The short answer: a custom AI tool costs you twice. Once to build it, which depends on how many steps and systems it touches, and then every month to run it, which depends on how often it runs and which AI models it calls. For a small business automating one repetitive job, the monthly running cost is often smaller than people fear, but only if the build is scoped tightly and the right model is doing each step.
| What drives the cost | Build (paid once) | Running (paid monthly) |
|---|---|---|
| Number of steps in the job | More steps, more to design and test | More model calls per job |
| Systems it connects to (CRM, Sheets, WordPress, email) | Each integration adds work | Usually small, sometimes a subscription |
| How messy the input is (PDFs, scans, voice notes) | Clean-up logic takes time | Bigger inputs use more tokens |
| Type of output (text, images, video, voice) | Templates and checks to build | Media credits cost far more than text |
| How often it runs | Barely affects the build | The biggest lever on the monthly bill |
| Human review steps | A little extra screen work | Saves money by catching bad runs early |
Prices in this post were checked in October 2026.
Why do the prices online disagree so much?
Search "AI development cost" and you'll find ranges that start in the thousands and end somewhere near a house. They're not wrong, they're just answering a different question.
Debut Infotech puts a simple LLM API integration at $5,000 to $20,000 for a starter build, and an AI agent or multi-agent system at $50,000 to $120,000 before it's even a full build. Webclues starts entry-level chatbots at $10,000 to $50,000 and puts enterprise-grade platforms at $200,000 to $500,000 and up. Groovy Web is the closest to a small business view: $2,000 to $8,000 for a single automated workflow, $8,000 to $25,000 for a suite of connected ones, and $25,000 to $60,000 and up for custom agents.
Those big numbers mostly describe enterprise machine learning: training models, labelling data, compliance reviews, teams of engineers. Most small businesses don't need any of that. They need one job, the one that eats ten hours of someone's week, to happen on its own. That's a much smaller thing to build, because the hard part (the model) already exists and you rent it by the use.
What are you actually paying for: build cost or running cost?
Build cost is the project. Working out the job step by step, writing the prompts, connecting the systems, building a simple screen for your team, testing it on real examples until it's reliable, and handing it over. You pay it once, though most tools get a few rounds of tuning in the first weeks.
Running cost is what the tool spends every time it works. It usually has three parts:
- Model fees (tokens). Text models charge per token, roughly a chunk of a word, for what you send in and what comes back. Output costs more than input.
- Media credits. Image, video, voice and caption models are billed per image, per second of video, or per character of speech. These are usually the largest line on a creative pipeline.
- Hosting and services. Somewhere for the tool to live, a database, file storage, email sending. For a small tool this is often modest, and sometimes it's already covered by platforms you pay for.
Agency quotes tend to blur these together. Keep them apart. A tool that's cheap to build but expensive to run is a very different decision from one that's the other way round.
Why doesn't my ChatGPT or Claude subscription cover this?
This catches almost everyone. The monthly plan you pay for ChatGPT or Claude is for the chat app: you, typing, in their window. A custom tool talks to the model through the API, the developer side, and that's a separate account with its own bill.
OpenAI says it plainly: "API usage is billed separately from your ChatGPT subscription." Anthropic says the same for Claude: "A paid Claude subscription enhances your chat experience, and the Claude API is billed separately." The one wrinkle: Anthropic notes that its Max and Team plans include some monthly API credits. Useful for testing, but don't plan a business process around them.
So when you budget, assume the API bill is new money, paid by the use, in dollars.
What does it cost per run? A real example
We run a free tool on our own site, the Pipeline Sketch. You describe a repetitive job in plain words and it sketches the AI pipeline that could do it: the steps, which kind of model handles each one, and where a human should check the work.
We measured what it costs us. Each sketch costs about $0.05 in model fees. The full version, emailed as a report with a PDF, costs about $0.07. Because it's public and anyone can use it, it also has a hard daily spending cap, so a busy day or a bad actor can't run up the bill. That cap is something we'd build into any client tool that faces the public.
Five cents is cheap. But notice what makes it cheap: it's text in, text out, and it runs once per request. Change either of those and the maths changes.
Two client pipelines we've built show the range. The OptionsLab Shorts pipeline makes short videos, and every Short calls several paid models: one for images, one for video, one for the voice, one for captions. Media models are where the money goes, so the cost per Short is driven by how many scenes it has and how long it runs. The Creative Championship match tool sits at the other end: a WordPress tool that reads a Wyscout PDF and makes one Claude call to write the match report. One text call per match is about as lean as AI gets.
How do I estimate my monthly running cost?
You don't need a developer to get a rough number. Use this:
Monthly running cost = (runs per month × cost per run) + fixed monthly costs
Cost per run is the sum of every model call in the job. For text, that's tokens in and tokens out at the model's price. At the time of writing (October 2026), Anthropic's API pricing lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and Claude Haiku 5.5, a smaller and faster model, at $0.10 in and $0.50 out for prompts up to 100,000 tokens. Jobs that don't need an instant answer can go through Anthropic's Batch API at a 50 percent discount on both.
An illustration, not a quote. These numbers are made up to show the method. Say you want a tool that reads each supplier invoice, pulls out the totals, and drafts a reply. Each run sends about 5,000 tokens in and gets about 1,000 back.
- On Sonnet 5.5: 5,000 input tokens is $0.01, and 1,000 output tokens is $0.01. That's about $0.02 per invoice.
- At 500 invoices a month, that's about $10 a month in model fees.
- Add whatever your hosting and storage cost each month, and that's your running bill.
The same job on Haiku 5.5 would cost a fraction of that, if Haiku is good enough for the task. Often it is for extraction and sorting, and often it isn't for writing that customers will read. Picking the right model for each step is one of the biggest savings in any build.
Now swap the job for a video. One 30-second Short with images, video clips, a voiceover and captions can cost many times more than a thousand text calls, because media credits are priced per second and per image. If your idea involves video or voice, get those per-unit prices first. They'll decide whether the idea pays.
What makes a custom AI build cheaper?
Start with one process. The biggest cost driver is scope. One job, done reliably, is a small build. "Automate our marketing" is not a job, it's twenty. Pick the one that hurts most, prove it saves time, then add the next.
Keep a human checkpoint. A review step, where someone approves the output before it goes out, is cheaper than trying to make the AI perfect. It also stops a bad run from turning into twenty bad runs and twenty sets of model fees. We put human gates in our pipelines on purpose.
Reuse what you already pay for. If your team lives in Google Sheets, WordPress or WhatsApp, build the tool around them instead of replacing them. The Creative Championship tool lives inside the WordPress site the team already uses, so nobody had to learn a new system.
Clean inputs. Typed forms are cheap to process. Blurry scans and voice notes need extra steps. If you can fix the input at the source, the build and the running cost both drop.
Match the model to the step. Use a big model only where judgement matters, and a small one for sorting, extracting and formatting.
Paying from Nigeria: why are AI bills in dollars?
If you're in Nigeria, plan for this up front: the AI providers bill in US dollars. Model fees, media credits and most hosting are charged to a dollar card or account, every month. Your build might be agreed in naira, but the running cost will move with the exchange rate.
Three things help. Know your monthly running cost in dollars before you commit, so a rate swing is a number you've already seen. Set a hard spending cap on every API account, the same way we cap our own public tool. And ask who holds the API accounts: they should be in your name, so the tool and its costs belong to you.
When should you not build a custom AI tool?
Sometimes the honest answer is: buy the software.
If a SaaS product already does your job well, for a monthly fee, and your process looks like everyone else's, buy it. Meeting notes, a website chatbot that answers FAQs, scheduling social posts, transcribing calls: there are good tools for all of these, and a custom build would cost more and do less.
Build custom when the job is specific to how you work. When the input is your own data (match stats, product catalogues, legal templates), when the output has to look and sound like your brand, or when you've tried the off-the-shelf tools and your team is still copying and pasting between them.
Where do you start?
Every build we do starts the same way: a fixed-price Workflow Audit of one process. We map the job step by step, work out which parts AI should do and which should stay human, and estimate what it will cost to run each month, before you commit to a build. Pricing for the build itself is scoped per project after a free call.
If you want a feel for it first, describe your job in the Pipeline Sketch. It's free, takes a minute, and costs us about five cents. You can see more of what we've built on our works page, or read about our AI workflows service.
Or skip ahead and book a call. Bring the one job you'd most like to stop doing by hand.
