Every article about AI development cost follows the same script: a table of feature-based prices copied from three other articles, a paragraph admitting that "it really depends on your requirements," and a contact form. If you are trying to put a real number into next quarter's budget, that is useless. We scope AI projects every week, so let us do this differently: actual 2026 ranges, and more importantly, the three variables that decide where inside the range you land.
One warning before the numbers. If a vendor gives you a confident fixed price before anyone has looked at your data, they are pricing the demo. The demo is about 20% of the work.
The short answer: in 2026, custom AI development costs roughly $4,000–$12,500 for a support chatbot with RAG, $7,500–$30,000 for a tool-using AI agent, $10,000–$35,000 for a voice agent, and $12,500–$50,000+ for a computer-vision system, built by a senior India-based team like AGI Software Solutions. A US or Western European agency quotes 4–8× more for the same scope. Add 15–20% of the build cost per year for inference, evaluation and maintenance.
Why "it depends" keeps happening
Traditional software gets priced by counting features: screens, roles, reports, integrations. Estimate hours per feature, multiply by rate, add contingency. That works because effort in a CRUD app is roughly additive.
AI projects break this model. We have scoped two support chatbots with near-identical feature lists where one cost four times the other. The features were the same; everything underneath them was not. In our scoping sheets, cost behaves like a multiplication of three factors, not a sum of features:
- Data readiness. Where does the knowledge the AI needs actually live? A clean product database plus a maintained help centre is one price. Four thousand PDFs, a wiki nobody has touched since 2022, and tribal knowledge in six people's heads is a very different price, for the exact same chatbot.
- Integration surface. How many systems must the AI read from or write to, and how hostile are they? Answering questions from documents is cheap. Checking a live order status inside a 2009-era ERP with no real API, while respecting per-user permissions, is where budgets actually go.
- Model strategy. Calling a frontier API, fine-tuning a model, or self-hosting open weights. This choice moves the build cost and completely changes the shape of the ongoing bill.
Multiply those three and you understand the enormous spread in every price table you have ever seen. A project that is easy on all three axes genuinely is a four-to-low-five-figure build. A project that is hard on all three is a low-six-figure programme, whatever the feature list says.
Nobody blows an AI budget on the model. They blow it on the data the model was supposed to understand and the systems it was supposed to talk to.
The three model strategies, priced honestly
API-based (Claude, GPT, Gemini)
You call a hosted frontier model over an API. Build cost is the lowest of the three tiers because the hard ML work is someone else's problem; your money goes into retrieval, prompting, tool wiring, and evaluation. Ongoing cost is usage-based inference. For the mid-market deployments we run, that is typically $150 to $3,000 per month, scaling with traffic. This is the right default for the large majority of projects, and it is where most of our AI and ML integration work lives.
Fine-tuned models
Fine-tuning makes sense when prompting plus retrieval provably cannot hit a tone, format, or domain-accuracy target. It adds $5,000 to $20,000 on top of the base build for dataset preparation, training runs, and the evaluation harness you now need to prove the fine-tune actually helped. Teams reach for this far too early. Exhaust the prompt-and-retrieval option first; it is cheaper to iterate and easier to undo.
Self-hosted open weights
Llama, Mistral, or Qwen-class models on GPUs you control. The build is the most expensive of the three because you are now doing MLOps: serving infrastructure, quantization decisions, failover, upgrades. The ongoing cost flips from usage-based to roughly flat: $1,000 to $8,000 per month in GPU spend, plus someone who can babysit it. It is justified by two things only: data-residency rules that forbid external APIs, or token volumes so high that flat infrastructure beats per-token pricing. Below a few million tokens a day, it rarely pays.
Realistic 2026 ranges by use case
These are build costs with a senior India-based team like ours. A US or Western European agency delivering the same scope will land four to eight times higher; we cover why below.
Support chatbot with RAG: $4,000-$12,500
Scoped means a bounded set of knowledge sources, a clear escalation path to humans, and one or two channels. The spread inside the range is almost entirely data readiness: clean docs sit at the bottom, PDF archaeology at the top. The enterprise version of the same idea, with SSO, audit logging, six knowledge systems, and a compliance review, runs $30,000 to $75,000+ with us; Western vendors quote multiples of that.
Tool-using AI agent: $7,500-$30,000
An agent that acts: creates tickets, updates records, issues refunds, drafts and sends. Every tool it can use adds an integration, a permission model, a set of failure modes, and test cases. Cost scales with the number of tools and the blast radius of a wrong action; guardrails and approval flows for write actions are where the money goes, and they are not optional. This is the core of our AI agent development work, and it is the category where underscoping hurts the most.
Voice agent: $10,000-$35,000
Everything the chatbot needs, plus telephony, interruption handling, and a latency budget under one second that changes the whole architecture. Voice also has far less forgiving evaluation: a clunky text answer gets re-read, a clunky spoken answer gets hung up on. The upper half of the range is mostly latency engineering and call-quality testing.
Computer-vision system: $12,500-$50,000+
The widest spread, because labeled data dominates. A common problem with strong pre-trained coverage, like standard document layouts or well-known defect types, sits at the low end. Custom detection on your specific factory line, with nothing labeled today, means the majority of the budget goes to data collection and labeling before anyone trains anything.
Document-processing pipeline: $6,000-$22,500
Invoices, claims, or contracts in; structured data out. The spread is driven by document variety and the accuracy target. Hitting 90% with a human review queue for the rest is the low end. Pushing toward 99% straight-through processing costs a multiple of that, because the last few points of accuracy are the expensive ones.
Why a US agency quotes 4-8× more for the same scope
Rates, mostly. A senior US engineer bills $150 to $250 an hour; a genuinely senior engineer in India bills $15 to $30. Both use the same frontier models, the same open-source retrieval stacks, and the same GitHub. The model does not know which timezone its prompt was written in.
What does not change with geography, and what nobody puts in the sales deck:
- Your data cleanup effort. Whoever builds the system, your team does the work of finding, correcting, and blessing the source data. That internal cost is identical at every vendor price point.
- Review loops. Your subject-matter experts have to judge outputs and correct them, usually across six to ten iterations. No agency can compress this, because the bottleneck is your experts' calendars.
- Calendar time. Delivery speed is limited by decisions and data access, not typing speed. Paying quadruple does not make your IT team grant API credentials faster.
The honest caveat: the comparison only holds between senior teams. A cheap offshore team of juniors reselling a thin wrapper will cost you more than any US agency once you count the rewrite. What you are evaluating is seniority and evaluation discipline, not a flag. Ask any vendor, anywhere, how they build eval sets and how many production AI systems they operate today; the answers separate teams faster than the rate card does.
The costs that surprise people
Four items keep ambushing budgets, whoever builds the system: the discovery and data-audit phase that has to happen before a real quote exists; the evaluation set, because someone must write and grade hundreds of test cases against real scenarios; the last 10% of accuracy, which routinely costs as much as the first 90%; and the inference bill, which grows precisely because the system is being used. We wrote a separate deep dive on these in the hidden costs of enterprise AI; if you are budgeting a project, read it before you sign anything.
How to get a real number instead of a range
There is exactly one honest way to turn the ranges above into a fixed price: a short, paid discovery phase. Ours runs one to two weeks and produces five things: an audit of where your data actually lives and what state it is in, an integration map with the authentication realities spelled out, a model-strategy recommendation with a projected monthly inference cost, an evaluation plan defining what "working" means in numbers, and a fixed-price build quote we will stand behind.
A vendor who skips discovery and quotes anyway has not removed the uncertainty; they have hidden it inside contingency you are paying for either way, or inside change orders you will pay for later. If you want a real number for your project, talk to us about a scoping engagement; the discovery fee is credited against the build if you proceed.
Where to start
- Write one page on where the knowledge lives. List the systems, the document stores, and who maintains each. This does more for quote accuracy than a thirty-page requirements document.
- Decide read-only versus write actions. An AI that answers is one budget class; an AI that acts is another. Knowing which you need roughly halves the quoting uncertainty on its own.
- Default to API-based models unless a regulator or your volume forces otherwise. You can always graduate later; you cannot easily un-buy GPUs.
- Budget 15-20% of the build cost per year for operations: inference, model updates, eval maintenance, and prompt drift. A launch budget with zero ongoing line is a red flag, including on your own spreadsheet.
- Pay for discovery before you commit to a build number. It is the cheapest week of the entire project, and it is the only week that makes every other number real.
Frequently asked questions
How much does it cost to build an AI chatbot?
A scoped support chatbot with RAG costs $4,000–$12,500 with a senior India-based team in 2026: a bounded set of knowledge sources, a clear escalation path to humans, and one or two channels. The enterprise version — SSO, audit logging, multiple knowledge systems, compliance review — runs $30,000–$75,000+. Data readiness drives most of the spread.
How much does it cost to build an AI agent?
A tool-using AI agent costs $7,500–$30,000 in 2026 with a senior India-based team. Cost scales with the number of tools the agent can use and the blast radius of a wrong action; guardrails and approval flows for write actions are where the money goes, and they are not optional.
What are the ongoing costs of an AI system?
Budget 15–20% of the build cost per year for inference, model updates, eval maintenance and prompt drift. API-based inference for mid-market deployments typically runs $150–$3,000 per month; self-hosted open weights flip that to roughly flat GPU spend of $1,000–$8,000 per month.
Why do AI development quotes vary so much?
Because cost multiplies across three factors instead of summing across features: data readiness, integration surface, and model strategy. Two projects with identical feature lists can differ by 4× on those factors alone — which is why any confident quote made before someone looks at your data is pricing the demo.
Is it cheaper to hire an offshore AI development company?
A senior US engineer bills $150–$250 an hour; a genuinely senior engineer in India bills $15–$30, using the same models and the same stacks — so the same scope lands 4–8× cheaper. The comparison only holds between senior teams: juniors reselling a thin wrapper cost more than any US agency once you count the rewrite.