AI Customer Insights Analysis: 2026 Guide for SMEs
Hephaestus Technology12 min read

AI customer insights analysis means using language models and text classification to sort unstructured feedback (support tickets, reviews, NPS comments, survey free text, call notes) into themes, sentiment and drivers you can act on. An SME can do it reliably by starting with one business question, building a small labelled sample to check the AI against, and keeping personal data under control. The legal obligations are lighter for text than for voice or video, where inferring emotions brings specific EU AI Act duties.
Key takeaways
- AI customer insights analysis turns unstructured feedback into themes and drivers. It works best on text you already collect.
- Start with one question, one or two data sources and a small codebook, and only then add tools. A general LLM, used under suitable terms, is often enough for a pilot.
- Never take AI themes or sentiment scores on trust. Compare them with a human-coded sample before anyone acts on them.
- GDPR applies because feedback contains personal data. Under the EU AI Act, text sentiment mainly triggers the AI literacy duty, while inferring emotions from voice or video brings much heavier duties.
- Insights only count when each theme has an owner, an action and a metric you will check again.
What is AI customer insights analysis, and what can it realistically tell you?
AI customer insights analysis is the automated coding of customer comments into categories (topics, sentiment, urgency, product area), followed by counting and comparing those categories over time. Two families of technique do the work. Classic text classification is trained on labelled examples and can be inexpensive to run at volume once trained. Large language models (LLMs) can read a comment, assign it to themes you define, summarise clusters and suggest new themes you had not thought of.
Realistically, it can tell you:
- What people talk about most, and how that shifts month to month.
- Which themes tend to appear alongside negative comments, such as delivery delays or onboarding confusion.
- Where feedback differs by segment, channel, product or region.
- Verbatim examples that make a theme concrete for the team.
The limits matter just as much. Sentiment scores can struggle with sarcasm, mixed comments ("great product, awful support") and industry jargon. Feedback is also a self-selected sample: the customers who choose to write are not necessarily representative of everyone. Treat the output as a structured way to listen, not as a measurement of all your customers.
Which customer data sources produce the most useful insights for an SME?
The most useful sources are the ones where customers explain why in their own words, and which you can link to an outcome such as churn, refunds or repeat purchase. For most SMEs that means:
| Source | Strength | Watch out for |
|---|---|---|
| Support tickets and live chat | High volume, specific problems | Skews towards problems; contains names, order numbers, sometimes payment details |
| NPS or CSAT free-text comments | Linked to a score, easy to segment | Short answers, survey fatigue |
| Public reviews (Google, app stores, Trustpilot) | Unprompted, comparable with competitors | Small samples, platform terms of use |
| Sales and account-manager call notes | Explains lost deals and expansion | Inconsistent note quality |
| Call recordings and transcripts | Rich detail | Highest privacy and AI Act risk if you analyse voice for emotion |
If your data is scattered across tools, sort that out first. Our guide on where to start with an SME data strategy in 90 days is a useful starting point before adding AI.
How do you set up AI customer insights analysis in 7 steps?
The setup below is the method we recommend for a first pilot. It is a practical sequence, not an industry standard.
- Pick one decision question. For example: "Why do first-time buyers not reorder?" A clear question decides which data and themes matter.
- Choose one or two sources and export a recent, representative batch, such as the last three to six months.
- Minimise and pseudonymise. Strip names, emails, phone numbers and order IDs before text reaches any external model, unless you have a clear need and a contract that covers it.
- Draft a codebook. Write 10 to 20 themes, each with a one-line definition and an example comment. Let the LLM propose extra themes, then have a person accept or merge them.
- Hand-code a sample. Two people from the team label a few hundred comments against the codebook. This becomes your test set.
- Run the AI and compare. Classify the same sample, measure agreement (see below), then adjust prompts or definitions until the results are good enough for your decision.
- Scale and schedule. Run the full dataset, then repeat monthly or weekly, re-checking a fresh sample each cycle so you catch drift.
Should you buy a feedback analytics tool, use a general LLM, or build a custom pipeline?
For most SMEs the answer is to pilot with a general LLM, buy a tool if the workflow sticks, and build only when your data or integration needs are unusual. The table below is our general assessment, not benchmarked data:
| Feedback analytics tool | General LLM (chat or API) | Custom pipeline | |
|---|---|---|---|
| Time to first results | Fast | Fastest | Slowest |
| Upfront effort | Configuration | Prompting and spreadsheets | Engineering and data work |
| Ongoing cost driver | Per-seat or per-volume subscription | Usage-based API fees plus staff time | Hosting, maintenance, people |
| Control over themes and accuracy | Medium, within the vendor's model | High but manual | Highest |
| Integrations (CRM, helpdesk) | Usually built in | Manual exports | Built to fit |
| GDPR checks | Vendor DPA, hosting region, sub-processors | API terms, data retention, training opt-out, region | Your own architecture choices |
| Best for | Teams that want dashboards without engineering | Pilots and low volume | High volume, multiple sources, specific workflows |
Check pricing on each vendor's current pricing page. It changes often, and volume tiers vary widely.
How do you check that the AI's themes and sentiment scores are accurate?
You check accuracy by comparing the AI's labels with a human-coded sample it has not seen, before anyone acts on the results. In practice:
- Measure human agreement first. If your two human coders often disagree, the codebook is unclear, and no model will fix that.
- Compare AI with humans theme by theme. Overall agreement can hide one important theme the model consistently misses.
- Use a chance-corrected statistic such as Cohen's kappa alongside simple percentage agreement. Some agreement happens by luck, especially when one theme dominates.
- Read the errors. Ten misclassified comments often reveal a vague definition or a prompt problem.
- Re-test on fresh data each cycle and after any change of model, prompt or product.
Our practical 7-step guide to LLM evaluation before launch can be adapted into a more structured testing routine. Agree in advance what level of agreement is "good enough" for each decision. A board-level pricing change needs more confidence than a tweak to an FAQ page.
What do GDPR, UK GDPR and the EU AI Act require when you analyse customer feedback with AI?
GDPR applies to almost all feedback analysis because comments contain or link to identifiable people. For text sentiment, the EU AI Act asks comparatively little beyond AI literacy. It adds significant duties if you infer emotions from voice or video. The points below summarise the regulation texts. They are not legal advice.
Under the GDPR:
- Purpose limitation and reuse. Article 5(1)(b) limits further processing to purposes compatible with the original one, and Article 6(4) lists the factors for judging compatibility. Reusing support tickets to improve service is often arguable. Document your reasoning anyway.
- Lawful basis. You need a basis under Article 6(1). Legitimate interests (Article 6(1)(f)) is one option for service improvement. It requires weighing your interests against customers' rights, and people can object to processing on that basis under Article 21.
- Transparency. Articles 13 and 14 require you to tell people how their data is used, so update your privacy notice to mention AI analysis of feedback.
- Vendors and LLM providers. Where a vendor processes data on your behalf, it is a processor, and Article 28 requires a contract with it. Article 28 does not name model training or retention specifically, but it is good practice to make sure your contract or terms cover both.
- Transfers. Articles 44 to 49 govern sending personal data outside the EU. Choosing EU-hosted processing can reduce transfer questions but does not remove them automatically. Remote access by a non-EU vendor or sub-processor may still involve a transfer, so check the sub-processor list.
- Automated decisions. Article 22 restricts decisions with legal or similarly significant effects based solely on automated processing. In our reading, aggregate theme reports usually fall outside it. Scoring or treating individual customers differently on that basis may not.
- Voice. Voice data becomes biometric data under Article 4(14) when specific technical processing is used to identify someone or confirm their identity. Article 9 then imposes stricter conditions on processing it for that purpose.
- DPIA. Article 35 requires a data protection impact assessment where processing is likely to be high risk. In our view, doing one for large-scale analysis or call recordings is sensible even when the case is borderline. Our step-by-step DPIA guide for AI features walks through the process, and our guide to using LLMs safely with customer data under GDPR is relevant if you send feedback to an external model.
Under the EU AI Act:
- AI literacy. Article 4 requires deployers to take measures to ensure sufficient AI literacy among the staff who use AI systems, whatever the risk class. That includes text-only feedback analysis.
- Emotion recognition definition. Article 3 defines an emotion recognition system as one that identifies or infers emotions on the basis of biometric data. Text-based sentiment on reviews and tickets is therefore generally outside that definition. Inferring emotion from customers' voices or faces is not.
- Transparency and high risk. Article 50(3) requires deployers of emotion recognition or biometric categorisation systems to inform the people exposed to them. Annex III lists emotion recognition among high-risk uses, in so far as its use is permitted under Union or national law.
- Workplace ban. Article 5(1)(f) prohibits AI systems that infer emotions in the workplace and in education institutions, except for medical or safety reasons. This matters if a call-analytics tool scores your own agents' emotions.
- Chatbots. Article 50(1) requires providers to design AI systems that interact with people so that those people are informed they are dealing with AI. If you collect feedback through a chatbot, check that the product you use meets this.
When do these apply? Under Article 113 of the original text, Article 4 and the Article 5 prohibitions have applied since 2 February 2025. The Article 50 transparency duties and Annex III high-risk obligations apply from 2 August 2026. The European Commission has proposed amendments (the "Digital Omnibus") that would delay some of these dates. Check whether they have been adopted before planning around them.
In the UK, UK GDPR shares the same core principles. However, the Data (Use and Access) Act 2025 amends it, including its rules on automated decision-making, a new "recognised legitimate interests" basis and purpose-compatibility rules. Its provisions commence in phases, and the GDPR article numbers above do not map one-to-one to UK GDPR, so check current ICO guidance. The EU AI Act is not UK law, but under Article 2 it can reach providers and deployers outside the EU where an AI system's output is used in the EU.
The relevant regulators are the Irish Data Protection Commission for Irish businesses, the ICO in the UK and the EDPB for EU-level guidance.
What does it cost, and how long before you see useful results?
A text-only pilot using a general LLM costs mostly staff time: exporting data, writing the codebook and hand-coding the test sample. API fees are usually the smaller part at SME volumes, but confirm this against current provider pricing. Costs rise with:
- the number of sources and integrations
- the volume and frequency of analysis
- the need for EU-only hosting or private deployment
- voice transcription and any emotion analysis, which also bring legal overhead
We have no reliable benchmark figures for timelines. A first pilot depends mostly on how quickly the data export and hand-coding happen. A recurring, integrated pipeline is a longer project than a one-off pilot. If you need engineering help, our build stage describes how Hephaestus approaches data and AI work.
How do you turn AI-generated insights into decisions instead of another dashboard?
Insights become decisions when each important theme has a named owner, a specific action and a metric you will revisit. A simple routine works:
- Monthly insight review. Cover the top five themes by volume and the fastest-growing theme, each with three verbatim quotes.
- One owner per theme. That person decides to act, investigate or deliberately ignore.
- Link to outcomes. Where possible, check whether customers raising a theme churn, refund or reorder at different rates.
- Close the loop. Track whether the theme shrinks after the fix, and tell customers what changed.
- Retire themes that no longer drive decisions, so the codebook stays sharp.
Frequently asked questions
Is a general chatbot like ChatGPT or Claude good enough to analyse customer feedback?
For a pilot, often yes, provided you remove personal data first and check the output against a human-coded sample. For recurring analysis, use the provider's API or a business plan under a contract that sets retention and training terms. Terms vary by plan, so check them before pasting in any customer data.
Does text sentiment analysis count as emotion recognition under the EU AI Act?
Generally no. The AI Act defines emotion recognition systems by reference to biometric data, which text reviews and tickets are not. Analysing voice or video to infer emotion is a different matter and brings transparency and high-risk obligations. The AI literacy duty in Article 4 applies either way.
Do I need customer consent to analyse support tickets with AI?
Not necessarily. Legitimate interests under Article 6(1)(f) GDPR is one possible basis for service improvement. It should be supported by a documented balancing test, a compatibility assessment for reuse and an updated privacy notice. Customers can object to processing on this basis, so record your reasoning.
How many comments do I need to hand-code to check accuracy?
There is no universal number. Code enough comments that every important theme appears a reasonable number of times, often a few hundred in total for a pilot. Increase the sample for rare but important themes.
Can a UK business use the same approach?
Yes, the method itself carries over, because UK GDPR rests on the same core principles of lawful basis, purpose limitation, transparency and DPIAs. However, the Data (Use and Access) Act 2025 changes parts of UK GDPR, including automated decision-making and legitimate interests, so the article references here do not map one-to-one. Check current ICO guidance, and check whether the EU AI Act applies if your AI outputs are used in the EU.
If you are weighing AI features or data work built around customer feedback, book a free 30-minute call with Hephaestus.
Cover photo by Mikhail Nilov on Pexels.