From a problem to a running system
Assess
We map your workflows, find where the return actually sits and cost the roadmap.
Build
We prove one use case properly, working, on your own data.
Deploy
We put it into production, integrated with your systems and signed off by your people.
Manage
We host it, monitor it, update it and support it on a retainer.
You see one owner from the first meeting through to steady state. That end-to-end ownership is the product.
We have not run this one. It is written out in full because the shape of the ninety days is what people ask about and a vague answer helps nobody. Every figure below is a starting condition of the example. None of it is a result we are claiming.
The starting position
- A builders merchant in north Dublin. Forty staff, one branch and a yard.
- About 1,400 supplier invoices and delivery dockets arrive every month as PDFs, email attachments and photographs taken by drivers.
- Two people in accounts key them into Sage. Month end closes about a week late.
- Nobody in the building has built software before and there is no IT department.
- The managing director uses ChatGPT most days and wants to know what it would take to use AI properly.
Assess
We sit in accounts for two mornings and watch the keying. We count how long an invoice takes, how often a line comes out wrong and what it costs when a supplier gets paid twice. Twenty suppliers turn out to account for most of the volume. We come back with a ranked list of what AI can take on here, a number against each one and a fixed price for the first build.
A baseline you can argue with, a ranked roadmap and a price
Build
We build against last year's invoices, which you already have. The twenty big supplier layouts first, then the long tail. Anything the system is unsure about goes to a queue and never into Sage. You see it running on your own documents in week five and tell us what is wrong with it.
A working pipeline on your own documents, with an exception queue
Deploy
It runs beside your two people for a fortnight and neither of them stops keying. We compare what the system produced against what they keyed, line by line, then show you the difference. When your accountant is willing to sign that off, the keying stops and the queue becomes the job.
A signed comparison, a live connection to Sage and a named owner for the queue
Manage
We host it, watch the queue and add new supplier layouts as they appear. You get a weekly note with volume, exceptions and anything that changed. If you decide to leave, the pipeline, the prompts and the data are yours and documented well enough for another firm to pick up.
A monthly retainer, a weekly report and an exit that is not a threat
What exists on day ninety
- Invoices and dockets reading themselves into Sage, carrying the fields your accountant cares about.
- A queue holding whatever the system was unsure about, owned by a named person, usually a handful a day.
- Your two people working that queue and the supplier questions. The keying is gone.
- A weekly report saying what it processed and what it could not.
- Documentation good enough that another firm could take the whole thing over.
This is an example and it is labelled as one because we have no clients yet. Its shape comes from the published research below, which is consistent about what separates the deployments that reach production from the ones that stall. When we have delivered work we are free to name, this section will be replaced by it.
We are new and we have no client results to show you. What we can show you is the public research, with the sources named and linked. These five figures are the reason the engagement is shaped the way it is.
- 89%
of Irish SME leaders use AI tools at work, saving about five hours a week each.
Opinium for OpenAI, 200 Irish SME decision makers, february and March 2026 - 20%
of Irish businesses have AI running inside the business. Among small firms it is 17%.
Central Statistics Office, Information Society Statistics Enterprises 2025, published February 2026 - 95%
of organisations studied got no measurable return from generative AI. 60% evaluated a tool, 20% reached a pilot and 5% reached production.
MIT NANDA, The GenAI Divide, preliminary findings, July 2025 - 67%
of work done with an outside partner reached deployment. Work built internally reached it about 33% of the time.
MIT NANDA, The GenAI Divide, preliminary findings, July 2025 - 71%
median productivity gain where the system handles the volume on its own and people review only the exceptions. All 51 deployments studied were built in small steps.
Stanford Digital Economy Lab, The Enterprise AI Playbook, 51 deployments, April 2026
The first two figures are the business we are in. Nearly every Irish SME leader has opened a chat window and very few have a system their company depends on. The other three are why we start with a two week assessment, why we build one thing at a time and why every agent we ship has a queue that a named person owns.
It starts with a two-week assessment.
The first conversation costs nothing. The assessment is a two-week decision and you own everything it produces.