AI solutions
AI automation
Routine that people do moves to the model: sorting requests, drafting replies, checking data before it goes out. People stay where decisions are made.
- Reply to your brief
- within 24 hours
- First call
- 30 minutes, no commitment
- First working version
- in 2 weeks
Process
What happens, and when
The work is split into stages, and each one has a named result. You see it yourself rather than reading about it in a report.
Days 1–3
Measuring manual work
We sit next to whoever does the work by hand and write it down as it is: how many times a day, how many minutes, where they have to jump into another system, where they actually decide something. That gives the “what it costs today” number — without it there is nothing to prove improvement against a month later.
You get — The “before” figure to compare against
Days 4–6
Splitting the routine
The routine splits three ways: what the model does alone, what it prepares for confirmation, and what it never touches — exceptions, money, telling a customer no. We take one routine end to end rather than half of five: a half leaves the person in both systems and saves nothing.
You get — The line between model work and human sign-off
Week 2
Plumbing and the queue
Most of the work here is not the model but the plumbing: where the task comes from, where the result goes, what happens on a duplicate and on a re-run. A model without a queue and without protection against double processing falls apart in live traffic on day two.
You get — A queue safe against double processing
Before launch
A week in parallel
For a week the model runs alongside the people: same work, but the output is compared rather than sent on. Where it matches, we switch over; where it diverges, you can see exactly which cases before a customer does.
You get — A divergence report from live traffic
After launch
The count after launch
We count the same number as at the start, plus the share of tasks handed back to a person. A rising share of hand-backs is not a fault but a signal that the flow has changed. The next routine is taken on once the first has paid for itself, not before.
You get — The same figure again and the hand-back share
What the work covers
Measuring the manual work
how often it happens, how many minutes one case takes, how many jumps between systems it costs — the “before” number the result is compared against a month later
Splitting the routine
the line between what the model finishes by itself and what a person signs off — with payments, refusals to a customer and anything irreversible kept on a separate list
Plumbing and the queue
where the task arrives from and where the result goes, protection against double processing, a retry after a failure; this is also where it is written down which fields reach the provider and which stay with you
The log and the rollback
every action is recorded along with what the model saw before it, finished work can be undone, and an uncertain result goes to a person before anything leaves the building
A week in parallel
the model runs on live traffic beside the people, its output goes into a divergence report, and only the cases with no divergence are switched over
The count after launch
the same measurement as in week one, how much of the flow comes back for a person to finish, and a list of the cases the model gets wrong most often
What we need from you
- Whoever closes this routine by hand today, and a couple of hours of their time in the first week
- Agreement to close one process completely instead of touching five halfway
- Access to the systems the task travels between, and a service account allowed to write to them, not only read
Projects in this area
E-commerce
Onvela
A cosmetics retailer whose first line of support is an AI bot. It does not recite the FAQ: it finds the order, checks delivery and payment, and answers from data.
18 seconds
to a first useful answer, where the peak-hour wait was 27 minutes
NestJS · PostgreSQL · Qdrant
Fintech
Tallim
Matching payments to invoices inside a payments company’s finance team: strict rules, similarity and a model each take their turn, and a person confirms what stays unclear.
31,000 payments
a month go through reconciliation; 1.9% of them reach a person
Python · PostgreSQL · OpenAI API
Construction
Prokta
Tender documents read for a general contractor’s commercial department: an archive of files goes in, a report of terms comes out, every line of it with a page reference.
91%
of the control sample’s critical terms the system found on its own
Python · Qdrant · OpenAI API
What people usually ask
Contact
Send a description of the task
A reply with the scope, the timeline and a budget estimate comes within 24 hours.
The first call is 30 minutes, with no commitment on your side.
01
You describe the task
Five questions in the form, or a plain email — whichever suits you.
02
We answer within a day
With the scope, the timeline and a budget estimate, based on what you told us.
03
We talk for 30 minutes
To clear up whatever is unclear. It commits you to nothing.