TL;DR
From a 5,000 prompt-layer tool to a 100-million-dollar foundation model — here's what each tier of AI actually costs to build, run, and keep running.
→ See how this applies to your business (free 30-min call)"How much does it cost to build an AI" is four completely different questions wearing one sentence, with answers spanning five orders of magnitude. Somebody asking it might mean a chatbot on their website or a competitor to GPT.
Here are the four tiers, what each actually costs to build and to run, and — the part almost every estimate omits — what it costs to keep alive.
Tier 1: Prompt-Layer Applications
What it is: a product that calls an existing model API with well-designed prompts, some retrieval, and a user interface. A document summarizer, a customer support assistant, an internal research tool.
Build cost: 5,000 to 40,000. Typically two to eight weeks of one or two developers.
Running cost: dominated by inference. For a business tool serving a few hundred users, expect 100 to 2,000 a month in model API costs. This scales with usage, not with revenue, which is a distinction that has surprised a lot of founders.
What drives the number:
The honest assessment: most business "AI projects" belong here and are quoted at three times what they should cost, because the vendor is pricing the word "AI" rather than the work.
Tier 2: Agentic Workflows
What it is: a system that takes multi-step actions with tools — reads a CRM, makes decisions, calls APIs, holds a conversation, writes results back. Voice agents, autonomous research agents, workflow automation with real branching.
Build cost: 15,000 to 150,000, depending mostly on integration depth and the reliability bar.
Running cost: 500 to 10,000 a month. Voice adds telephony and speech costs on top of inference, and voice agents burn tokens faster than chat because they run continuous context.
What drives the number:
This is the tier where most genuine business value currently sits, and where we spend most of our time.
The gap between a demo that works and a system that runs unattended for a year is roughly 5x the cost, and it's entirely in the failure paths.
Tier 3: Fine-Tuning and Custom Training
What it is: taking an existing open-weight or hosted model and training it further on your data.
Build cost: 25,000 to 250,000. The compute is often the smallest line. Data preparation dominates — cleaning, labeling, formatting, and validating a training set is where the months go.
Running cost: if you're self-hosting, GPU inference for a mid-sized model runs 500 to 5,000 a month depending on throughput and whether you need dedicated capacity. Serverless GPU is cheaper at low volume and more expensive at sustained load.
What drives the number:
The honest assessment: fine-tuning is dramatically over-prescribed. In the large majority of cases we've evaluated, better retrieval and better prompting beat fine-tuning at a tenth of the cost. Fine-tuning genuinely wins for narrow, high-volume, latency-critical tasks with stable requirements — classification at scale, a specific output format, a domain vocabulary the base models handle poorly. It rarely wins for general capability.
Tier 4: Foundation Models
What it is: training a large model from scratch.
Cost: tens of millions to hundreds of millions of dollars for a competitive frontier model, plus a research team you'll be competing with well-funded labs to hire.
The honest assessment: if you're reading a blog post to determine this number, this isn't your tier. Fewer than a hundred organizations globally should be doing this.
The Costs That Aren't in the Quote
Whatever tier you're in, these are systematically omitted from build estimates:
Rule of thumb: budget 30 to 40% of the initial build cost annually for maintenance. Projects budgeted at zero maintenance are the ones that get quietly switched off in month nine.
How to Cut the Cost Legitimately
Narrow the scope hard. "Qualify inbound leads and book the good ones" is a fraction of the cost of "an AI assistant for our business," and it's the one that produces revenue.
Tier your models. Use a frontier model for reasoning and a cheap fast model for classification and routing. This alone often cuts inference cost by 60 to 80%.
Rent the infrastructure, own the logic. Don't rebuild telephony, speech, or transport. Do own your qualification criteria, conversation design, and data.
Escalate generously in v1. Handling 70% of cases and handing off the rest costs a fraction of handling 95%, and it ships months earlier.
Validate manually first. Run the process with a human and a written script for two weeks. If it doesn't work manually, automation makes it fail faster.
Cache and batch. Prompt caching and batch processing reduce inference costs substantially for repeated context.
What to Expect at Common Price Points
How to Read a Quote
When you get a proposal, the useful signal isn't the total — it's whether the vendor can decompose it. Ask for the breakdown and look for these lines:
Two red flags worth walking away from: a vendor who won't tell you which model they're using, and one who quotes a fixed price without asking to see your data. The first means you can't reason about cost or capability; the second means the number is fiction.
The Return Side of the Equation
Cost only means something against value. Run the recovery math before the build math.
A business missing 50 inbound calls a month, closing 30% of answered calls at a 1,200 average ticket, is losing roughly 216,000 a year to a phone that doesn't get answered. A system that recovers a third of that returns 72,000 annually against a build in the low tens of thousands and a run cost in the hundreds per month.
That's the calculation that should drive the budget — not a comparison against other vendors' quotes.
We build in tier 2, deliberately: AI caller agents that reach every inbound lead within 90 seconds, qualify against the criteria a business defines, and book qualified leads straight onto a calendar, writing everything back to a GoHighLevel pipeline the client owns.
If you want a real number for your specific situation — including an honest answer if the return doesn't justify the build — [book a free strategy call](/book).
Free Weekly Briefing
One AI Marketing Tactic.
Every Tuesday. Free.
What's actually working across our client accounts right now — ROAS moves, follow-up sequences, creative angles. The stuff that isn't in any blog post yet.
No spam. Unsubscribe anytime. 1,200+ business owners already in.