the genuinely hard cases, not the easy ones dressed up as hard because the system wasn't tuned.
The output is searchable retrieval with the original page attached to every result. Nobody has to trust a summary. They see the source page a search pulled the answer from, every time.
04 What got built: the rest of the system
Tenant communications draft themselves from the underlying data, lease terms, payment status, open maintenance items, but nothing goes out without a human reading it first. Drafting the first version is the automation. Approving it stays a person's job.
Rent and arrears reconciliation runs against the lease and payment records automatically, flagging mismatches for someone to look at rather than letting them sit unnoticed in a spreadsheet until quarter-end.
An owner dashboard sits on top of all of it: occupancy, arrears, open items, in one place instead of a phone call away.
05 Results
Manual back-office work is down more than 40%. Occupancy moved from 60% to 84%. Rental income is up as much as 30%. The 70-year paper archive is searchable now, with the source page attached to every result.
None of that came from replacing the office manager. It came from taking the parts of her job that were pure lookup and repetition off her plate, so the parts that needed judgment got her full attention.
06 What it doesn't do
This is worth being straight about. The system does not replace judgment on lease negotiations, disputes, or anything with legal exposure. It drafts; a person approves. It does not guess on documents it can't read confidently, low-confidence items go to a human, full stop.
Some of the oldest handwriting in the archive, 1960s and earlier, still defeats OCR often enough that those pages go straight to manual review rather than getting a machine-generated guess filed as fact. That's a deliberate limit, not an oversight. A wrong answer filed with confidence is worse than a slow right answer.
07 Why this is the first case study, not a pitch deck line
This system ran for a real operation before Marwin took on a single client. The offer that came out of it, Operations Scan, Automation Sprint, Operations Partner, is the same shape of work: find where paper and manual lookup are eating a team's time, build the narrowest thing that fixes it, keep a human in the loop wherever judgment is required.
If your back office looks like this one did, worth a short conversation before anything gets scoped.
See what a scan finds in your operation
20 minutes, no deck. Book a slot and bring one process that's still running on paper or someone's memory.
Book an Operations Scan