from $2,400
The repetitive path between your systems
Built once and handed over with documentation a person on your team can read and change.
Most businesses run for more hours than they are staffed for.
The second shift is the one nobody staffs, and the one your customers still expect somebody to be covering.
We build the automations and assistants that carry the work through those hours. Then we stay on and operate them, because software nobody is responsible for stops being an asset within about a quarter.
- Work arrives from intake channels and from systems you already run.
- It is classified by deterministic rules first, and by a model only where rules cannot decide.
- Action is taken using a fixed, named set of tools and nothing else.
- Anything below the confidence threshold escalates to a named person rather than proceeding.
- Every action is written to an audit log you can read afterwards.
The same details get retyped into three systems before anyone touches the actual work.
The first hour of every morning goes into turning email into spreadsheet rows.
An assistant was set up once and nobody is certain what it can still reach.
Quoting runs through one person, and it stops when they are away.
What an unattended run looks like
A worked example rather than a recording, at the level of detail we would show you in week three. The fourth step is the one worth reading: the assistant gets it wrong, notices, and hands back. That behaviour is designed in, and it is most of what separates a system you can leave running from one you cannot.
0.00s
Input
Message arrives at the shared inbox with a PDF attached. Sender is not in the supplier table.
0.31s
Plan
Extract the invoice fields, match the supplier, check the purchase order, then queue for payment. Stop before payment, because that tool is not in scope.
1.44s
Tool call
extract_document returns a total of 4180.00 and a currency field that is empty.
1.52s
Caught and corrected
The first pass assumed the local currency. A guard rejected it, because currency was absent from the source rather than equal to the default. Assuming here would put a wrong number on an invoice, so the run stops instead of guessing.
1.55s
Escalate
Routed to the named owner for accounts payable with the extracted fields, the source PDF, and the one question that needs a human answer.
1.56s
Log
Written to the audit record: what it read, what it decided, which guard fired, and who it handed to.
Four shapes of work, priced from a floor
Scope moves the final number a long way. These are floors, published so you can tell before a call whether we are the right size for the problem.
from $6,000
An assistant with a defined job and a defined reach
It runs unsupervised because the limits were decided in advance and written down, not because we assume it behaves.
from $4,500
Your documents made answerable
Answers arrive with the source attached, so a person can check the claim instead of trusting it.
from $1,600 / month
We keep what we built accurate
A vendor changes a format, your process shifts, and a system that was right in March is wrong by August with nobody noticing until a customer does.
What we will not do
We will not sell you a model
The model is the cheap, interchangeable part. Nearly all of the work is the boundary around it: what it can reach, what it remembers, what it may do without asking, and what it costs to run.
We will not automate a process nobody can explain
If the current process lives only in one head, we write it down first. Automating an undocumented process makes the confusion faster.
We will not leave you unable to turn it off
Every system we ship has an off switch, a named owner, and a fallback the business can survive on. Software you cannot stop is not an asset.
We will not hide what it costs to run
You get the running cost before you commit, measured against your own volumes. Ongoing spend that surfaces after signing is a sales tactic.
Answered before you have to ask
What happens to our data?
Systems run in your accounts under your keys wherever the vendor allows it. We do not train anything on your data, and we do not retain copies after handover beyond what you ask us to keep for support. Where a third-party model provider sits in the path, you are told which one and what reaches it before anything is connected.
What happens when it breaks at 2am?
It fails to a documented fallback, usually the manual process it replaced, and it alerts a named person rather than a dashboard. On a retainer that person can be us. Without one it is somebody on your side, and part of handover is making sure they know what to do.
Are you going to leave us with something we cannot maintain?
This is the right question to ask any contractor, and the reason handover is a named week rather than a final email. You get the system, the documentation, the reach list, the measured running cost, and the off switch. The test we hold ourselves to is that a competent person on your team could pick it up from the written record alone.
Do we have to replace our existing stack?
No, and we will push back if you suggest it. The work sits on top of what you already run. A migration and an automation project at the same time is two risks multiplied rather than added.
How do we know it is actually better?
Week three runs the automation in parallel with the manual process and compares the two on your own inputs. If the comparison says it is wrong often enough to matter, you hear that from us, and the decision to continue is made on evidence rather than on how much has been spent so far.