The home page is the short version. This is the whole argument: where the manual hours hide, why the last attempt to remove them failed, where these projects stall, what we would refuse to build, and what we have already shipped.
We find the same four shapes in almost every operation we open up. Each one looks like a job for a person because the input is messy, and each one stopped needing a person the moment machines could read a sentence.
Invoices, delivery notes, claims, contracts, customs paperwork.
Someone opens the PDF, reads it, and types eleven fields into a system that will not accept the twelfth.
The document is read, the fields are extracted and checked against what you already hold, and only genuine mismatches reach a person.
Shared mailboxes where every message has to be read before it can go anywhere.
Two people spend their mornings triaging, because keyword rules cannot tell an angry customer from a routine reorder.
Each message is classified on what it actually says. Where the answer is known it comes back drafted. Where it is not, it escalates with the context attached.
The ERP, the CRM, and the spreadsheet one person maintains and nobody else understands.
A human is the integration. They copy, paste, reconcile, and are the only reason the two systems agree.
The transfer runs on a schedule, disagreements are surfaced as a short list, and the spreadsheet stops being load bearing.
Internal queries about policy, specification, price lists, or what was agreed in a contract.
The answer exists in a folder somewhere, so people ask the one colleague who knows where it is.
The question gets answered from your own material, with the source attached so the asker can check it.
If you have already sat through one automation programme that went nowhere, this is what happened to it.
Old automation meant writing rules, and rules hold only while the input holds still: fixed fields and formats, in a single language. Most real work is not like that, which is why the manual hours survived every previous attempt to remove them. What changed is the input problem.
If that history has made you sceptical of this conversation, we think you are right to be. The failure was in the method, and the method has changed.
Rules had to be rebuilt for every language, so anything outside the head office language stayed manual. That constraint is gone. Pick one and watch the right-hand side stay exactly where it is.
Incoming, as written
Auf der Rechnung fehlt die Bestellnummer.
German
Understood as
The reasons are well documented and they repeat, and almost none of them are about the model. Every one below is avoidable, and avoiding them is most of what we are actually being paid for.
95% of enterprise generative AI pilots produced no measurable impact on profit and loss. MIT Project NANDA (opens in a new tab)
40% of agentic AI projects are expected to be cancelled before the end of 2027. Gartner, June 2025 (opens in a new tab)
Nobody agreed what better meant
A pilot with no agreed before-number gets judged on impressions, and impressions fade. We count the hours first and you sign off on that count, so the comparison at the end is arithmetic rather than opinion.
The data was never ready
This is the expensive one, and it always surfaces later than you would like: the documents are inconsistent, the access does not exist, and the field everyone relies on turns out to be typed by hand. We check it in the first week, while walking away still costs you almost nothing.
Compliance arrived after the build
From 2 August 2026 the EU AI Act's transparency duties and its full penalty regime apply. Classification, logging, human oversight and where your data physically sits are design decisions here, settled before anything is built. We have run SOC 2 programmes, GDPR work and audit readiness, so this is not a phase we bolt on at the end.
The bill grew while nobody was watching
Token spend and cloud spend both drift upward without anyone deciding they should. We instrument cost from the first deployment and set a ceiling you approve. Bringing an existing bill back down is work we have done before, and the number for it is in the record further down.
It was built beside the work
A tool people have to remember to open is a tool they stop opening by March. Automation belongs in the system where the work already happens, which usually means the integration is most of the job and the model is the easy part.
Then the sponsor left
Plenty of working systems die because the only person who understood them changed jobs. You get the runbook, the documentation and every credential. If we vanished tomorrow it would keep running, and your own engineers could change it.
Nobody should have to sign a transformation programme to find out whether this works on their process. Each step is priced on its own and ends with something you keep.
Forty-five minutes, no charge
You describe an ordinary week. We ask what gets redone, what gets chased, and what stops working when one particular person is on holiday. There is nothing to prepare and no deck to sit through.
Ends with a straight answer on whether there is anything here worth automating. Sometimes there is not, and we would rather say so on a call than a quarter later.
Roughly two weeks, fixed fee
We sit with the process where it actually runs, count the hours, and check whether the data behind it is in a state anyone can build on. Then we write down what we would automate, what we would leave alone, and what each is worth.
Ends with a written list, the hours behind every line of it, and a ranked order to work through. It is yours whether or not you build anything with us.
Fixed price, live in 90 to 120 days
One process, built into the systems your people already use, running against real volume and measured against the baseline you already agreed. We settle the compliance posture and agree a spend ceiling before the first line of code.
Ends with that process in production, plus the runbook, the documentation and the credentials. Keeping us on afterwards stays optional and always will.
Monthly, cancel any month
Models change, suppliers change their APIs, and your own process changes whether you planned it or not. Someone watches output quality, spend and the edge cases, and the next process on the ranked list gets built when you want it.
Ends with nothing, ideally. If you find yourself thinking about the retainer, it is not doing its job.
Every fee is fixed and quoted after the call, against a scope you have read first. We do not bill by the hour, because the hours are the thing you are paying us to remove.
Automating the wrong process is worse than leaving it, because you pay to build it and then pay again to watch it.
Work that runs a few times a month
The build costs more than the hours it saves, and it will still need maintaining in three years. Keep doing it by hand.
Processes that only look manual
Some steps exist because two teams disagree about who owns the decision. Software will not settle that argument, it will just make it faster.
Anything where a rare wrong answer is expensive
If a mistake is both costly and unlikely to be noticed, the checking costs more than the doing. That work stays with a person.
Judgement your customers can feel
There is a line where efficiency starts reading as indifference. We would rather tell you where we think it sits than find out with your customers.
He came to the rescue on our 8 year old monolith, moving the database to AWS, optimizing queries and implementing caching that exceeded users expectations. Our users are ecstatic.
Your operation runs in whatever languages your customers and suppliers use, and that is precisely where the old tooling gave up. Ours does not, because nothing we build is written against a language in the first place.
One sentence is enough. You will get a straight answer on whether it is worth automating, what it is costing you now, and what we would do first. If you would rather see the arithmetic on your own numbers first, the cost model is on the home page.