If you run a SaaS product or a company weighing AI tools, you have probably already heard the question "can we do this with AI?". Your first step is to write down three tasks that repeat in your work, then answer for each: who would check the result, and what does a mistake cost? If you have no answer to the second question, AI is not yet the right tool for that task.
This text separates what works in production from what is marketing, shows how to choose between an external API, a purpose-built service and your own model, and lists the costs and checks worth doing before release. If you are building an AI feature rather than buying one, see the SaaS product lifecycle guide and the SaaS MVP development framework. The wider context is in the AI-augmented development guide.
Where does AI in a SaaS product really deliver, and where is it oversold?
AI delivers where the result is easy to verify: extracting data from documents, classification, summarisation, semantic search and anomaly detection; it is oversold in chatbots without your own documentation and in dashboards nobody opens.
How many European companies already use AI?
Eurostat reports that in 2025, 19.95% of EU enterprises with ten or more employees used AI technologies, up from 13.48% in 2024: 55.03% of large, 30.36% of medium and 17% of small enterprises. Among enterprises using AI, 34.70% used it for marketing or sales and 31.05% for organising business administration processes (Eurostat, Use of artificial intelligence in enterprises, 2025). The article is about the EU, so do not transfer the figures to the local market without checking.
My conclusion is cautious: adoption is growing, but most enterprises still do not use AI, so there is no reason to feel pressure to add it without a clear use case.
What does "SaaS with AI" mean in practice?
The phrase covers everything from a chat widget to a full inference pipeline. The concrete, useful cases are these.
Document processing. You upload a contract or an invoice, extract structured data and flag deviations. The alternative is manual entry. Amazon Textract, for example, extracts text, forms and tables, and has dedicated operations for invoices and receipts (AWS Docs, What is Amazon Textract).
Classification and tagging. Categorising support tickets, flagging content that breaks the rules, tagging uploaded files. Accuracy can be measured, so evaluation is simple.
Summarisation. Long documents, meeting transcripts and activity logs become a short summary. Users adopt this quickly because the alternative is reading everything.
Semantic search. Search by meaning rather than exact keywords. Supabase describes it as part of a toolkit for AI applications on Postgres and pgvector (Supabase Docs, AI). The implementation is well known: embeddings, vector storage and retrieval.
Anomaly detection. Unusual patterns in usage, transactions or sensor data, especially where edge cases have real consequences, for example in fintech and logistics.
Text drafts. First drafts of reports or messages adapted to the product's context. The user edits instead of writing from scratch, and the result is reviewed before use.
Where is AI oversold?
A chatbot that does not know your product. Without retrieval from your real documentation (RAG) it will invent answers and route users worse than a static FAQ. When RAG makes sense and when it does not is explained in the post RAG for B2B SaaS.
"AI insights" nobody reads. A summary on a dashboard nobody opens does not make the dashboard useful. If the data is not usable for decisions, AI does not fix that.
An LLM call for a plain condition. If the decision is a few if-else branches, an LLM adds delay, cost and non-determinism with no benefit.
Recommendations for a small catalogue. With around fifty items, sorting by popularity is usually enough and needs no infrastructure. That is my estimate, not a measured fact; measure on your own data.
How do you choose between an external AI API, a purpose-built service and your own model?
Start with an external API because it is fastest to try, use a purpose-built service for a narrow task, and build your own model only for proven reasons.
| Option | When it is the right choice | What the choice costs |
|---|---|---|
| External LLM API (Anthropic, OpenAI, Google) | You are still discovering what users want, the feature is not a key differentiator, you need to ship fast | Data leaves your infrastructure, you depend on a vendor |
| Purpose-built service (for example Textract for documents) | The task is narrow and clear, you need an SLA | Less flexibility outside that task |
| Your own or fine-tuned model | General models perform poorly on your domain, you have proprietary data or a ban on external processing | High upfront cost, your own maintenance |
Most early products should stay in the first row of the table until they get a concrete reason to leave it. For regulated data (health, finance) an external API needs a data processing agreement before the first call.
One documented example from my work: the Calia assistant in Callidus (React + Firebase + Stripe Connect, a UK clinic SaaS) runs on Vertex AI through the @google/genai SDK, with thirteen regex patterns against prompt injection, per-hour request limits and a 90-day GDPR cleanup. It is an example of a protective layer around a model, not a recommendation of a particular vendor.
How much does an AI feature cost and how do you keep it under control?
AI cost is driven mainly by the tokens you send and receive, so reduce it with a smaller model, cached repeated context and batching of tasks that do not need an instant answer.
What does a vendor publish about prices?
According to Anthropic's pricing page on 30 September 2026, Claude Haiku 4.5 costs 1 USD per million input tokens and 5 USD per million output tokens, and Claude Sonnet 5.5 costs 2 USD and 10 USD. Reading from the prompt cache costs 0.1 of the base input price, a five-minute cache write costs 1.25 of the base price, and the Batch API gives a 50% discount on input and output. In the documentation's example, processing 10,000 support tickets of about 3,700 tokens per conversation on Haiku 4.5 comes to about 37 USD.
Prices and models change quickly: the same page lists retired models and notes that newer models use a tokenizer that produces about 30% more tokens for the same text. So check prices on the vendor's page before planning, and I do not quote prices of other vendors here because I did not verify them. A detailed saving technique is described in the post AI SaaS: caching and model routing.
What are the other costs?
Embedding storage. For semantic search you store and query vectors; the cost depends on the database and volume.
Latency. A call to an external model adds waiting time that depends on the model and request size. Measure it on your own case; synchronous AI features often need a different UX from asynchronous ones with a notification.
Evaluation and monitoring. Output quality has to be tracked continuously, and the tools and time for that are part of the cost.
Project scope. As a supporting feature (automatic tagging, note summaries) AI adds integration, error handling and fallback behaviour. As a core feature it changes the whole testing plan because the product loop depends on model reliability. I deliberately give no figures in weeks because they depend on the case; the terms are explained in SaaS MVP and MVP timeline.
What should you check before an AI feature goes to production?
Before release, check that the core product works without AI, that the feature has a fallback, that decisions can be explained, that accuracy is measurable and that the pipeline handles malformed real-world input.
- Does the product work without AI? AI should improve a product that already has users. Low engagement is rarely solved by adding an AI feature.
- Can the feature fail without harm? If the model is unavailable, the feature should degrade cleanly: show a cached result, an empty state or manual handling.
- Can users understand the decision? In sensitive areas (finance, health, hiring) the result must be explainable. Design explainability before you build.
- Can you measure accuracy? "It seems right" is not a metric. Define what good output looks like and how you will notice degradation when users or the model change.
- Is the data pipeline ready? A demo uses clean inputs, production gets malformed data and edge cases. Build validation and cleaning before the AI layer.
What do regulations require of AI features in the EU?
According to the European Commission's page on the AI framework, the AI Act entered into force on 1 August 2024, prohibited practices and AI literacy obligations apply from 2 February 2025, rules for general-purpose models from 2 August 2025, the main transparency rules from 2 August 2026 and high-risk systems in sensitive areas from 2 December 2027 (European Commission, AI Act). The same page also mentions the so-called AI Omnibus, so check the current deadlines before deciding. The Commission describes transparency as a requirement that people know when they are interacting with an AI system.
How do you prepare for model changes?
Models and prices change; the vendor pricing page I reviewed already lists retired models. So keep the AI call behind a single layer in your code, so that replacing a model does not require rewriting the product.
Your next step: take one of the three tasks from the start of this text, collect about twenty real examples and mark the correct answer by hand. That is your first accuracy test and it shows whether the task is suited to AI at all. If you want us to assess together what is worth building in your product, write to me.
Part of the AI-Augmented Development: Building Software Faster Without Cutting Corners guide.