An AI agent is a system that takes in data, makes a decision with a confidence score, produces an output, and hands back to a person when the answer is uncertain or the stakes are high. It is not a chatbot. It is the engine behind a workflow.

What is the difference between agentic AI and generative AI?

Generative AI produces content when prompted: text, images, code. Agentic AI pursues an outcome: it takes a job, works through the steps using your data and systems, and delivers a result, escalating to a person when unsure. ChatGPT writing an email is generative. A system that reads a request, checks your ERP, drafts the quote, and routes the odd one to your estimator is agentic.

What is the difference between an AI agent and automation like Zapier or Copilot?

Automation tools follow fixed triggers and templates. Copilot Studio and Zapier have grown genuinely capable, and for many workflows they are the right answer. A custom agent earns its keep when the job needs deeper integration, decision logic tested against your own historical cases, an audit trail, and clear ownership. If a flowchart covers your process, use the automation tool. We will tell you which one you need.

How much does it cost to build an AI agent?

A proof of concept typically costs €8,000 to €30,000 and runs 3 to 6 weeks. Production builds mostly land between €25,000 and €100,000. Discovery is €5,000. We give a firm price at the end of Discovery, based on a written specification.

What does an agent cost to run after go-live?

Three parts: hosting, model usage, and optional support. Hosting and model usage are itemised for your volumes during Discovery, so you see the running cost before you commit to the build. Ongoing monitoring and improvement, if you want us to do it, is €1,500 to €3,000 per month.

How long does it take to build an AI agent?

A proof of concept takes 3 to 6 weeks. A production agent typically takes 10 to 20 weeks from Discovery handoff, depending on integrations and compliance requirements, including a calibration period before full reliance.

Can an AI agent work with our existing ERP or business systems?

Yes. Agents we have shipped connect to Sage, SAP Business One, Odoo and custom-built platforms, and we support Microsoft Dynamics and other common commercial ERPs. The agent usually sits alongside your existing system through secure, read-only access by default; write access, where a workflow needs it, is scoped, logged and agreed in the control sheet. Any API or middleware work is identified during Discovery.

How do you keep an AI agent safe and under control?

Scope and control are designed before the build starts. The agent's read access, write access, confidence thresholds, and human review path are all agreed and documented in a control sheet. Every decision is logged with its inputs, and every override is recorded. You can audit what the agent has done, and you can stop it.

What AI models do your agents use?

We build on OpenAI, Anthropic, and open-weights models, choosing per task. The value is in the integration, the decision logic, the guardrails, and the connection to your data. We do not change a production model automatically; candidate updates are tested against your evaluation set first.

Do you work with UK companies?

Yes. We are Ireland-based and deliver across Northern Ireland and the wider UK, with clients in the EU and US as well. Most work is delivered remotely with on-site visits where the project needs them. Government grant supports apply to eligible Irish and Northern Irish businesses only; pricing and process are the same for everyone else.

When is an AI agent the wrong answer?

When the rules are simple and stable, when a good product already covers the job, when the data is not captured reliably, when the process changes constantly, or when nobody can own the human review. In those cases we recommend something cheaper, or nothing. We tell you during Discovery, and some Discoveries end exactly that way.

Custom AI Agent Development Company

Most AI agent demos look impressive because the process is clean. Most real businesses are not. Real agents deal with messy job categories, half-filled fields, and exceptions nobody wrote down. That is the work. The demo is the easy part.

Key facts

Based in: Longford, Ireland. Serving Ireland, the UK, the EU, and the US.
Client size range: 15 to 13,000+ employees. Function-level engagements at larger organisations.
Typical first agent: €8,000 to €30,000 proof of concept. Production builds €25,000 to €100,000+.
Typical timeline: 3 to 6 weeks for a proof of concept. 10 to 20 weeks to production.
Invoicing: Euro, sterling or US dollars.
IP ownership: Yours. The code, the prompts that run it, evaluation sets and documentation, handed over at completion.
Hosting: Ireland or EU only.
Delivery: Senior-led team. Mostly remote, on-site where the project needs it.

Which conversation are you here for?

Different people come to this page for different reasons. Go straight to yours.

You run the business

Start with the quoting agent, one real decision traced start to finish, and the prices. Jump to the quoting agent →

You are the technical half

The engineering notes further down are written for you. Jump to the engineering notes →

You lead a function inside a large organisation

Start with how we work from the UK, Europe and the US, and the procurement facts. Jump to working with us →

The Jobs Agents Are Built To Take

These are the patterns that come up again and again. If yours is a different shape, tell us anyway.

Decisions queue behind one person

Quotes, approvals, classifications. The person is good at the job. That is the problem. Everything waits for them, and when they are out, it waits longer.

Documents get read by hand

Invoices, specs, certs, delivery dockets. Someone opens each one, finds the numbers, types them somewhere else, and hopes nothing was missed.

Nobody watches the data until it bites

Reject rates drift. Margin slips on one product line. The signal was sitting in the data for weeks. Nobody had time to look.

Off-the-shelf tools stop at the surface

Copilot and Zapier are genuinely useful for fixed steps. A custom agent earns its keep when the job needs a tested judgement call on your own operational data, with an audit trail behind it.

You cannot hire your way out

Experienced people are hard to find and slow to train, and the volume keeps coming.

One Decision, Start to Finish

The clearest way to show what a production agent is: walk through one real run of the quoting agent we built for a 60-person building services contractor. Details are simplified and redacted. The shape is real, and the full case study is here.

A quote request arrives with a job spec attached.
The agent reads it and matches it against five years of completed jobs in the company's ERP, through secure, read-only access.
It drafts an itemised quote and assigns a confidence score showing how closely the job matches historical patterns.
This one lands below the agreed threshold. Unusual job category. So the draft, the score and the reasons route to the senior estimator instead of going out.
The estimator adjusts one line and approves. Fifteen to twenty minutes, not most of a morning.
Every step, input and override is logged. The audit trail is part of the product.

The measured results, tracked over a four-week baseline and the first six weeks live: complex quotes went from 1 to 2 hours to 15 to 20 minutes. Simple reactive quotes from 20 to 30 minutes to under 10. Turnaround to the client from 24 to 48 hours to same day. Roughly 15 to 20 hours of estimating admin freed every week. That is 80% faster quoting, and the senior estimator still owns the judgement calls. That is the design, not a limitation. Figures last checked July 2026.

"The big thing for us was that Declan's pricing knowledge, all the stuff that was only in his head, is now in the system. The rest of the lads can actually produce accurate quotes without waiting for him."
Operations Director, Building Services Firm. Company name withheld at the client's request; a reference call is available.

And the honest bit, because you should ask every vendor what broke: in the first weeks, every AI draft was independently checked against a manual estimate by the senior estimator while the confidence thresholds were tuned. That four-week calibration period is now standard on every agent we build. Trust is earned on real data, not assumed.

When an Agent Is the Right Answer, and When It Is Not

An Honest Assessment

Agent territory: the task needs judgement on your data. Quoting, triage, cross-checking, exception handling, monitoring for what matters. High volume, real patterns, decisions a person currently makes by digging.

Not agent territory:

The rules are simple and stable. If a flowchart covers it, RPA or Zapier is cheaper and easier to maintain.
A product already does it well. If off-the-shelf software genuinely covers the job, buy the product. Custom earns its keep on judgement, integration depth and ownership, not on principle.
The data does not exist, or is not captured reliably. Fix the capture first. An agent can only work with data that exists and can be trusted.
The process changes every few months. An agent encodes a process. Unstable process, stale agent.
Nobody can own the review. Early agents need human review while confidence is earned. If nobody has capacity to own that, do not start.
The stakes rule out partial autonomy. Some decisions should not be delegated at any confidence level. We will say so.

We tell you this during Discovery, and some Discovery engagements end with us recommending ordinary software instead of an agent. That is the point of Discovery. We would rather lose the build than ship an agent that gets quietly switched off.

What Kinds of Agents We Build

Five kinds of AI agents for business, and the jobs they do.

Decision agents

Read historical data and produce a decision with a confidence score. Quoting, pricing, classification, triage. Anything below the confidence threshold routes to a person.

Document agents

Extract, validate and cross-check information from PDFs, forms and records, then pass it where it needs to go. With a full audit trail.

Workflow agents

Carry a multi-step process from trigger to outcome. Human checkpoints sit wherever the stakes justify them, not autonomy for its own sake.

Monitoring agents

Watch operational data streams and flag what matters. Anomalies, drift, thresholds crossed. Ordinary days stay quiet.

Conversational agents on your own data

Question-answering over your own databases and documents in plain English, with the same permissions, logging and grounding as everything else we build. Not a public chatbot.

Bounded Autonomy, In Practice

The design questions that matter are not about the model. They are about scope and control. Every agent we ship has a control sheet agreed before build:

What it can read. Which systems, which fields, which records.
What it can change. Usually less than it can read. Read-only is the default; write access is scoped, logged and earned.
What always needs a person. Some actions require approval at any confidence level.
The confidence threshold, and where uncertain work goes.
What gets logged. Every decision, every input, every override.
The stop condition. How it halts, and who can halt it.
The named owner on your side, and the review cadence.

If you cannot audit an agent, you cannot trust it. This is what the industry now calls governance. We call it how you build software that handles money and reputation.

Our Own Agent Ships in September

The strongest proof we can offer is not a slide. It is work you can inspect: the client agent above, in production today, and the product we are shipping for ourselves.

Reactable AI is the Deep Purple team's own product, funded by Enterprise Ireland's Pre-Seed Start Fund. We built and ran the original version commercially, then took it down to rebuild it properly as an agent. The new Reactable launches in September 2026, built first for salons and personal-care businesses. The owner photographs their work, the agent turns it into posts and runs the schedule, and one hard rule carries through the whole design: nothing publishes without the owner's approval. Autonomy is earned as approvals build trust, never assumed, and the owner controls every widening of permissions. The site is live at reactable.ai; judge the plan on this page and the product when it lands.

Building and running the first version, with our own revenue and support burden on the line, taught us what breaks after week one: permissions get tested by real users, monitoring earns its keep, and cost per task matters more than any demo. Retiring our own product to rebuild it as an agent was the expensive, correct call. Those lessons are in every agent we build for clients.

Before You Buy From Anyone

One caution from inside the industry. Gartner estimated in June 2025 that only around 130 of the thousands of vendors claiming to sell AI agents were building genuine agentic systems. That number gets quoted everywhere now, including by vendors it describes. So do not take the sentence, take the test: ask every vendor to show you a system that has run in production for months, and ask what broke. Ours has. The quoting agent has been live for months, and our answer to what broke is two sections up.

From Idea to Production Agent

This is what our AI agent development services look like, start to finish.

Discovery

Map the workflow, the data, the decision points, and the review path. €5,000 as a standalone engagement. What we need from you on day one: a named workflow owner, access to representative historical cases, and read-only system access for review. Output: a specification and a firm build price, not a vague proposal. Some Discoveries end with us recommending something simpler. You keep the specification either way.

Proof of concept

A working agent on your real data, tested against acceptance criteria agreed before we start. Typical investment €8,000 to €30,000, running 3 to 6 weeks, and most first agents land in the lower half of that range. You see the confidence scores and the failure cases before committing further. What moves the number: the state of your data and the number of systems the agent must read.

Production build

Integration, permissions, audit trails, user training, deployment, and a calibration period where drafts are checked against your current process until the thresholds are tuned. Most agent builds land between €25,000 and €100,000. A single-system agent usually sits in the lower half of that range; multi-system or regulated builds in the upper half. Typically 10 to 20 weeks. Larger programmes are scoped separately.

Run

Operate it yourself with full documentation and runbooks, or retain us for monitoring and improvement at €1,500 to €3,000 per month. Model and hosting usage costs are separate, itemised for your volumes during Discovery, so the running cost is visible before you commit to the build. Performance improves only when feedback is reviewed, changes are tested, and someone owns that process.

All prices ex-VAT and stated in euro. Sterling and US dollar invoicing available. The price is the price for everyone. Irish and Northern Irish businesses may qualify for government supports that reduce the out-of-pocket cost, and we confirm eligibility during Discovery.

Working With Us From the UK, Europe and the US

Most of our delivery is remote, and it is specific, not vague. A named senior lead. A weekly working session with a demo of what changed. Shared documentation you can read at any time. On-site visits where the project needs them. Every project is delivered by senior engineers with over ten years of experience each. You work directly with the people doing the work.

For UK businesses: same language, same working hours, a short flight, and the team has delivered for UK organisations before. Data transfers between the UK and Ireland run under the UK adequacy decisions, renewed in December 2025, and the exact route for your data is recorded in the DPA and its schedules.

For EU businesses: Ireland is an EU member state. EU data residency, hosting in Ireland or the EU only, DPA as standard, euro invoicing.

For US businesses: US dollar invoicing, and the team has delivered for US-headquartered organisations including Qualcomm and ResMed. Our afternoon is your East Coast morning, a solid daily overlap; the West Coast overlap is thinner and we plan around it. Hosting stays in Ireland or the EU. For many US companies that is a feature, not a limit: if you serve European customers, your system already lives under EU rules. If your policy requires US-based hosting, we are not your builder, and we will say so on the first call.

The procurement facts. Purpledecks Limited, trading as Deep Purple AI Consulting, Company No. 513907, registered in Longford, Ireland, operating since 2012. Insurance and company documentation available for vendor onboarding. We do not hold ISO 27001. What we do instead: DPA as standard, privacy risk screening on every use case, EU-only hosting, role-based access, and a full audit trail on every agent decision. In practice that means encryption in transit everywhere, encryption at rest where the sensitivity of the data warrants it, databases locked down with network isolation and tightly scoped access, defined backup and recovery, and incident notification commitments written into the DPA. A security summary is available for your vendor review process.

Organisations this team has delivered for include RTÉ, ResMed, Qualcomm, Musgrave, Eir, Sage, IDEMIA and Dentsu. That roster is 14 years of software delivery for complex organisations, not a list of agent clients. The agent proof is the two systems above, and we will name delivery references at your scale when you ask.

Your Data, Your Agent

Your business data is not used to train public AI models. Application hosting is in Ireland or the EU. Where a model provider processes data, the DPA and its data-flow schedule state exactly what goes where, including every subprocessor. Every use case is screened for privacy risk before build, and a Data Protection Impact Assessment is completed where the processing is likely to put people's rights at high risk. After full payment, you own what we build for you: the code, the prompts that run it, the evaluation sets, the runbooks and the documentation. Everything needed to operate, maintain and extend the system without us. The tools and methods we used to build it stay ours, the same way buying the car does not buy the factory. If a build tool is needed to run or maintain your system, it transfers with it. Every agent logs inputs, outputs, confidence scores and overrides. On the EU AI Act: the transparency rules apply from 2 August 2026, and the EU's 2026 amendments moved the high-risk obligations out to December 2027. Where your use case falls under the Act, we assess its classification with you during Discovery rather than waving a compliance badge at you.

Architecture and Engineering Notes

For the technical reader. Maintained by Barry Gough, CTO, who architected Reactable AI. Last reviewed 24 July 2026. Every line below is delivery practice, not aspiration.

Models. We build on OpenAI, Anthropic and open-weights models, chosen per task. Smaller models route simple steps. Larger models carry the reasoning. We do not change a production model automatically: candidate updates are tested against the evaluation set and released only when they meet the agreed thresholds.

Single agent first. Most jobs need one well-bounded agent, not a fleet. Multi-agent structure only where the workflow genuinely branches.

Integration. Shipped: Sage, SAP Business One, Odoo, and custom-built platforms. Supported: Microsoft Dynamics and other common commercial ERPs, assessed case by case. API-first, with event triggers and middleware where APIs are thin. Read-only is the default; write access is scoped, logged and earned.

Evaluation. Every agent ships against a test set built from your historical cases, with acceptance thresholds agreed before build. Changes to prompts or models are re-tested against the set before release.

Observability. Every decision is logged with inputs, output, confidence and any human override. Traces are reviewable. Rising review rates and drift get flagged, not discovered.

Failure handling. Provider outages and rate limits degrade to the human path, not to silence. Work queues, people are notified, nothing is dropped quietly.

Portability and exit. You own the project: the code, the prompts that run it, the evaluation sets and the runbooks. Model choice is per task, not a platform commitment, and the handover documents how a model swap would run.

Protocols. We use the Model Context Protocol where it fits the client stack, and we build our own MCP connectors, including for our internal systems. Protocols outlive frameworks.

Barry Gough is the CTO of Deep Purple AI Consulting. With an MSc in Computer Science from University College Dublin, where machine learning was a core focus of his studies, and over 20 years building production software systems, Barry brings formal ML training and deep hands-on engineering experience to every AI and data analytics engagement.

Barry completed his masters at UCD in 2011, studying ML algorithms, statistical modelling and data-driven systems just as big data techniques were maturing and deep learning was about to transform the industry. Barry joined Purpledecks in 2016 and has led the technical delivery of enterprise projects incorporating machine learning, computer vision, data classification, predictive features and recommendation engines for commercial clients across Ireland and the UK.

In 2023, Barry architected and built Reactable AI from the ground up as an internal Deep Purple product, a system that generates and optimises marketing campaigns across channels against measured results. Reactable AI was one of Ireland's earliest production deployments of autonomous AI agents, requiring him to design systems where AI made real decisions with real consequences.

At Deep Purple, Barry leads all technical delivery: AI system architecture, machine learning model development, data pipeline engineering, and manages a team of experienced ML engineers and applied statisticians. His combination of formal ML education, years of incorporating AI into commercial projects and hands-on experience architecting autonomous AI systems means clients work with a technical lead who can make genuine engineering decisions about AI.

Custom AI Agent Development