"Can we add AI to it?" is now the most common line in the web app briefs we receive. Sometimes it's the right call. Often the client isn't sure what they want the AI to do, only that their competitors seem to have it.
We built AI features into five client apps this year. Two of them are used every single day. One was quietly switched off after six weeks, and that one taught us the most. It was a "smart assistant" for a distributor's dealer portal, and dealers simply didn't trust it. They kept phoning the sales team anyway. What they actually wanted was a faster way to check stock, and that didn't need AI at all.
Here's the practical framework we've used since.
Start with a boring problem
The best AI features we've shipped solve dull, repetitive tasks:
- Summarising long text. Support tickets, meeting notes, customer feedback.
- Drafting replies that a person then reviews and sends.
- Classifying and routing incoming enquiries to the right team.
- Searching your own content in plain language, like an FAQ or a document library.
- Extracting data from uploaded invoices, forms or PDFs.
The worst ones try to be impressive: a chatbot that "knows everything about the company" with no clear job. Our favourite feature this year is almost embarrassingly simple. For a logistics client's support desk, the app writes a three-line summary at the top of every long email thread. Agents told us it saves them "the first two minutes of every ticket". Nobody calls it AI. They just call it "the summary".
A good test: can you describe the feature in one sentence without the word "AI"? "Suggest a reply to each support ticket" passes. "Add an AI assistant" doesn't.
Use your own data, carefully
Most useful business features need the AI to work with your content: your products, your policies, your past tickets. The common approach is retrieval: when a user asks something, the app first finds the most relevant documents in your own database, then gives only those to the model along with the question. This keeps answers grounded in your actual information and makes it much easier to show sources.
It also means the quality of your answers depends on the quality of your content. Outdated FAQs produce outdated answers.
Running costs are real, but manageable
AI features are billed per use, usually by the amount of text processed. For most of our small and mid-sized clients, monthly AI costs for features like summaries and search land somewhere between a few hundred and a few thousand rupees. But costs can spike if you're careless. We always:
- Set a hard monthly spending limit with the AI provider.
- Use smaller, cheaper models for simple tasks like classification.
- Cache results so the same document isn't summarised twice.
- Limit how much text each request can send.
- Log usage per feature so we can see what's actually being used.
Privacy and the DPDP Act
If your app sends personal data to an AI provider, you need to think about consent, data storage and where the data is processed. Check the provider's data retention and training policies, choose business plans that don't train on your data, and avoid sending personal details the task doesn't need. Mention AI processing in your privacy policy in plain language.
Keep a human in the loop
For anything customer-facing, we design AI features to suggest, not decide. The AI drafts the reply; a person clicks send. The AI tags the ticket; a person can change the tag. This catches mistakes, builds trust with your team, and gives you real feedback on how well the feature works. In one app, we track how often agents edit the suggested reply before sending it. When that number started climbing, it told us the product information the AI was reading had gone out of date, long before any customer complained.
Measure it
Before building, agree on what success looks like. Minutes saved per ticket, faster response times, fewer "where is my order" emails. If after two months the numbers haven't moved, it's fine to switch the feature off. The best AI feature is one people actually use, not one that looks good in a demo.
