Add AI to your product when it makes a task your users already do faster or better, using your own data, and the result feeds straight into something the product does. Skip it when a normal feature, like a form, a search box, or a status page, would do the job more reliably. "Add AI" on its own isn't a feature.
Pressure to add AI usually comes from outside: a competitor launched something, an investor asked, or a customer mentioned ChatGPT. Those are fair prompts, but they don't tell you what to build. The AI features people actually use remove a slow, repetitive step from work they already do. The ones they ignore are usually a chat box bolted on because it was expected. Below: how to decide, three tests, features that tend to work, when a chatbot is the wrong choice, the risks, and how to add AI to a live product. For prices, see what an AI feature costs.
AI or a normal feature: the short answer
AI helps when
The task is slow thinking work
- Users spend real time writing, reading, searching, or sorting
- The answer has to come from your own data or documents
- The result feeds straight into something the product already does
- An occasional imperfect answer is acceptable, or a person checks it
A normal feature is better when
The task has a right answer
- The task is picking dates, quantities, or options
- People want to find a specific page or record
- The answer is a fact your database already knows
- Every answer must be exactly right, every time
The cut
AI is good at language and judgement: reading, writing, summarizing, sorting, and matching. It's poor at exact facts it wasn't given. A database answers "where is my order?" perfectly; an AI model guesses. Use each for what it's good at.
Three tests an AI feature should pass
1. You can name the task without saying AI
"Draft a reply to this customer from their order history" is a task. "An AI assistant" isn't. If you can name who does the task today and how long it takes them, you can measure whether the AI helps.
2. It needs your data
A general model already knows the internet. It doesn't know your products, policies, or customers. Features that combine the model with your own data are the ones competitors can't copy by opening ChatGPT.
3. The result leads to an action
If users copy the AI's answer into another tool, the feature stops halfway. The best AI features end in something the product already does: a reply sent, a record updated, a request routed.
AI features that usually work well
| Feature | What it replaces |
|---|---|
| Answers from your own content | Staff answering the same questions by email |
| Summaries of long records | Reading a full history before every call |
| Drafted replies and messages | Writing each one from scratch |
| Sorting and tagging incoming requests | Someone triaging an inbox by hand |
| Pulling details out of documents | Copying fields from PDFs into forms |
| Personalized content at scale | Writing one version per customer |
The Chat Company is the first row turned into a product: businesses get a chatbot that answers from their own website, now handling 100K+ messages a day. On a confidential B2B sales platform I worked on, AI sits inside sales workflows, helping personalize outreach rather than living in a separate chat window.
When a chatbot is the wrong AI feature
Chat is one way to use AI, and often not the best one. It's the wrong choice when:
- The task is a form. Dates, quantities, and options are faster to pick on a screen than to type in a conversation.
- People want to find a page. A good search box, with typed results and links, is quicker and cheaper than a bot.
- The answer is a status."Where is my order?" is a database lookup. A model may phrase it well and still get a detail wrong.
In these cases, AI works better behind the scenes, for example by suggesting the right option or sorting results, while the screen stays simple. Users rarely ask for a new way to use your product. They want the existing one to take less effort.
Risks to plan for when you add AI
- Wrong answers.Models sometimes state things confidently that aren't true. Showing the source, and keeping a person in the loop for anything important, limits the damage.
- Data leaking between customers.In a product with many customers, each one's data must stay separate. On The Chat Company, every lookup is filtered by business for exactly this reason.
- A growing bill. Usage-based pricing grows with use. Caps and tracking keep it predictable.
- Provider changes. Models are updated and retired, so the feature should be able to switch provider easily.
None of these are reasons to avoid AI. They're reasons to build it properly rather than as a quick experiment left running in front of customers.
How to add AI to a product that's already live
- Pick one slow task. The one users spend the most time on, or complain about most.
- Measure it today. How long it takes, and how often it happens, so you can prove the AI helps.
- Test on real data first. A quick prototype with your own records shows whether answers are good enough.
- Launch to a small group. With a way to report bad answers, and a fallback if the AI fails.
- Measure again, then expand. Keep it if it saves time. Remove it if nobody uses it.
This keeps the product live and stable the whole time, and means you only pay for AI that earns its place.
How to tell if an AI feature is working
An AI feature can look impressive in a demo and still go unused. Decide how you'll judge it before launch, using numbers you can already see:
- Use. What share of the people who could use it do, week after week, not just in the first few days.
- Time saved. How long the task took before, and how long it takes now.
- Acceptance.How often users keep the AI's draft or answer, versus rewriting or ignoring it.
- Cost per use. The monthly AI bill divided by how often the feature is used.
If use drops after the first month, or most answers get rewritten, the feature needs better data or a different design, or it should go. Removing an AI feature nobody uses is a good decision, not a failure.
Should you charge extra for AI features?
If you run a SaaS product, AI adds a cost that grows with every use, so it's worth deciding how to cover it. There are three common approaches:
- Include it in every plan, with sensible limits. Simple, and best when the AI makes the core product better for everyone.
- Put it in a higher tier. Works when AI mainly helps heavier users or larger teams.
- Charge by usage, often as credits. Fair when some customers use far more than others, but harder to explain.
Whichever you choose, track AI use per customer from the start, as The Chat Company does, so you know what each customer costs you before you set the price.
When AI is the wrong next step
Sometimes the honest answer is "not yet":
- There's nothing to work from. A chatbot over five thin website pages has nothing useful to say. Improve the content first. Write the answers customers ask for most, and the AI will have something to work from later.
- Nobody can name the task.If the brief is only "add AI", the feature will be a demo with a monthly bill. Spend a week noting which tasks slow your users down, and come back with one.
- There's no product yet. Launch the core job first. See what to put in a first version.
A checklist before you build an AI feature
See how I add AI to products on my AI features page, or send me your checklist, and I'll tell you whether AI is the right tool for the job.
