What it is: adding AI features to the software you already have, such as summaries, search over your own help docs, a support assistant, drafting, or pulling data out of documents. Most of the work is connecting your app to an existing model from OpenAI, Anthropic or Google through its API, not building a model.
What it costs: two bills. First the build: development firms publish ranges of about $10,000 to $150,000 or more, and Upwork lists AI engineers at $35 to $60 an hour. Then the monthly model bill, charged per token: for the same feature it can be a few dollars or a few hundred, depending on the model.
How to choose: ask the four questions below, and ask for one small feature shipped first rather than an AI strategy.
AI Integration Services at a Glance
| How you buy it | How it is priced | Best for | Watch out for |
|---|---|---|---|
| AI consultancy or agency | Project quote, often $10,000 to $150,000+ | A large AI project with a fixed date and several specialists | Strategy and workshops billed before anything ships |
| Freelance AI engineer | Hourly, $35 to $60 on Upwork | One well-defined feature, if someone can review the code | You do the vetting, testing and managing |
| In-house hire | Salary, US software developer median $135,980 | AI at the heart of your product, for years | Months to hire, and a full salary before the first feature |
| Development subscription | Flat monthly fee, requests one at a time | AI features added to a live product, alongside the rest of your work | One request at a time, not a team working in parallel |
| Off-the-shelf AI tool | Monthly per seat or per use | A standard job, like a support chatbot on your help docs | It works on its tool's data, not inside your product |
Before you hire anyone, ask these four questions:
- Question 1, "Which feature would you ship first, and how will we know it works?" A good answer names one narrow feature and how success is measured. A vague answer about "an AI strategy" is a warning.
- Question 2, "How will you test the answers?" You want a set of real examples from your product, checked before launch and again after every change to the prompt or the model.
- Question 3, "What stops a user from tricking it, and where does our customers' data go?" OWASP ranks prompt injection as the top risk for these apps. Expect limits on what the model can read and do, and a clear answer on the provider's data terms.
- Question 4, "What will it cost to run each month?" Ask for an estimate of the model bill at your usage, and how they will keep it there.
Want AI features shipped into the product you already have?
A senior developer on subscription: $2,995 a month, as many requests as you like, each shipped within 2 business days, and your money back if the first 14 days don’t prove it.
In this article
1.
2.
3.
4.
5.
6.
7.
8.
9.
This guide is for SaaS founders adding AI to a product that already exists. For choosing a provider for any kind of software work, see how to choose a software development company. If an AI tool wrote your app and it now keeps breaking, start with vibe coding cleanup.
What AI Integration Services Include
Most AI features in a SaaS product follow the same pattern: your app collects the right text or data, sends it to a model with instructions, checks what comes back, and shows it to the user. A full service covers each step.
Picking the feature
Choosing one job the AI should do well, for one kind of user. The best first features are narrow and already asked for, like summarizing long threads or answering questions from your help docs.
Connecting a model
Calling a model from a provider such as OpenAI, Anthropic or Google through its API, from your server. This includes choosing the model, handling errors and timeouts, and keeping the API keys secret.
Feeding it your data
Models know nothing about your product or your customers. To answer from your content, the app searches your documents or database for the relevant pieces and sends them along with the question. You will hear this called retrieval-augmented generation, or RAG.
Instructions and limits
The written instructions (the prompt), the format the answer must come back in, and limits on what the model can see and do inside your app.
Testing the answers
A set of real examples with good answers, run before launch and after every change. AI output varies, so this replaces the usual "it worked when I tried it."
Shipping and watching it
Releasing it to some users first, logging what it does, tracking the model bill, and improving the prompt on real usage.
Common AI Features in SaaS Products
| Feature | What it does | How big a job |
|---|---|---|
| Summaries | Turns long threads, calls or documents into a short summary | Small: one model call on text you already store |
| Drafting | Writes a first draft of a reply, description or report for the user to edit | Small to medium |
| Sorting and tagging | Classifies tickets, leads or records and routes them | Small to medium, with good test examples |
| Pulling data out of documents | Reads invoices, forms or emails and fills in fields | Medium: accuracy needs careful testing |
| Search and answers over your docs | Answers questions from your help center or the customer's own data | Medium to large: needs search over your content |
| Assistant or agent | Chats with users and takes actions in your app for them | Large: the most testing and the strictest limits |
Start at the top of this table. A small feature that ships teaches you more about what your customers want from AI than a large one that is still being planned.
What AI Integration Costs
AI features cost money twice: to build, and every month they run.
The build. Published prices come mostly from firms selling the work, so treat them as rough. Biz4Group, a development firm, puts AI integration into existing software at $10,000 to $150,000 or more. Rapid Dev, another, puts a simple chatbot integration at $15,000 to $35,000 over 4 to 8 weeks. Upwork lists AI engineers at $35 to $60 an hour. Hiring in-house means a salary first: the US Bureau of Labor Statistics puts the median software developer's pay at $135,980 a year, before benefits.
The running cost. AI providers charge by the token, a small chunk of text. OpenAI's rule of thumb is that 100 tokens is about 75 words of English. You pay for the text you send and, at a higher rate, the text the model writes back.
Here is what one feature costs to run on Anthropic's current list prices. Say each summary reads a 2,000 word support thread (about 2,700 tokens) and writes 200 words back (about 270 tokens).
| Model | List price per million tokens (in / out) | Per summary | 10,000 summaries a month |
|---|---|---|---|
| Claude Haiku 5.5 | $0.10 / $0.50 | about $0.0004 | about $4 |
| Claude Sonnet 5.5 | $2 / $10 | about $0.008 | about $80 |
| Claude Fable 5.1 | $10 / $50 | about $0.04 | about $400 |
The figures are our own arithmetic from Anthropic's published prices. Your numbers will depend on the length of the text and how often the feature is used, but the lesson holds: the model you pick can change the running cost a hundredfold for the same feature, so test the cheaper models first. Anthropic also takes 50% off for work that can wait and run in batches.
Why AI Projects Stall
Nearly everyone is trying AI, and far fewer have it working at scale. Gartner named the reasons it expected projects to be dropped: poor data quality, weak risk controls, rising costs and unclear business value. Those are large-company figures, but the same four things stall a small SaaS feature:
- Unclear value. "Add AI" is not a feature. "Summarize every support thread before the customer calls" is.
- Messy data. If your help docs are out of date, an assistant built on them will be confidently wrong.
- No limits. A feature that can read anything and do anything is a security and support problem waiting to happen.
- No cost check. A demo on the most capable model can be fine at ten users and painful at ten thousand.
The Risks to Check Before You Launch
- Prompt injection: the model only sees the data the current user is allowed to see, and cannot take actions they could not take themselves.
- Customer data: you have read the AI provider's data terms, and your own customer contracts allow sending that data to it.
- API keys live on your server, never in the browser or the mobile app.
- Wrong answers: the feature was tested on real examples, and users can see where an answer came from or edit it before it is used.
- Costs: a monthly spending limit is set, and usage is logged per customer.
- A fallback: the rest of the app keeps working if the AI provider is slow or down.
On data, OpenAI says it does not use data sent to its API to train its models by default. That answers one question, not all of them, so still check how long the provider keeps the data and what your customers agreed to.
How to Choose an AI Integration Provider
You get a team and a project manager. Check how much of the quote is strategy and workshops, and how soon something real ships.
Cheaper per hour, but you do the vetting, testing and managing. See agency vs freelancer for the trade.
The most control, at a full salary and months of hiring. See how to hire a software developer.
A flat monthly fee for requests worked one at a time. An AI feature is one request, or a few, in the same queue as everything else. See developer as a service.
Whoever you hire, a fixed price for an AI feature is hard to set, because nobody knows how good the answers will be until they are tested on your data. Our guide to fixed price software development covers why, and software development contracts covers getting the code assigned to your company in writing.
Red Flags in an AI Integration Proposal
- A strategy phase with nothing shipped. Weeks of workshops before any feature reaches a user.
- Training a custom model first. For most SaaS features, an existing model plus your own data is faster and cheaper.
- No plan for testing answers. "We'll try it and see" is not a test.
- No estimate of the monthly model bill. You find out what it costs when the invoice arrives.
- API keys or prompts in the browser. Anyone can copy them, and run up your bill.
- The code and the AI provider account in their name. Both should belong to your company from the first day.
How StackTie Fits
StackTie is a development subscription. You send requests, they are worked one at a time, and each one comes back in days as a pull request with a plain-English summary of what changed. An AI feature is a request like any other: we pick the narrowest version worth shipping, connect it to a model, test it on real examples from your product, and keep the model account and API keys in your name. Bigger AI work, like an assistant that takes actions in your app, is split into smaller requests that ship one after another. One senior developer, Michael, leads every build, and everything we build belongs to your company and lives in your own accounts. There is a money-back guarantee if it is not a fit. It is not the right choice if you need native iOS or Android apps or a whole team working in parallel. The price and what is included are on the homepage.
