Insights · 12 min read
Generative AI development explained: which use cases pay back first, what a project costs, and how to start without burning the budget.
By GGP Editorial
When a company calls us about generative AI, they usually say the same thing in different words: we want the thing ChatGPT does, but for our business. That sentence hides the real questions, which are what to build, what it costs, and whether it will pay for itself.
Generative AI means software that creates text, images, voice, or code from a prompt rather than just classifying or predicting. It runs on large language models and other foundation models. Building on those models is its own discipline, with a cost structure that looks nothing like the predictive machine learning projects companies ran five years ago.
I write from the vendor side. GlobeSoft is a software company founded in 2018 with 40-plus engineers and more than 300 delivered projects, and a growing share of that work is generative AI: retrieval systems over client documents, internal copilots, voice agents, and AI features inside FinTech and property products. This guide covers what generative AI development involves, where it returns money first, what it costs, and how to start without burning the budget.
Generative AI development is building software around a foundation model so the model can do useful work on your data. The model is one component. The product is everything around it: where the model sits, how your data reaches it, how its output is checked, and how a person actually uses the result.
Three examples make it concrete.
A support copilot that drafts a reply to a ticket, reading the customer's history from your database. A document assistant that answers questions against your internal policies and cites the page. A property platform that lets a manager type "show me units with rent overdue more than 30 days" and turns that sentence into a database query.
In each case the model writes or acts, but the value comes from your data and your workflow, not from the model's training data. That distinction drives every cost decision that follows.
Generative AI is also not the same as the predictive AI companies used to buy. Predictive machine learning answers questions like "is this transaction fraud" or "what will demand be next month." Generative AI answers "write this" and "do this." The two share little in terms of architecture, data needs, and testing, so do not let a vendor sell you one as the other.
The use cases that pay back fastest share one trait: they remove repetitive reading and writing from a workflow that already exists. They do not invent a new workflow, because new workflows need adoption, and adoption is slow.
| Use case | What the model does | Pattern | Typical effort |
|---|---|---|---|
| Document intake and extraction | Reads invoices, claims, contracts, pulls fields | Prompted API or RAG | Low to medium |
| Internal knowledge assistant | Answers questions over policies and docs | RAG | Medium |
| Customer support copilot | Drafts replies, summarizes tickets | RAG | Medium |
| Code and testing assistant | Writes and reviews code | Tool calling / agent | Medium |
| Voice agent | Handles calls or orders end to end | Agent plus voice | High |
| Content and marketing drafts | Generates copy from brand guidelines | Prompted API, sometimes fine-tuning | Low to medium |
Document-heavy back offices are the safest first bet: insurance claims, contract review, compliance screening, invoice matching. These teams read documents and copy fields all day. A retrieval system with good extraction cuts that sharply, and the result is easy to measure: documents processed per person per day.
Internal support is the second. HR, IT, and finance teams answer the same questions repeatedly. A retrieval system over policies and tickets deflects a real share of them.
The mistake is starting with the glamorous, high-risk use case, like a fully autonomous agent that makes decisions. Start with the boring document workflow. It ships faster and builds the internal trust you need for bigger bets. For a wider view of where AI returns real ROI, see our piece on AI automation for business.
Two things drive cost: the scope of the integration and the seniority of the team, in that order. The model itself is a small line item. Most of the budget goes to data work, integration, evaluation, and the months of iteration it takes to make a system reliable enough for a business to depend on.
The table below shows the monthly rate of a dedicated engineer through a China-based team like ours. These are the same published rates we quote international clients, and they include project management and QA.
| Level | Monthly rate (China-based, dedicated) |
|---|---|
| Junior (1 to 3 years) | $2,000 to $3,500 |
| Mid (3 to 6 years) | $3,000 to $5,500 |
| Senior (6+ years) | $5,000 to $8,000 |
| Architect / tech lead | $7,000 to $10,000 |
A generative AI project is rarely one engineer. A realistic first production system is a small team: a senior engineer to own the architecture, one or two engineers to build, and part-time product and QA.
Rough build ranges, based on the same cost logic as any software project plus the model line item:
These are ranges, not quotes. The only number that matters comes from a scoped proposal, which I wrote about in how to estimate a custom software project.
Running costs are smaller than most people expect and bigger than they plan for. Hosted model APIs charge per token, roughly a few dollars per million input tokens and more per million output tokens depending on the model. For most internal tools the model bill is a few hundred dollars a month. What people miss is the engineering cost: the model needs monitoring, prompts need updating as models change, and the evaluation set needs to keep growing. That ongoing work is a real budget line.
If you want the full breakdown of how AI budgets are built, read our AI development cost guide.
The model is the easy part. The four things that decide whether a generative AI project ships, and whether it stays live, sit around the model.
Evaluation. You cannot tell whether a model is doing a good job by reading three sample answers. You need a set of test questions with known-good answers, and you run the model against them every time you change a prompt, a model, or a retrieval setup. Without this, every tweak is a guess.
Guardrails. A model that drafts an email is low risk. A model that can update a record or send a message is a different animal. Decide what the model is allowed to do, and check every output before it acts.
Data. Where does your data go when it leaves your system and hits a model API? The answer differs by provider and contract. If your data is regulated, or your customers expect it to stay with you, this question comes before the build, not after. Some teams route only non-sensitive text to a hosted model and keep the rest on an open model they run themselves.
Latency and volume. A support agent will wait two seconds for a drafted reply. A search box will not. Where the model sits in your user's flow decides whether you need streaming, a faster model, or a cached answer, and each choice has a cost.
Three decisions set the cost and the risk before a line of code is written.
Retrieval versus fine-tuning. Retrieval feeds your documents to the model at query time. Fine-tuning trains the model further on your data. Most teams should reach for retrieval first; it is cheaper, easier to update, and covers most business needs. Fine-tuning is for a narrow, repetitive, high-volume task with clean examples. The full decision is in RAG vs fine-tuning.
Hosted model versus open model. A hosted API is the fastest way to start and the cheapest at low volume. An open model you run yourself gives you control over where data goes, which matters for regulated or sensitive workloads.
Build, buy, or extend. Build when the AI is your product or a durable competitive difference. Buy when a mature vertical tool already does the job. Extend when you have a solid internal system and want to add a retrieval layer or a copilot on top. Most companies should land on extend. Our build vs buy AI software guide works through this with more concrete questions.
A working generative AI team is smaller than most people expect, but the roles matter more than the headcount.
You need a senior engineer who has shipped a production LLM system before, not just someone who has used a chatbot API. You need someone who owns the data, because data quality decides the outcome more than model choice. You need a product owner inside your business who can define what good means for a specific workflow. And you need QA that includes evaluation, because you cannot unit-test a language model the way you test a CRUD app.
If the vendor cannot name the senior engineer who will lead your project and put them on a call, treat that as a warning sign.
The fastest way to spend a lot and ship nothing is to skip the unglamorous parts. Here is the sequence that has worked across our projects.
Pick one workflow with real volume and a clear before number. Documents processed per person per day, tickets deflected, hours per invoice. That number is how you justify the next phase.
Write the evaluation set before the code. Write down twenty questions the model should answer correctly, with known-good answers. If you cannot do that, you are not ready to build.
Keep a human in the loop for anything that acts. Decide what the model can do, check every output before it acts, and keep a person approving anything risky until the guardrails have proven themselves.
Run a short paid pilot. Two to four weeks on a defined slice of the workflow tells you more than any vendor deck. It should end with a working prototype on real data and a written list of what is missing for production.
Then decide. If the pilot showed a real gain, scope the production system with governance and evaluation included from day one. If it did not, you spent a small amount to learn that cheaply.
When you are ready to bring in a team, read how to choose an AI development company. If you already have software and want to add generative AI to it, start with how to integrate an LLM into existing software. And if this is one step in a larger program, the pilot-to-production path in enterprise AI development is worth reading.
It is building software around a foundation model so the model can write, read, or act on your data. The model is one component; the development work is the data, integration, evaluation, and guardrails around it.
Predictive AI classifies and forecasts. Generative AI creates text, images, voice, or code. They use different models, different data, and different ways of testing. A fraud-detection model and a document-writing copilot are different projects.
A prompted-API feature runs roughly $10,000 to $30,000. A retrieval system runs $30,000 to $80,000. A tool-calling agent starts around $80,000. Running costs are mostly engineering and monitoring; the model API itself is usually a few hundred dollars a month for internal tools.
No. Retrieval feeds your data to the model at query time and covers most use cases. Fine-tuning helps only for a specific style or task at high volume. Details in RAG vs fine-tuning.
Sometimes. A meeting summarizer or a standard chatbot is better bought than built. Build when the AI is your product, a competitive difference, or something that must fit your data and compliance rules. See build vs buy AI software.
A prompted-API feature can go live in two to four weeks. A retrieval system usually takes six to ten weeks. A tool-calling agent with guardrails runs three to five months. The schedule is set by your data and integrations, not by the model.
It depends on the provider and your contract. For regulated or sensitive data, check where the data goes, whether it is used for training, and whether you need to run an open model in-house instead.
GlobeSoft is a software development company founded in 2018, with 40-plus engineers and more than 300 delivered projects for over 100 clients across the US, Brazil, South Africa, and Singapore. We build custom software, AI products, SaaS platforms, FinTech systems, and ERP and CRM on a stack that includes Java, Spring Boot, Go, Python, React, Vue, and Node. Generative AI work sits inside the AI development services we ship, from retrieval systems to copilots to voice agents.
If you are planning a generative AI project, talk to us about your project. Tell us what workflow you want to improve and what your data looks like today, and we will tell you honestly whether generative AI is the right first step, and what it will cost.
Tell us what you are building and where you are today. We typically reply within 24 hours.
