AI automation services
AI automation services for websites and web apps
One job at a time is how I add AI to a website or product, such as sorting enquiries or drafting replies that a person checks.
- The task in writing: inputs, output, what a good answer looks like, what happens on a wrong one
- Claude, Gemini or OpenAI connected through your own account and API keys
- A second model on standby and a monthly spending limit
- Prompts, a set of real examples to test against, and a review step
- Logs of every input and every output
- Deployment, monitoring, a written runbook, and fixes for anything that breaks soon after launch
Try it
One job, through the pipeline. Try knocking a model out.
Pick a job and watch it run. If the first provider is slow, the backup takes over. The meter shows your monthly limit.
- 01 · input412 findings from a crawl
- 02 · model
- Claude
- Gemini
- OpenAI
- 03 · reviewinternal → logged, spot-checked monthlyauto, with a log
- 04 · output…
The counts and the meter are illustrative. What's real is the shape: a written job, a backup model, a review step where it matters, and a cost limit you set.
Overview
Where AI automation services fit
AI automation services, as I build them, give a model one task inside your product. Say you send owners a long SEO audit they can't follow. The task is to turn it into a short fix list in plain English. Or your inbox fills with the same questions every week. Then the task is to sort each enquiry and draft an answer.
- AI integration for websites and web apps, called from your server
- Office jobs that come round every day, like sorting enquiries or summarising reports
- Your own account and API keys, so the bill and the data stay with you
- Logs of every request and every answer
What a scope from me looks like
Sample scope for adding AI to a live site
Included
- The task in writing: inputs, output, what a good answer looks like, what happens on a wrong one
- Claude, Gemini or OpenAI connected through your own account and API keys
- A second model on standby and a monthly spending limit
- Prompts, a set of real examples to test against, and a review step
- Logs of every input and every output
- Deployment, monitoring, a written runbook, and fixes for anything that breaks soon after launch
Phase two
- A second task, once the first one has been measured
- The model reading your own documents before it answers
- Fewer human reviews, once the error rate is known
Not doing
- A chatbot with no task behind it
- Replacing a person your customers need to reach
- Sending your customers' data for model training without their consent
Most lines in this scope are about wrong answers. Each one says what a good answer looks like, and what happens when the model gives a bad one.
Process
How a feature gets built, step by step
Start collecting real examples early. From step two onwards, everything is tested against them.
- 01
Discovery call
We talk through what the AI should do, who it's for and what a wrong answer would cost you. Two or three real examples help, even rough ones.- You
- Describe the job in one sentence.
- Get
- Notes on the job, the people it serves and the cost of a mistake
- 02
Strategy & Wireframes
Your description becomes a written job, with a line for every input and every output. You sign it off before anything is wired up.- You
- Send real examples of the inputs, including the messy ones.
- Get
- A written scope naming the job, the model and the monthly limit
- 03
Design & Develop
The model is called from your server through your account, never from the browser. Prompts are tested on your examples. The backup model and the cost limit are wired in before anything else.- You
- Compare the outputs with your examples and mark the wrong ones.
- Get
- A staging version you can try with real inputs
- 04
Review & Revisions
You use it on real work. I adjust the prompts and the review step until the error rate is one you are willing to sign off.- You
- Flag every bad answer you see.
- Get
- An error rate measured on your own work
- 05
Launch & Handover
The feature goes live behind the cost limit. The logs, the runbook and the API keys stay in your name.- You
- Set the monthly limit and give the final sign-off.
- Get
- The live feature, its logs, the runbook, and my fixes if it misbehaves early on
- 06ongoing
SEO & Support
Providers update their models and change their prices. A prompt that worked at launch can start giving weaker answers after an update. So after each update I re-run your test examples, and every month I read the outputs and the bill.- You
- Tell me who reads the monthly figures on your side.
- Get
- Test scores after each model update, and the month's spend
- Discovery callboth
- Strategy & Wireframes
- Design & Develop
- Review & Revisions
- Launch & Handoverboth
- SEO & Support
Tech stack
AI features for a web app — the layers underneath
The model is the smallest part of the build. Most of the hours go into what surrounds it.
| CRITERION | What I use | Why |
|---|---|---|
| Models | Claude, Gemini and OpenAI, through their APIs | A second model stays ready if the first is slow or fails |
| Your API | Next.js route handlers, or Node and Express | The model is called from your server with your keys, so nothing secret sits in the browser |
| Data | MongoDB, plus your own documents when the job needs them | Answers come from your content rather than from whatever the model remembers |
| Cost | A budget per job and a monthly cap | Each job has its own budget, so one busy feature can't use up the month for the others |
| Testing | Real inputs, each with the answer you'd accept | A prompt change is measured against them before it ships |
| Review | An approval queue in your admin, with edit and send buttons | Each edit is logged next to the model's draft, so you can see what got corrected |
Why it matters
When the model gets it wrong
A model can be wrong and still sound completely sure of itself. It writes a smooth paragraph either way, so reading the tone won't catch the mistakes. Before launch I run it on your own examples, count how often it's wrong, and show you that number.
That number decides how much checking the feature needs. It goes into the scope. You sign it off before launch.
Your keys, your bill
The API keys are opened in your account, not mine. So the provider bills you directly, and your customers' text goes through an account you control.
When it's running
A normal day with the feature switched on
Enquiries arrive, and the model sorts them before you open the inbox. A question it can answer gets a draft reply, and a person reads it before it goes out. If the provider has a slow day, the backup model takes over. When the month's limit is reached, the feature stops and tells you.
- 900+ projects
- since 2017
- 4.9★
- 400+ five-star reviews on Fiverr
- 9+ years
- building WordPress and Next.js sites
Support is the whole team, not one person — reach us any time, and someone picks it up.
“The freelancer demonstrated excellent problem-solving abilities, tackling challenges head-on and coming up with innovative solutions, maintains clear, concise, and prompt communication. Any questions or concerns I had were addressed quickly and professionally.”
Schedule a Meeting
Describe the job you have in mind in one sentence. I'll tell you on the call whether it needs a model at all.
A real office, real people
Led by me, built with my team — in Kolkata, India.
The team: three developers, a designer and three SEO specialists in the Kolkata, India office, plus around six people working remotely. Where the model's output reaches your customer, a person reads it first — and every feature has a cost cap.
Sandipan Das
You talk to me; I’m the one accountable for the work.
- Availability
- The team works on your schedule, in your time zone.
- Support
- Support is the whole team, not one person — reach us any time, and someone picks it up.
Pricing
Two costs: the build and the model's monthly bill
An AI feature has two costs, and both are written down before work starts. The build follows the scope, line by line. The monthly model bill belongs to the provider, not to me. You set its limit, and you can read every call in the logs.
The test examples are yours to collect. For each real input, you write down the answer you'd accept. I can't write those for you, because I don't know what your customer should be told. That set is what lets you trust the feature with real customers.
For how those lines turn into a price, see the quote page. The feature itself normally sits inside a Next.js app or behind an API. If the whole site should be drafted with AI and then reviewed, that is AI website development.
- Build cost: one line for each part of the scope
- Model cost: monthly, capped by you, visible in the logs
- Test examples: your time, and the part most worth spending it on
Next step
One sentence is enough to start
Write down what the AI should do, and for whom. Add two or three real examples, the messy ones included, and we start from those.
Which model would you use — Claude, Gemini or ChatGPT?
- Different jobs suit different models. Long, careful writing suits one; quick, cheap sorting suits another. So I try two or three on your own examples and keep the one that gets the fewest wrong. Claude, Gemini and OpenAI all connect from your server with your own keys, and a second one stays ready for the day the first provider is slow.
What will it cost me every month?
- You pay the model provider directly, and the amount grows with how often the task runs. Because the usage is yours, I don't put a figure on it. You set a monthly limit, and the logs show every call, so the bill doesn't arrive as a surprise.
Can it answer my customers by itself?
- For low-stakes questions, such as opening hours or where to find a form, the model can reply on its own. Questions about money or a complaint go to your staff as a draft. The scope lists which questions fall on which side, using the error rate from testing.
Will my customers' data be used to train the models?
- Your data goes to the provider only to do the task, through an account in your name. The scope names the provider and the data setting it uses. I never send your customers' data for model training without their consent.
What if the AI gets something wrong?
- When a model gets an answer wrong, one of three things happens, and the scope says which. A person can take over. A safe default answer can be shown instead. Or a second model gets the same question. Every case lands in the log, so you can see how often each one happened.
Published · Updated