Insights · 5 min read
Most companies don't need an AI transformation. They need one workflow that eats hours of someone's day. Here is what a real AI build includes and costs.
By GGP Editorial
Most companies do not need an "AI transformation." They need a specific workflow that currently eats hours of someone's time, and they want software to do part of it. When a client asks us about AI development services, the first conversation is almost always about that one workflow, not about models.
That framing saves a lot of money. Models are cheap to try. The expensive parts are the data, the evaluation, and the last-mile integration into a system people already use.
A real AI build has more in common with normal software engineering than people expect. The model is maybe twenty percent of the work.
These ranges are for a production feature, not a hackathon demo.
| Scope | What you get | Typical cost |
|---|---|---|
| Internal Q&A | Chat over your own docs and policies | $10,000 to $25,000 |
| Document processing | Extract fields from forms or invoices | $15,000 to $40,000 |
| Customer-facing assistant | Support or sales bot with guardrails and handoff | $25,000 to $80,000 |
The gap between internal and customer-facing is mostly risk. When the model is wrong in front of a customer, you need guardrails, a handoff path, and someone reviewing the failures. That is engineering time, and it is where cheap quotes quietly run out.
AI systems drift. Documents change, user questions change, and the model vendor changes their API. Budget for someone to read the failure logs every week and tune the prompts or the retrieval. A system with no owner degrades in a couple of months and becomes a liability.
We treat this as maintenance, not an afterthought, and we tell clients the honest number up front. A feature that works on day one and is wrong by month three is worse than no feature.
For a lot of workflows, the fastest path is calling an existing model through an API and wiring it into your product, not training anything from scratch. Fine-tuning or training a custom model is worth it only when you have a large, clean dataset and a problem where the off-the-shelf answer is clearly not good enough.
Most internal Q&A and document-extraction projects never need a custom model. They need better retrieval and a good evaluation set. Spending on a custom model before those are sorted is the classic way to overspend on AI.
In our own work, the AI projects that survive tend to share one shape: a narrow, repetitive task with a clear input and a clear output. Extracting fields from supplier invoices. Answering repeated support questions from a policy document. Flagging transactions that look out of pattern in a finance system.
These are not glamorous. They are just work that used to take a person an afternoon and now takes a few seconds. That is where AI earns its keep, and it is the kind of project we are happy to scope in an afternoon with a client.
One more thing worth asking: who owns the evaluation set? If the answer is nobody, the project will not end well. A partner who treats evaluation as a first-class deliverable is one who has shipped this before.
Ask any AI vendor to point at a system they shipped that is still running a year later, and ask what its error rate is. Most cannot answer the second question, and that tells you more than any slide deck.
We are a software company first, which means AI work sits inside real systems: finance tools, POS, CRM, ERP, IoT dashboards, mini programs. The model is one component among several, and the others have to keep working when the model is wrong.
We run a China-based engineering team with English and Portuguese communication and overlap hours for clients across the Americas, Europe, Africa, and Asia. If you have one workflow in mind, that is the right place to start the conversation.
Tell us what you are building and where you are today. We typically reply within 24 hours.
