What it costs to add AI to an app you already have

$14,000 – $20,000 over 9–13 weeks for an assistant that answers from your own data, on an app that’s already live. The model calls are the cheap part. Getting your data to them is the work.

$14,000 – $20,000: An assistant answering from your own data
$15,000 – $22,000: That plus search that actually works
9–13 weeks: Start to shipped

Written by

WhatWillMyAppCost

We build the estimator this site runs on, and we design and build custom mobile and web software for a living. Every figure below is the estimator’s own output, except the subscription prices, which are read off each vendor’s pricing page and dated.

Your board asked what you’re doing about AI. A competitor shipped something last quarter and now it comes up in every meeting you have with them. And the quotes you’ve had back run from a few thousand to a few hundred thousand, with no obvious reason for the gap.

Here’s what it costs to add one to an app that already exists, why the range is that wide, what it costs every month afterward, and how to work out whether you need one at all.

What does adding AI to an app cost?

$14,000 – $20,000, over 9–13 weeks, for an assistant your customers can ask questions in plain English that answers from your documents and their account. That assumes the app is live, the screens are designed, and there’s a working API behind it.

$14,000 – $20,000: An assistant on an app that already exists
$15,000 – $22,000: With search rebuilt at the same time

If your API is partly built, or the data the assistant needs to read is spread across systems that don’t talk to each other, $18,000 – $26,000 is nearer the mark. That’s the usual reason these quotes vary so much, and it has nothing to do with the AI. Price your own with the description already filled in.

Why is the range so wide?

Almost none of the cost is the AI. Calling a model is a few lines of code. Any developer can demo one before lunch. Everything around it is what takes the time.

What the demo shows: A box you type into, an answer that comes back, and something that sounds convincing. Genuinely a morning’s work, and it looks finished.
What ships: Getting your documents and records into a form the model can read. Deciding what it’s allowed to see for each customer. Checking whether an answer is good enough to show before you show it. Handling the times it isn’t. And a way for somebody to see what it told people last week.

The one question that decides most of the cost: which of your data does it need to read, and where does that live right now? If the answer is one system somebody can query today, this is cheap. If it’s four systems and a shared drive, that’s the project, and the AI is the last week of it.

What does it cost to run every month?

Less than people fear. It grows with use, not features. You pay per question asked and per document processed. For most business apps that lands somewhere between a small monthly bill and a noticeable one.

Watch that it scales with your success. A feature ten people use costs almost nothing. The same feature at ten thousand users is a real line item, and unlike hosting it never flattens out. Ask whoever builds it to put a cap in from the start, so a bad week can’t produce a surprising invoice.

Get the estimate as cost per question instead of cost per month. Per month depends on assumptions about usage that nobody has yet. Per question you can multiply yourself once you know how often people actually use it.

Do you actually need one?

Only if your customers ask something your app can’t answer, and you can name it. Sounds obvious. It’s the step almost everybody skips.

Good reasons. People email support the same question forty times a week and the answer sits in a document they can’t find. People give up on your search because it only matches exact words. Somebody spends hours a week summarizing things a machine could summarize.

Bad reason: it should have AI in it. An assistant nobody asked for gets used twice, then quietly costs you money every month for as long as the app is live. If search is the real problem, fix search. That’s part of what the $15,000 – $22,000 figure above covers, and for a lot of apps it’s the whole answer.

Worth doing before you spend anything: read a month of your support inbox and count how many messages one good answer would have prevented. That number tells you whether this pays, and it takes an afternoon.

Adding an assistant to an app that already exists runs $14,000 – $20,000, and the cost is in your data, not the model.

Name the question it has to answer before you brief anybody. If you can’t name one, or if the honest answer is that search is broken, fix search and save the rest.

FAQ

Questions about adding AI

How much does it cost to add AI to an existing app?

$14,000 – $20,000 over 9–13 weeks for an assistant answering from your own data, on an app that’s live with a working API. $15,000 – $22,000 if search gets rebuilt at the same time. $18,000 – $26,000 if the data it needs is spread across systems that don’t talk to each other.

Why are the quotes I’m getting so different?

Almost always because of where your data lives, not because of the AI. A shop assuming one clean system and a shop assuming four systems and a shared drive are pricing completely different projects. Tell each of them exactly what the assistant has to read and the quotes will move toward each other.

What does it cost to run each month?

You pay per question asked and per document processed, so it grows with use, not with features. Small at first, and a real line item at scale. Ask for the figure as cost per question and put a cap in from day one.

Can we just use ChatGPT instead of building something?

For internal work, often yes, and it costs a fraction of this. The reason to build is when it has to answer from your own data, respect who’s allowed to see what, and live inside the app your customers already use. If none of those apply, buy a subscription.

What if it gives customers a wrong answer?

It will, occasionally, and how you handle that gets built in from the start. Common answers are showing the source it drew from, keeping it to topics you have documents for, and giving people an obvious way to reach a human. Decide which of those you want before the work is scoped.

Is our data used to train the model?

Not under the standard commercial terms of the major providers, which say inputs are not used for training. Read the terms of whichever one gets proposed, and if you handle health or financial records, raise that in the first conversation and not the last one.

Should we fix search first?

Frequently, yes. A lot of what looks like a case for an assistant is really a case for search that matches meaning instead of exact words. It costs less, it breaks less, and people already know how to use it.