How to Read an AI Project Quote

A bigger number on the quote doesn't mean a better project. You need to read what the numbers mean to make good decisions.

The Moment You Feel Powerless in Front of a Quote

When you first receive an AI project quote, most decision-makers share the same experience. There are many line items, the terminology is unfamiliar, the amounts are large — and you have no basis to judge whether it's expensive or cheap, necessary or unnecessary.

So you fall into one of two extremes. Trust the vendor and approve as-is, or cut by gut feeling. Neither is a good decision.

This article dissects the common line items in AI project quotes, showing where money actually goes and where there's room to negotiate.

The Basic Structure of a Quote

AI project quotes, regardless of scale, generally follow a similar structure.

① Consulting/Analysis: Current state assessment, requirements definition, data analysis

② Development/Build: Model development, system construction, integration work

③ Data: Data collection, cleaning, labeling

④ Infrastructure: Servers, cloud, API costs

⑤ Testing/Validation: Performance testing, user testing

⑥ Maintenance: Operations, monitoring, updates

Understanding each item's share of total cost and what it means is when the quote starts making sense.

Dissecting Each Item

Consulting/Analysis — Looks Expensive, But Don't Cut It

Often 10–20% of total cost. It's the item most likely to get the reaction: "We're just having meetings — why is this so expensive?"

But skipping or shrinking this stage causes far greater costs later. Poorly defined requirements → wrong development → complete rework. The cost of this pattern is multiples of the consulting fee.

Checkpoint: Are the deliverables for this stage clearly defined? "Current state analysis report," "data diagnostic results," "functional requirements document" — specific outputs should be listed. If it just says "analysis and planning," ask for details.

Development/Build — The Biggest Chunk, the Haziest Item

30–50% of total cost. And the hardest item to evaluate.

The key issue: diverse tasks get lumped together under "development." Building the AI model itself, connecting it to existing systems, and creating the user interface — these three have different natures and difficulty levels.

Checkpoint: Request a breakdown of the development item. How much for model development, system integration, and UI development separately. And confirm: "Are we using an off-the-shelf model, or building from scratch?" Custom-built models can cost anywhere from several times to tens of times more.

Data — The Underestimated Cost Black Hole

It looks small on the quote, but it's the most common cause of budget overruns in real projects.

If data exists in clean form, costs are low. In reality, data collection, cleaning, format standardization, and labeling (humans classifying data so AI can learn) consume unexpected time and money.

Checkpoint: Check the "assumptions behind the data cost estimate." If there's a note saying "assumes data is already cleaned," that means cleaning costs will be additional. And clarify who does the cleaning — the vendor or you.

Infrastructure — Small Initially, Grows in Production

Cloud server costs, GPU fees, API call costs. They don't look big during initial build, but can spike dramatically in production based on usage.

Generative AI in particular often has per-API-call pricing. Monthly costs with 10 users versus 1,000 users are vastly different.

Checkpoint: Request "estimated monthly operating costs based on projected usage." And get scenarios for what happens when usage doubles, quintuples, or grows tenfold.

Testing/Validation — Dangerous If Missing

If this item is completely absent from the quote, it's a warning sign. They might say "it's included in development," but it needs to be a separate line item to ensure adequate time and resources.

Checkpoint: Verify that test scope (functional, performance, security), test owner (vendor, you, third party), and test duration are specified.

Maintenance — Skip This Question and It Hurts Later

AI isn't build-once-and-done. Model performance degrades over time (as data distributions shift), new requirements emerge, and system environments change.

Maintenance costs typically run 15–25% of initial build cost annually. Leave this out of the initial quote, and it reappears a year later as "additional costs."

Checkpoint: Confirm that maintenance scope (monitoring, retraining, incident response), response times, and cost structure are included in the contract.

Common Fluff Found in Quotes

Not every item is necessary. These tend to be over-included.

Excessive customization: Fine-tuning or custom model development included for problems that a general model + RAG could handle. As we covered in a previous post, most companies get sufficient results at Stage 2 (RAG).

Unnecessary dashboards: A sleek visualization dashboard is included, but ask — will anyone actually look at it daily? Existing spreadsheets or BI tools might suffice.

Over-provisioned infrastructure: Server configurations assuming massive traffic from day one. Starting small and scaling as needed is the advantage of the cloud era.

A Practical Method for Comparing Quotes

When comparing quotes from multiple vendors, looking only at the total leads to bad judgment.

Compare by line item. Vendor A is 70 million won, Vendor B is 50 million won — but A includes data cleaning and maintenance while B doesn't. Align the scope before comparing.

Focus on what's missing. Check what's absent from the cheaper quote. Missing items come back later as additional costs.

Calculate Total Cost of Ownership (TCO). Compare not just initial build cost, but the sum of 3 years of operations, maintenance, and infrastructure. Some structures have low upfront costs but high operating costs.

If You Can Read the Quote, You Can Negotiate

When you understand what each line item means, your negotiation language changes.

Not "Can you give us a discount?" but rather: "We'll handle data cleaning internally, so please remove that item," "We don't need this infrastructure scale for the pilot phase, so let's downsize," "How does the cost change if we set the maintenance SLA at this level?"

These conversations become possible. And vendors are more honest with customers who can have these conversations.