---
title: Will Adding AI Make My Product Better?
url: https://varstatt.com/jurij/p/what-do-most-founders-get-wrong-about-ai-products
author: Jurij Tokarski
date: 2026-03-12
updated: 2026-10-26
description: Investors ask for an AI story, so the plan is a chatbot and an agent. AI adds value only on top of what you know; anyone can call the same model.
section: Blog (https://varstatt.com/jurij/archive)
tags: Q&A (https://varstatt.com/jurij/c/questions)
---

<Quote>
  We run a small SaaS for property managers. Investors keep asking about our AI story, so we're planning a chatbot and an agent. Will that make the product more valuable?
</Quote>

Only if it runs on what you know about the problem. A chatbot and an agent are features any competitor can add, because anyone can call the same model.

What nobody can copy is your knowledge of property managers: how they work, where they lose time. AI makes that knowledge faster. Without it, the AI story is a demo.

So the question I'd ask first isn't which AI to add. It's which job your knowledge already does better than anyone else's.

## Anyone Can Call the Same Model

AI is a tool, not a product. Adding it doesn't make a product more valuable by default.

A tender consultant with years of expertise uses AI to transcribe and analyze documents. The value isn't the AI. It's his workflow and expertise, encoded into the software.

He was winning tenders before AI existed, and AI lets him scale.

A platform for doctors' clinical notes works the same way. AI transcribes the recordings, and the product works because of deep knowledge of how doctors work.

For most products, AI is a supporting feature, not the core. Customers care whether the problem gets solved, not whether AI or cheap freelancers solved it. Property managers are no different.

[Worth Building](/principles/discovery/worth-building) asks whether the problem deserves a solution, whatever powers it. [Find the Core](/principles/discovery/find-the-core) asks which capability is the advantage: for most products, the domain knowledge. [Business Cost](/principles/philosophy/business-cost) is the reminder that technology for its own sake is self-indulgence.

That leaves what building the AI part takes.

## The Model Is Rented, the Work Isn't

A chatbot and an agent both run on hosted models from OpenAI, Anthropic or Google. The model is rented through an API, the same one any competitor can rent.

The work is everything around it: the prompt, the data the model reads, the screens, the failures and the bill. That's where your knowledge goes in.

Two titles name two different jobs here, and job posts mix them up. An ML engineer trains models, usually in Python, which fits when the product needs a model of its own. An AI developer, or LLM developer, ships features on the rented ones.

A chatbot and an agent are the second job: AI integration. You bring the domain knowledge, the provider brings the model, and the developer wires both into a workflow users already have. What both jobs cost is in [what a freelance developer charges](/jurij/p/what-a-freelance-web-developer-actually-charges).

Of the two features on this plan, the agent is the one to question first.

## An Agent Comes Last

One call to a model with good context does most jobs. An agent earns its place when the steps can't be written down in advance.

An agent is a model that can call tools, like your API or a function that sends an email. It decides on its own when to call them.

The model is rarely the hard part. In [Discovery](/discovery), the model could call a report tool that compiled a PDF and emailed it.

The prompt told it to always ask for the user's email first. In testing, it called the tool with `founder@example.com`.

Stronger wording in the prompt didn't hold. A check in the tool's own code did.

From then on, a placeholder address came back as an error, so the model asked for a real one. [Three bugs that were my prompts](/jurij/p/three-bugs-that-were-actually-my-prompts) has the code.

That's the rule I build agents by. A prompt asks; code enforces.

So before building this agent, write down the steps it would take. The steps that can be written become code, and the model handles the ones that need judgment.

Every tool call also carries data out of the product. That is the next thing to settle.

## Send the Model Less

Every prompt carries user data to a provider. So send it less, and decide where it may go.

In a product for property managers, that data is their records. Three decisions about it come before the first call goes out.

The first is where the key lives. I once reviewed an app with its OpenAI key in the React frontend, open to anyone with browser developer tools.

The fix needed [a backend for every AI call](/jurij/p/when-polish-over-security-costs-real), estimated at two to three weeks. The key belongs on the server, on your own provider account, like [every other key](/jurij/p/who-owns-the-code-my-developer-writes).

The second is what gets logged. A logged prompt is user data, kept as long as the logs are.

Before the cutover of [a rebuild I wrote up](/jurij/p/soapnoteai-v2-now-serves-100-percent-of-users), the hardening included an audit of what got logged. Prompts and transcripts had to stay where they belonged.

The third is where the data may go. On another feature, two providers filled a document form correctly, but neither could process it in the region the customer's policy required.

It [worked but couldn't ship](/jurij/p/it-works-you-cant-ship-it). Now I ask that question before I pick the approach.

With the data safe, one risk is left: the answer itself.

## Plan for the Wrong Answer

A person checks what goes out, and the code catches what fails. The model will be wrong sometimes, and the API reports success either way.

Wrong answers come in two kinds.

The first is wrong and plausible. In one pipeline, a comparison rate from an early step came back three steps later as the user's budget. [Labeling each piece of context](/jurij/p/when-llm-remembers-too-much) fixed it.

The second is wrong and silent. An empty reply arrived with HTTP 200, got saved, and [the chat never recovered](/jurij/p/silent-ai-response-corrupted-conversation-history). JSON cut off at the token limit threw no error and [surfaced three layers away](/jurij/p/200-ok-data-wrong).

The first guard sits in the code, for both kinds. An empty reply is never saved, and a cut-off reply raises an error. Failures are [monitored from the first deploy](/principles/diligence/monitor-day-one). A user sees an error message, not a blank answer.

The second guard sits in the product: the user decides. The note assistant in the rebuild I wrote up proposes each change as a diff, and the user accepts or undoes it. Earlier versions stay recoverable after the session ends.

That's where domain knowledge comes back in. The person who catches a wrong answer is the one who knows the work: the consultant, the doctor, the property manager. Build the workflow around their judgment, and the model only has to propose.

## The Company Owns the Chatbot's Answer

Who answers for a wrong one has already been tested. In February 2024, a Canadian tribunal ruled on the chatbot on Air Canada's website.

It had told a passenger he could claim a bereavement fare after the trip, within 90 days of buying the ticket. The airline's policy didn't allow that. Air Canada argued it couldn't be held liable for what its chatbot said.

The tribunal disagreed: "It should be obvious to Air Canada that it is responsible for all the information on its website." It ordered the airline to pay about C$812, as [Daily Hive reported](https://dailyhive.com/canada/air-canada-chatbot-decision) on February 15, 2024.

The sum was small, and the point wasn't. A wrong answer in a product for property managers goes out under the product's name, whichever model wrote it.

## How I Build an AI Feature

I build AI features on hosted models, in JavaScript and TypeScript, inside the team's own product and repo. [AI development for teams](/x/teams/ai-development) starts with a prototype on real data. Before the build, I price one request at the expected volume.

An AI agent development company sells an agent as a separate project. I build it as tasks on the team's board, and the team reviews every pull request.

A prototype that shows the feature isn't worth building is a result too: [saying no](/principles/discovery/saying-no) early costs less. The price is on the [homepage](/).

## What the Founder Should Do Now

Pick the one job your knowledge makes better, and build that before the chatbot.

Start with two questions. What do you know about property managers that a competitor doesn't? Which task in their week does the product make faster or better?

If the answers need the word AI, there's no product concept yet.

Then write that job down for a developer, with real examples of its inputs. [What to send a developer first](/jurij/p/what-makes-a-good-project-brief) covers the rest of the brief.

That job, built on real data, is the AI story investors are asking about. The chatbot and the agent can come after it, if users still want them.
