I skipped the totals and read the scope. How much does it cost to add AI to an existing product? One PDF said integration, $12,000, and the only line under it was ChatGPT. The other said platform, $180,000, and week two was a bar called agent orchestration. People already logged in. Neither scope said what they would finish this week that they can't finish now.
I build this as AI-powered products, on SaaS and web products that already run. Whether the job belongs at all is when AI belongs in a product. A generated trial you want made real is an AI-generated app or a developer. A product that doesn't exist yet is how much a web product costs. This page is the feature on top of one that already has users.
How much an AI feature costs to add
Dollars, from briefs in 2026 where the product was already up. I price the feature you can ship, and the bill that shows up after people use it. A quote for a whole new product is the last row. That fee is on the other page.
| What they asked for | What it costs | What the fee covers |
|---|---|---|
| One call on a screen you have | $5,000 to $20,000 | A button on a screen that already exists. The answer drops into a field. I've shipped that in a week, and I've taken three when the field has rules the model keeps breaking. About $50 to $500 a month until a lot of people hit it. The $12,000 PDF was this row. Under the fee, one bullet said knowledge base, and no page was attached. |
| Search over your own documents | $15,000 to $60,000 | The answer has to come from your docs, and the screen has to show which page. People call that RAG. Four to ten weeks when someone can already point at the page. About $200 to $2,000 a month, more once the index is large. Their Drive had 140 files. Half were named Final. I don't quote the cheap end of this row until those duplicates are gone. |
| An assistant inside the product | $20,000 to $80,000 | Chat on a page a logged-in person already uses, and it knows which account is talking. Six to fourteen weeks. About $500 to $3,000 a month. If nobody has named an action it can take, you bought the first row and a text box. |
| An agent that can change a record | $40,000 to $150,000+ | It changes a record. Files the ticket, or edits the order, and some nights it will do the wrong one. Eight weeks when the action is one. Past twenty when it can move money. The monthly number is the retries, and I've seen that land anywhere from $1,000 to $10,000. The $180,000 PDF was this row with an admin copied on top. |
| A new product sold as a feature | The other receipt | Login, records, the job the company is. That fee is how much a first version costs. I leave it off this table. A button on a product people already open is the rows above. |
An agency on the same row is often twice the number. You pay for a room. That gap is a freelance developer or an agency. The model name barely shows up in it. Whether I can deploy does. If someone else holds the keys, you're in taking over a product without a rewrite, and I price that before I price the prompt.
Week two on the $180,000 PDF was a bar labeled agent orchestration. The product already changed that status with a dropdown.
What moves an AI feature into another price
I write a brief before I quote. A day rate doesn't tell you which row you're on. Fixed price or hourly waits until I can name the screen. Before that, the deposit is on a logo.
The policy still lives in five people's heads
Search is the second row when someone can point at the page that already says the answer. If it lives in five heads and a shared drive from 2022, the first weeks are me reading that drive. I'll say so on the call. If the chat is already on the homepage in week six and I still haven't found the policy, I guessed, and you paid for the guess. Training your own model is off this table. For a first feature I put your pages in front of a model you pay for, and I leave a way to change vendor without rebuilding the screen.
The new verb is refund
Refund is the verb that shows up in the brief. The moment it can change a record, you left the assistant row. A paragraph someone copies into the admin is a demo. I need a switch that turns the feature off without taking the product down, and a person on your side who actually reads what it did. Who types, and who holds the keys, is what to ask before you hire. The $12,000 quote had none of that and still promised the agent. The second invoice is where that promise gets paid.
ChatGPT is the biggest word on the slide
One customer in the database is a different build from a product where customer A must not see customer B's tickets. If a regulator already reads the product, I price the log of what the model saw, and where that file lives. I don't bury those hours in a line called integration. A quote that can't name the screen is a logo.
What the AI feature costs every month
The table is what I charge to build it. The model bill shows up every month after that, and it should be in an account you can open. Before anyone approves the work I write two figures from the traffic you already have. What a quiet Tuesday costs. What it costs the week the agent tries the same job again because the first answer was wrong. I don't paste a vendor rate card into the proposal and call that a forecast. Those rates move. Retries move the bill more.
The first row stays in that $50 to $500 until a lot of people hit the button. That's what I see on products that are already up. Vendors change that price. Search costs more as you add pages, because the index grows. An assistant that asks the model to rewrite your shipping policy on every message pays for that paragraph every time, so I cache the policy. On one agent the statement was a bit over $6,000, mostly retries. The build had been $40,000. Nobody had capped what one account could spend in a session. I put the cap in before that feature goes on for anyone else.
The API key sits in an account you can open. I don't resell the tokens. If the only person who can see that invoice is me, you don't own the feature. That handoff is who owns the product after launch. I'll cut the feature until the monthly number is one you can read on the statement.
When I won't quote an AI feature
Some of these I send back. The number is real. The job is one I wouldn't take.
Nobody can name who it's for
A competitor shipped a bubble. Or an investor asked where the AI was. If you can't name who is worse off without it, I don't price it. That call is whether AI belongs in the product. A filter and a list often finish the same job. I watched a team pay a model to sort a column the table already sorted.
Five pages and a shared drive
Pointing a model at five marketing pages doesn't make a knowledge base. Write the pages, then ask me again. And if nobody on your side will look when a customer pastes a sentence your policy page never said, leave the feature off. The cheap quote stays cheap until that screenshot lands in your support inbox.
There is no login yet
Nobody comes back tomorrow. This table is the wrong page. A first version is how much a web product costs, and what belongs in that first version is the cut before any model. A site a stranger visits once is how much a custom website costs. Store listings are how much a mobile app costs. I won't put an agent on a brochure so a deck can say AI.
When I do take it, I ship one screen behind a switch. Wrong answers, you turn the switch off, and the product they used yesterday is still there. A few of the live ones are in selected work. Send the PDFs and the URL if you want me to mark them, same as any other quote. I want the sentence that wasn't in either scope. Who opens it, and what they finish that they couldn't finish last week. More of AI feature cost is that argument on other briefs.
