Documents, requests, receipts, forms. They turn up without warning and belong to nobody in particular. They end up with you, because you are the one fixed point in the business. We build the systems that catch what arrives, so it stops routing through you.
Show Me What This Looks LikeNobody designed it this way. A file arrives. A client asks for something. A receipt turns up in a text message. Whoever notices first deals with it, and the person who notices most is you.
This never gets fixed because it cannot be handed to software the way your reporting can. Work that runs on a schedule can be automated end to end, because you know its shape before it arrives. This work you do not.
So we build it the other way round. Everything sorts, extracts and proposes. Nothing files, sends or approves. A person clears the queue in a couple of minutes, and only then does anything actually happen.
That sounds like extra work. It is the opposite, and here is why it matters more than it looks.
Tested on ten thousand real receipts, the best models scored around three-quarters correct overall, and closer to half on the individual line items. More telling than the score is how they failed. When they could not read a field, they made up a plausible value instead of leaving it blank. In hundreds of cases they quietly adjusted line item prices so the total would add up.
The document balances. It is wrong. Nobody looks again.
That is the shape of every failure in this kind of work. Ask a search tool for every document about something and it returns some of them, presented as all of them. A sync fires when a record is created but not when it changes. A file lands under the wrong client and is not lost, it is invisible, until a dispute.
A larger company catches these, because somebody checks. At fifteen people nobody does. So the checking step has to be built in from the start, which is what makes these systems safe to actually rely on.
Accuracy figures from an independent university study of 10,656 real receipts, 2026. Vendors in this category advertise 99% accuracy, measured on their own test sets with their own definition of a correct field.
There is a notetaker in your calendar that nobody decided to put there.
Someone installed it on a personal plan, it joins everything on the schedule, and it has been sitting in client calls for months. Ask where the transcripts live, who can search them, or whether the client on the other end agreed, and most firms cannot answer.
We turn off auto-join, so recording becomes something a person chooses rather than the default. Then notice in the invite, and a line to say out loud in the first minute. If anyone declines, the recording does not happen and somebody takes notes. Then a deletion clock, client transcripts kept apart from internal ones, and a name against who owns all of it.
This one usually costs less than nothing, because it collapses three or four personal subscriptions into one.
Files arrive as scan001.pdf and get dragged somewhere by whoever saw them first.
We set up one place things land. Everything arriving gets read, sorted into a fixed list of document types, and given a proper name and a destination.
Then it stops. The system stages the change and waits. Somebody clears the queue in a two-minute pass each morning, and only then does the file move.
Two rules make it hold. The sorter picks from a closed list or says it does not know, so it never invents a category. And anything it is unsure about goes to a person by default, along with anything from a company it does not recognise. No best guesses.
We also build a monthly spot check on a handful of filed documents. Without it, the review queue turns into a rubber stamp within a quarter, and the control exists only on paper.
Four things are sitting out there with somebody else and you remember three.
It is not just unsigned agreements. It is the onboarding form, the insurance certificate, the tax form, the question you asked about an invoice. A firm your size has ten to thirty of these open at once. Nobody tracks them consistently.
We build one register, filled automatically from the systems that already know. It shows what is open, how old it is, and whose it is. The assistant reads the shared inbox to notice when something opens or closes, then drafts the chase in your own voice.
It does not send. Chasing somebody who already did the thing is worse than not chasing at all, so a person hits send.
There is an address three people half-watch and everyone assumes somebody else has it.
Most of the sorting here should never involve AI, and saying so is the point. Mail from a known client goes to that client's owner immediately. A reply on an existing thread stays with whoever owns the thread. Those are rules, not judgement.
On top sits a clock that cannot be talked out of anything. If a message has not been actioned within a set window, it escalates to a named person, no matter how it was categorised. That is the safety net.
Only what is genuinely ambiguous goes to the assistant. It sorts into a handful of categories, plus an explicit "not sure" that always goes to a human. It summarises long threads for whoever picks them up, and flags any deadline mentioned in the body.
It never replies. Not even an acknowledgement. An address you publish is an address anyone can write instructions into, and a system that only reads cannot be talked into anything.
Three people took clients to dinner in March and the receipts turned up in June, as photos.
We move capture to the moment of spend, match against the card feed automatically, and nudge whoever spent when a charge has no receipt after a few days.
The piece worth building beyond what the standard tools do is the business purpose line, drafted from what was already in the calendar. It is the field people skip and the field anyone reviewing the books actually wants.
Then the checks, which are not optional given how these models fail. A person confirms anything above a threshold, anything in a foreign currency, and anything billed back to a client. That last one has no exceptions and no minimum. A date read a month wrong does not surface until long after you could have fixed it.
Most of what we build is specific to how you run.
None of this may be your bottleneck. That is normal. Every engagement starts by finding where the week disappears, then building against the two or three places that keep costing you.
What arrives unannounced, who it lands on, and what it costs to handle. No tooling decisions yet.
The highest-leverage two from the map, working in your stack, inside the engagement.
Documentation, templates, and the written rules for what the system is allowed to do on its own. Your team runs it without us.
Async by default. Written updates, recorded walkthroughs, working systems you can inspect. We take a small number of clients at a time, which is why the work goes deep rather than wide.
A week of data entry eliminated. CRM auto-enrichment and contact scoring from records that fill themselves from public signals.
A week recovered, per person. A six-step automated intake workflow built for a high-volume intake operation.
Both are intake and records work, the same shape as everything above. We do not have a published result yet from a document or expense build. When we do, it goes here with a real number.
Tell us what arrived unannounced last week and who ended up dealing with it. We will send back a short breakdown of what could be handled before it reaches you, what should stay on your desk, and what we would not automate at all.
Send Me The Breakdown