How we use AI, and what we build with it.

We use AI in our own design and engineering work. Inside a client's system we use it only where a step needs to read or write language, and a person approves what matters.

What AI-native means hereAI is part of how a small team builds this much.

We use AI tools in our own design and engineering work. That is what the phrase on our homepage means in practice, and it is part of how a small team builds this much.

In the systems we build for clients, the approach is narrower. Automation handles work that follows defined rules. Where AI adds value, its role has clear limits and an accountable person approves anything that commits the business.

Where AI earns its place

Work where the input varies: enquiries that arrive in different forms, documents with no fixed layout, and drafting. Sorting, drafting, pulling out details and classifying are good candidates. We test whether the result actually improves the task once a person's review and corrections are counted.

Where ordinary software does better

Calculations, scheduled transfers, required fields and known approval rules run as ordinary software and automation. AI adds review effort and running cost, so the case for a build includes both before it is approved.

Where judgement stays with people

Decisions that need professional judgement or commercial authority stay with the business. AI can prepare material for review; the person with the authority decides the outcome.

How it sits in a systemAI does a narrow job inside a workflow a person controls.

AI does a narrow job inside a workflow a person controls.

AI does a narrow job inside a workflow a person controls.An enquiry or document arrives from a source agreed for the task. AI reads it and proposes fields, a classification or a draft. The application checks the proposal and holds anything that commits the business until a named person approves, corrects or rejects it. Only then does ordinary software carry it into the record. Failures go to an exception queue with a named owner.
AI does a narrow job inside a workflow a person controls.An enquiry or document arrives from a source agreed for the task. AI reads it and proposes fields, a classification or a draft. The application checks the proposal and holds anything that commits the business until a named person approves, corrects or rejects it. Only then does ordinary software carry it into the record. Failures go to an exception queue with a named owner.
Key
AI has a defined task and limited access.Ordinary software carries approved work onward.
The system in words

An enquiry or document arrives from a source agreed for the task. AI reads it and proposes fields, a classification or a draft. The application checks the proposal and holds anything that commits the business until a named person approves, corrects or rejects it. Only then does ordinary software carry it into the record. Failures go to an exception queue with a named owner.

Select a stop to see what it does.An illustration of the pattern, not a client's system.

Limited sources and tools

AI receives only the sources and tools agreed for the task. The application checks what it produces.

A person approves

Anything that moves money, contacts a client or changes an important record waits for a named person, who sees the source and the proposal and can edit, reject or stop it.

Failures have an owner

Rejected output, supplier failures and repeated errors go to an exception route with a named owner.

In systems we have builtWhere AI does one job, and where it is left out.

Advisory firmApproved conversation

450+ messages handled a month, every reply approved by a person.

AI drafts each reply from the assessment the enquirer completed, and the draft waits in a review queue. Nothing reaches the sales system until a person approves it. Message volume from the system's logs, averaged over recent months.

Read the approved conversation study

PapercutEnquiry to invoice

AI reads each print enquiry and pulls out the specification.

That is its only job. The numbered lead, the quote versions, the production project and the invoice connection run as ordinary software.

The Papercut project

Youth sport coachingParent-managed accounts

The child is pseudonymised before AI drafts a training plan.

A coach-assist model drafts plans from the coaching record, and the child is pseudonymised before any information reaches it. Parents control the accounts.

Read the coaching study

Professional servicesProposal to agreement

Signing is recorded by automation, not judged by AI.

Automation records engagement and signing, and every agreement signed through it carries a full audit trail. The system does not rely on a model to decide whether a proposal has been accepted.

Read the proposal study

GovernanceThe business keeps the authority over its systems.

The same standards apply to AI as to the rest of a system. They are set out in full on our security page.

  • Access follows the roleEach role reaches only the records and actions it needs. Administrative access sits behind a separate boundary.
  • Approval comes before the actionA named person approves anything that moves money, contacts a client or changes an important record.
  • The record keeps the decisionEach action links its input, who acted, the approval, any edits and the outcome.
  • Changes are reviewedChanges to permissions, integrations or models are reviewed and tested before release, with a rehearsed way back.
  • Tested on real casesBefore acceptance, AI is tested on representative cases, including incomplete inputs and conflicting evidence. The acceptance record states the accuracy needed, the correction effort and the running cost.

Responsibilities under the EU AI Act, for established service firms in Ireland.

Our paper, published 16 October 2026, with sources checked on that date.

It looks at the EU AI Act through the operating decisions an Irish service firm faces. It is general information and not legal advice: qualified legal advisers assess how the Act applies to a particular use. We supply the records of system boundaries and supplier responsibilities that review needs, and build in the requirements it agrees.

Before any AIMap the process first.

Process mapping follows a record from the first request to completed, paid work, through each decision, and notes who has authority and where information is typed in again. Exceptions get named owners, and the existing cost is measured before anything is designed. That gives automation, and AI, a defined job and gives the business a basis for judging the return.