Four Quotes on the Approver's Desk
A manufacturing decision-maker told me this last week: "I got four PoC quotes. The lowest was ₩4M, the highest ₩80M. A 20× spread. What am I supposed to compare?"
The question itself already reveals half the problem. The reason the numbers on those quotes vary by 20× is that each vendor is selling something different under the same word — "PoC." And most approvers sign off on the approval without being clear, even to themselves, about what they are actually buying.
The first article in this series laid out a rough sense of total budget by company size. Today we drill into the first gate on that path — the four-week PoC budget. What actually separates the ₩8M option from the ₩40M option? And does that difference matter to the decision-maker?
A PoC Is Not an ROI Study
Start with the most common misconception: the belief that a PoC will answer "how much is this AI worth to our company." It will not. Four weeks is not enough time.
The real purpose of a four-week PoC is much simpler. Find out which of your internal datasets are dirtier than you expected. Verify that the pipeline actually runs once you relax the accuracy target. And get a rough sense of what a single query will cost per unit once the system moves into production.
If you walk out of four weeks with those three things, the PoC was a success. ROI only begins to take shape at the pilot stage — three months or more. This is where I see approvers and vendors talk past each other most often. The approver wants to see ROI; the vendor only needs to sell as far as "a demo that runs."
What Four Weeks Buys — Three Deliverables
A properly scoped four-week PoC should hand the approver three things:
- A validated data model — a schema tested against real internal data, plus a document identifying which fields are missing and which are noisy
- An inference architecture decision — on-premises versus cloud, open model versus API, with the reasoning behind the call
- A per-query unit cost estimate — a cost curve across monthly volume scenarios assuming real production use
Without those three, no deliverable — however polished the demo — gives the approver a basis for the next decision. A single chatbot running on a screen will not defend next quarter's budget.
The Right Budget — Between ₩2M and ₩40M
From what we see on the ground, the standard range for domestic PoCs is ₩2M–₩8M. Local quote data compiled by windyflo this year supports that band. But most of that price range covers only shallow PoCs — a single workflow, a single data source, largely wrapping existing API calls.
Globally, a realistic four-week PoC budget lands at $15,000–$30,000, or roughly ₩20M–₩40M. At that level, consulting-led scoping, data cleansing, and inference architecture validation are all included. And in that range, the biggest cost driver is not the model itself — it is data quality, integration complexity, and the required accuracy threshold.
Put differently, the gap between the ₩8M option and the ₩40M option is not about AI model performance. It is about whether the budget also covers data cleansing and integration validation. This is exactly the point the approver should press the vendor on: "How many person-hours of data cleansing are built into this quote?"
Why 60% of PoCs Stall
According to Gartner's 2025 report, roughly 60% of enterprise AI projects are halted at the PoC stage. MIT NANDA's 2025 study is more sobering: only 5% of integrated AI pilots produced measurable P&L impact, while the remaining 95% delivered no measurable results.
What is interesting is the failure pattern. It is not model performance. It is the inability to embed a validated AI into an actual business workflow, data quality that turned out worse than assumed, and business value that was never clearly defined from the start. The same three causes keep repeating.
Gartner also projects that more than 40% of agentic AI projects will be canceled by the end of 2027. A large share of those cancellations trace back to the same condition — "costs climbed higher than expected and the business value blurred." Read the other way, projects that invested seriously in scoping are the ones that stay inside their budgets. Adding a week of scoping to the front of a four-week engagement is by far the better trade.
What the Approver Should Be Asking
Before signing the approval package, the questions compress into three. How many hours of data cleansing are actually in this quote? At the end of week four, is the deliverable a demo or a document? Does it include a monthly unit-cost estimate for the production scenario? Any quote without clear answers to those three deserves a second look, regardless of the price tag.
The next article picks up after the PoC succeeds — do you build your own LLM, or subscribe to an API? We dig into where the actual break-even point sits.
If you need to map out a four-week scenario at the right scale for your company, consider a cost sizing and PoC scoping consultation. We will help you build the framework for comparing vendor quotes from the very first meeting.