Custom AI for SMEs: 6 Steps Beyond ChatGPT and Copilot
Hephaestus Technology10 min read

To move from ChatGPT and Copilot to custom AI for SMEs, choose one costly, repetitive process and measure how it performs today. Then build the lightest custom tool that improves it, run it as a short pilot with GDPR checks, and compare the results with your baseline. In our view, value you can prove comes from fixing and measuring one process, not from adding more licences.
Key takeaways
- SFA 2026 research on 404 Irish small firms reports that 92% use generic tools and 30% use customised models. The Irish Times says these are shares of the 18% of firms that use AI.
- Only 6% of 285 leaders at firms with 50–500 staff, surveyed for BGF, can measure AI ROI with confidence. Set a baseline first.
- Start with one frequent, repetitive process and the lightest custom option that fixes it.
- Pilot with human review and the GDPR points the Data Protection Commission has raised.
- Decide to scale, change or stop against your baseline.
Why do generic AI tools stop delivering value for SMEs?
Generic tools make individuals faster at drafting and admin, but survey data suggests most small-firm use stays at that level. The Small Firms Association's 2026 survey, as reported by ThinkBusiness, found that 92% use generic off-the-shelf tools such as ChatGPT and Microsoft Copilot, while 30% use customised AI models, up from 22% in 2025. The survey covered 404 companies and was conducted by Amárach Research.
The base matters. The Irish Times reports that only 18% of the firms surveyed incorporate AI into most or some of their practices, and it presents the 92% and 30% figures as shares of those AI users, not of all small firms.
Most use is light-touch. The same SFA research found that 65% use AI for managing digital files, and 63% each for sales and marketing content and admin workflows.
Measurement is weak even among larger firms. A BGF/Censuswide survey asked 285 CEOs, MDs, CPOs and CTOs at Irish and British firms with 50–500 employees and £5m–£50m turnover. It found that just 6% can measure AI ROI with confidence. These firms are bigger than most SFA members, so treat the figure as a warning rather than a direct read-across.
Our reading is that the tool is rarely the problem. The problem is the lack of a target process and a number to move. For the wider picture from GOV.UK, Eurostat and CSO data, see our AI adoption statistics for Irish, UK and EU SMEs.
What counts as a custom AI system, and how is it different from ChatGPT or Copilot?
In this guide, a custom AI system means an AI tool set up around a specific business process. It is connected to your own documents, data or software and measured against a business outcome. ChatGPT and Copilot are general-purpose, and each employee decides how to use them. A custom system has fixed instructions, a defined input and output, and an owner.
Custom does not have to mean building your own model. The Data Protection Commission notes that AI systems "can also be further specialised" by re-training, augmenting or fine-tuning them with specific datasets, often by using cloud-based AI products or licensing an existing model. Some firms do build their own. The Irish Times reports that 7% use tools they developed themselves. This guide focuses on adapting an existing model to your process.
Step 1: How do you audit what your team already uses AI for?
Start with a short survey or a 20-minute conversation with each team. Ask five questions:
- Which tasks do you use AI for, and how often?
- Which tool and which account (personal or company)?
- What information goes in, including any customer or employee data?
- What do you do with the output, and who checks it?
- Where does it save time, and where does it create rework?
The DPC suggests understanding what personal data a tool uses, where that data goes and whether the provider retains it. The output of this step is a simple list of tasks, tools, data types and risk flags. That list becomes your shortlist for Step 2.
Step 2: Which process should you customise first?
Pick a process that is frequent, repetitive, costly in staff time and easy to check, with a manager who wants it fixed. We suggest looking for these traits:
- It happens often, ideally many times a week, not once a quarter.
- There is a recognisable right answer, so errors are easy to spot.
- The inputs already exist in documents, email or a system.
- It does not make decisions about individuals. The DPC warns that relying on AI output without "critical human analysis" can create automated decision-making risks.
Examples that fit include answering repeat customer questions, drafting quotes from templates, triaging inbound email and extracting data from invoices. Choosing one process also addresses a common blocker: in a Google/Amárach survey of 400 Irish SMEs, 16% said they are unsure where to start.
Step 3: How do you set a baseline so you can prove the value?
Measure the process as it runs today for one to two weeks before changing anything. 34% of BGF respondents cite ROI uncertainty as a barrier, and a baseline reduces that uncertainty. Track:
- Volume: how many times the task happens per week.
- Time: average minutes per task, including checking.
- Quality: error or rework rate.
- Speed: time from request to completed response.
- Cost: staff hours multiplied by loaded hourly cost.
Then write down one target in a sentence, for example "cut average handling time by a third without increasing errors". Base the number on your own data, not a vendor benchmark.
Step 4: Which custom option fits: configured assistant, knowledge base, workflow automation or agent?
Choose the lightest option that hits your target, because each step up adds build effort, data work and maintenance. In our view, this also eases the skills barrier.
That barrier shows up across surveys:
- SFA research: ThinkBusiness reports that 31% of firms cite lack of technical expertise. Published reports differ on which group of firms this share covers.
- Google/Amárach: 27% cite a lack of skills.
- BGF: among the larger firms, a skills gap is cited by 40%.
The table below reflects our editorial judgement, not survey data.
| Option | What it is | Typical use cases | Data needs | Relative effort |
|---|---|---|---|---|
| Configured assistant | A shared assistant with fixed instructions, templates and reference files | Proposals, policy-consistent replies, standard reports | A few curated documents | Low |
| Knowledge base (RAG) | An assistant that retrieves answers from your own document library and cites them | Internal FAQs, product and support questions, onboarding | Organised, current documents with access controls | Medium |
| Workflow automation | AI as one step in a triggered sequence across your systems | Email triage, invoice extraction, CRM updates | Structured inputs and system integrations | Medium |
| Agent | AI that plans and takes several actions with tools, within set limits | Multi-step research, order follow-up, scheduling | Integrations, permissions, logging and guardrails | High |
Step 5: How do you run a low-risk pilot that respects GDPR and customer data?
Run a time-boxed pilot with a small group, and have a person review every output until the error rate is known. This targets the barrier Irish SMEs cited most in the Google/Amárach survey: fear of making mistakes, at 30%. Among BGF's larger firms, skills gaps (40%) and security (38%) ranked higher. We suggest a pilot of four to eight weeks. That is our recommendation; set the length around how often the process runs.
Before any personal data goes in, work through the points the DPC raised in its July 2024 blog post on AI, LLMs and data protection. It is a blog post rather than formal guidance, so also check for newer DPC or EDPB material.
- Controller role: an organisation using an AI product that relies on personal data could be a data controller, and if so a formal risk assessment should be considered.
- Rights requests: if someone asks for access to or deletion of their data within the AI system, can you do it?
- Retention: without a retention schedule, you risk non-compliance with the storage limitation principle.
- DPIA: in the section addressed to organisations creating or repurposing AI products, the DPC says a data protection impact assessment may be needed. This applies especially where the technology is new to you, where datasets are combined, or where minors or vulnerable people are involved. If your pilot configures or builds a tool, we suggest treating this as relevant.
We also suggest using test or minimised data where you can. GDPR is not the only rulebook. The EU AI Act sets separate obligations that this guide does not cover, so check how it applies to your tool before you scale.
If you want outside help testing a use case before you commit budget, Hephaestus offers an Idea Validation stage.
Step 6: How do you decide whether to scale, change or stop?
Compare the pilot against your baseline and apply rules you agreed before the pilot started. BGF data shows the gap between piloting and scaling, at least among larger firms. 13% are actively scaling proven use cases, 21% have at least one in production, 15% are prototyping or piloting and 17% are still exploring. We suggest these decision rules:
- Scale if you hit the target, errors are no worse than the baseline and users want to keep it.
- Change if you hit part of the target. Adjust the data, instructions or option, then run one more short cycle.
- Stop if the gain is marginal after two cycles. Record what you learned and move to the next process on your shortlist.
What does custom AI for SMEs cost, and how can you keep it affordable?
Costs depend on the option you choose, so price the pilot, not the vision. The cost lines we suggest planning for are:
- tool licences
- configuration or build work
- integration with your systems
- usage-based running costs
- ongoing maintenance and security
- staff time for review
Cost is a real barrier, but surveys rank it below others:
- Google/Amárach: cost was cited by 24%.
- SFA research: ThinkBusiness reports 17% citing finance and 28% citing lack of time.
Three habits keep spending under control. Reuse the licences you already pay for. Start with the lightest option. Fix a pilot budget with a stop rule.
For skills, Google's March 2026 release announced up to 10,000 AI scholarships, alongside free AI Works for Ireland events with the Local Enterprise Offices. Those events ran between March and May 2026. Check with Google or your Local Enterprise Office whether scholarship places are still open.
What is the checklist you can start today?
- [ ] List every AI tool and account in use, and what data goes into each.
- [ ] Name one person who owns the AI pilot.
- [ ] Shortlist three repetitive processes and pick one.
- [ ] Measure volume, time, errors and cost for that process for one to two weeks.
- [ ] Write one target sentence and agreed scale, change and stop rules.
- [ ] Choose the lightest option from the table.
- [ ] Check controller role, rights requests, retention, DPIA need and EU AI Act obligations.
- [ ] Set the pilot dates and the review meeting.
Frequently asked questions
Do we need to build our own AI model?
Not necessarily. The DPC notes that AI systems can be specialised by re-training, augmenting or fine-tuning, often through cloud products or licensed models. We suggest starting by configuring an existing model around your documents and processes. Some firms do build their own: the Irish Times reports 7% use tools they developed themselves.
How long should an AI pilot take?
We recommend four to eight weeks after a one- to two-week baseline. That is long enough to see real volumes and error rates, and short enough to limit spend. It is our suggestion rather than a published rule, so set the length around how often the process runs.
Who should own a custom AI project in a small business?
The person who owns the process being improved, not just the IT contact. Among BGF respondents at firms with 50–500 staff, 64% have appointed a senior leader responsible for AI strategy. In a small firm, that can be the owner or an operations lead with a few protected hours each week.
Can we put customer data into a custom AI system?
Only after checking your obligations. The DPC says an organisation using AI that relies on personal data could be a data controller, and if so should consider a formal risk assessment. Confirm you can handle access and deletion requests, set a retention schedule and check whether a DPIA is needed.
What if the pilot does not deliver?
That is a useful result, not a failure, if you set stop rules in advance. Record what limited the gain, such as data quality, the option chosen or user adoption. Then move to the next process on your shortlist.
If you would like an honest view of your idea, the risks and what we would do first, book a free 30-minute call with Hephaestus.