AI product development for startups
The lab
Where does AI actually help a startup?
There are six places where AI already does good work. Pick a specimen: alongside, you'll see what it does, what it takes to build and the risks to plan for.
Customer support
What it doesAnswers common questions by chat or email from your own documentation, and hands anything it can't resolve to a person.
What it takes
- Data
- Help articles, past replies and terms of service, kept current and owned by someone.
- Model and API
- A language model over an API, with retrieval over your content (RAG), wired into your support channel.
- Evaluation
- A set of real questions with known good answers, re-run every time the prompt or the model changes.
RisksConfident wrong answers, promises the company can't keep, personal data turning up in the conversation.
Custom AI agentsDocument search
What it doesFinds the answer inside contracts, manuals and archives and quotes the passage it came from, so the reader can check.
What it takes
- Data
- Your documents in a readable format, plus the permissions for who may see what.
- Model and API
- Text indexing (embeddings) and a model that answers only from the passages it retrieves.
- Evaluation
- Test questions paired with the document you expect: you measure whether it finds it and cites it properly.
RisksOld documents contradicting new ones, access to confidential files, loose citations.
Data extraction
What it doesReads invoices, forms, emails and PDFs and pulls out structured fields for your back office or CRM.
What it takes
- Data
- A sample of real documents, messy ones included, with the correct fields filled in by hand.
- Model and API
- A model with structured output (a JSON schema) and automatic checks on every field.
- Evaluation
- Share of fields right across the sample, plus a human review queue for uncertain cases.
RisksSilent errors in figures and dates, formats it has never seen, personal data that falls under GDPR.
Custom web apps and internal toolsProcess automation
What it doesConnects your tools and handles the repetitive steps on its own: routing requests, updating records, drafting replies for approval.
What it takes
- Data
- The process written out step by step, with who decides what and the exceptions that really happen.
- Model and API
- A workflow orchestrator such as n8n, or an AI agent with narrow access to your tools.
- Evaluation
- Trial runs on real cases, a log of every step and a human-intervention rate you work to bring down.
RisksWrong actions taken automatically, permissions that are too broad, costs that climb with volume.
How AI agents workAI features in your product
What it doesPuts AI inside your app: summaries, suggestions, natural-language search, generated text for your users.
What it takes
- Data
- Your users' data, with their consent and clear limits on what reaches the model.
- Model and API
- Model calls from the backend, never the browser, with per-user usage limits and caching where possible.
- Evaluation
- Automated tests on typical and edge cases, then real usage: who tries the feature and who comes back to it.
RisksAPI costs growing faster than revenue, slow responses, reliance on a single model provider.
SaaS developmentAnalysis
What it doesReads feedback, reviews, tickets and usage data and groups them by theme, so you see quickly what customers are asking for.
What it takes
- Data
- Clean exports from your sources, dated, and stripped of personal data you don't need.
- Model and API
- A model to classify and summarise; the numbers come from code, not from the model.
- Evaluation
- Spot checks on the categories it assigns, compared against what the team already knows.
RisksPersuasive conclusions drawn from thin data, figures the model makes up, themes ignored because they are rare.
The choice
Should you build, buy or use an API?
For almost every startup, the sensible start is a model used through an API inside your own product. Buy a ready-made tool where AI isn't your edge; train your own only with unique data and a strong reason.
| Criterion | Buy a ready-made tool | Use a model through an API | Build your own model |
|---|---|---|---|
| What it is | Software with AI already inside: you configure it and use it. | Your code calls a provider's model and uses the answer in your product. | You train or fine-tune a model on your data and run it yourself. |
| When it makes sense | Internal work most companies share: meeting notes, writing, standard support. | Features that are part of your product and must behave your way. | Data only you have, high volumes, requirements general models don't meet. |
| Control | Low: you decide only what the vendor allows. | High over the flow, prompts, data sent and interface; the model stays the provider's. | Complete, along with all of the responsibility. |
| Costs | A per-seat subscription, predictable. | Pay as you go: estimate it per user and keep it in check. | The highest: data, compute, people and upkeep over time. |
| Speed | The fastest. | Fast: quick to prototype, and you can switch models if needed. | The slowest. |
| Main risk | Vendor lock-in and company data leaving the building. | Dependence on the model provider and variable costs. | Spending heavily before you know it's needed. |
In most cases the middle column is where to start.
The path
How does a startup get started with AI?
Start small and measure: four phases, each ending with something concrete in your hands.
Audit
We look at your product and processes and pick where AI saves time or creates value, using data you already have.
You come away withA shortlist of use cases ranked by value and feasibility.
Prototype
We build a first version of the chosen case on a model through an API, with real data and a bare-bones interface.
You come away withA prototype to try with the team or a handful of customers.
Evaluation
We measure quality against real examples, estimate the cost per user and decide what it can do alone and what a person must check.
You come away withA test set that stays with the project, and an estimate of running costs.
Production
We ship the feature into the product with logging, spending caps, monitoring and an easy way to report mistakes.
You come away withA live feature you can measure and improve.
Experience
Who handles AI in the group?
Within the group, AI is the day job of Giallo Studio, MPN's AI agency for Italian SMEs, which writes custom software only when it is genuinely needed.
For a startup we put that experience next to the people who build web products and apps, on the technologies we use every day. Already have a prototype built with AI tools? We can start there: prototype to production.
- AI audits
- We map your processes and data and show where AI pays off and where it doesn't.
- Process automation
- Flows that connect the tools you already use, with a person wherever a decision is needed.
- AI agents
- Agents working inside your tools with narrow permissions and every step logged.
EU AI Act and GDPR. An AI feature may fall under the EU Artificial Intelligence Act and, if it handles personal data, the GDPR. We factor them into the project from the start, but we don't give legal advice: for an assessment, please speak to your lawyer and the official texts.
Questions about AI for startups
How can a startup use AI?
In two ways: inside the product, with features such as search, summaries and suggestions, and inside operations, automating support, data extraction and repetitive work. Start with a single use case that has clear value, and measure it.
Should we use a model through an API or build our own?
At the start, almost always a model through an API: it is quick to try and you can switch provider. A model trained or fine-tuned on your data makes sense when you hold data nobody else has, run high volumes or have requirements general models don't meet.
How much does it cost to run AI in a product?
You pay per use: the cost depends on how many requests each user makes, how much text goes to the model and which model you choose. That's why we estimate the cost per user during the prototype and add usage limits and caching before going live.
Does customer data end up training AI models?
It depends on the provider: their terms say whether data you send can be used for training, and they are worth reading before you choose. We send the model only the data the task needs; for GDPR obligations, please consult your lawyer.
Does the EU AI Act apply to my startup?
It may: it depends on what your AI system does and who uses it, and the obligations differ a great deal between a chat assistant and a high-risk system. We don't give legal advice, so for an assessment please speak to a lawyer.
Further reading
AI and automationHow startups can use AI: where it pays off and how to pick your first use caseRead 
AI and automationThe EU AI Act for startups: what it requires, when it applies and what to do nowRead 
AI and automationGDPR and LLM apps: what founders must get right before shipping AI featuresRead 
AI and automationLLM application development: architecture, API costs and evaluation, a practical guideRead
Got a use case in mind?
We reply within one business day.
Tell us about it