KleioKleio

Thesis & research · Pricing

What an AI thesis project actually costs.

Nobody publishes a price list for this, and the reason is that the range is enormous. Here is what actually drives the number, so you can scope a project to the budget you have.

Kleio Software9 min read

This is the question everyone asks first and almost nobody answers publicly, which is frustrating when you are a student trying to work out whether this is even possible on your budget.

The reason nobody publishes a number is not evasiveness. It is that two projects described in the same sentence can differ by a factor of ten in actual work. “A machine learning model for my thesis” covers both a prediction model on a clean spreadsheet and a computer vision system on images that do not exist yet.

So instead of a number that would be wrong for most readers, here is the structure underneath the number. Once you can see which variables your project triggers, you can have a specific conversation with anyone you approach — and you will be able to tell whether their quote is reasonable.

What actually drives the cost

Five things, in roughly descending order of impact. Most students assume the model is the expensive part. It is usually the cheapest part.

The state of your data — by far the largest variable
A clean tabular dataset in one spreadsheet with consistent columns is the cheap case. Data spread across scanned PDFs, inconsistent forms, or several sources with mismatched units can easily be more work than the modeling. If you can only influence one thing, influence this.
The type of problem
Tabular prediction and classification are the most economical. Text work costs more. Computer vision costs the most, largely because of labeling effort. If images are involved, expect the number to roughly double against an equivalent tabular project.
Whether it needs to be deployed
A trained model plus evaluation is one scope. A model wrapped in a working web interface your panel can operate is meaningfully more. It is also, in our experience, the thing that most changes how a defense goes.
How much has to be explained to you
This should be non-negotiable and it does take time. You need to understand the architecture well enough to defend it. Any quote that does not include a walkthrough is quoting a different, lesser thing.
Your timeline
Six weeks before your defense is normal planning. Two weeks before is compressed work at a premium, if it can be taken on at all. Booking early is the single most effective way to reduce what you pay.

How to think about the range

Rather than a peso figure that would mislead, here is the useful frame: research work is quoted on effort, and effort is measured in weeks. Ask any prospective collaborator to break their quote into the phases below. The breakdown tells you far more than the total does.

A tightly scoped tabular project on data that is already clean sits at the low end of whatever range you are quoted. A computer vision project on unlabeled images, with deployment, sits at the high end. Most projects land in between, and most of the spread is data preparation.

Phase 1 — Data assessment
Reviewing what you have and confirming it can support the question. Should be quick, and some of it should be free. If nobody offers to look at your data before quoting, that is a warning sign.
Phase 2 — Data preparation
Cleaning, structuring, encoding, feature engineering. The most variable phase and usually the largest.
Phase 3 — Modeling and evaluation
Baselines, the chosen model, tuning, and honest evaluation with proper metrics. More predictable than people expect.
Phase 4 — Deployment
Wrapping the model in an interface with a public URL. Optional, but this is what your panel interacts with.
Phase 5 — Walkthrough and defense preparation
Making sure you can explain every decision. Never optional.

How to make your project cost less

Almost all of the leverage is in things you control, and almost all of it has to be exercised early.

Come before your data collection is finished. A thirty-minute conversation about what to record and how to structure it routinely removes a week of cleaning later. This is the highest-return thing on the list and it usually costs nothing.

Narrow the question. Predicting one well-defined outcome is dramatically cheaper than a system that does three things. Panels reward specificity anyway, so this improves the thesis as well as the price.

Bring your data in one file with consistent columns if you possibly can. Time you spend consolidating spreadsheets is time nobody has to bill you for.

Say your budget up front. It is not a negotiating weakness; it lets the scope be shaped to fit rather than quoted and then argued down. A good collaborator will tell you honestly what is achievable inside it, including if the answer is nothing.

What to watch out for when comparing quotes

Some things should make you pause regardless of price.

A quote given without anyone looking at your data. The data is the main cost driver, so a number produced without seeing it is a guess that will be revised upward later.

No walkthrough included. If nobody is going to explain the model to you, you cannot defend it, and you have bought a liability.

An offer to write the thesis itself. That is a different service, it is an academic-integrity problem, and it is not what any legitimate technical collaborator does. Building the system is engineering; writing your research is yours.

A price far below everything else you have been quoted. On research work this usually means the evaluation will be superficial, which is precisely the part a panel examines hardest.

Get a number for your actual project

Bring what data you have, even if it is messy, and your defense date. The call is free, and if the project will not work as scoped you will hear that rather than a quote.

Book a free discovery call

Related questions

The follow-ups people ask after reading this.

Why will nobody give me a price without a call?

Because the state of your data drives most of the cost, and nobody can assess that from a description. A quote given without seeing the data is a guess that gets revised upward once the real work becomes visible. A short conversation usually produces a number that holds.

Are research projects priced differently from commercial work?

They should be. Student budgets are not business budgets, and a scope can be shaped to fit one. The practical approach is to state your budget early so the project is scoped to it, rather than quoted at commercial rates and then cut down.

Does deployment cost much more than just the model?

It adds meaningfully to the scope, but it is usually the highest-value part of the spend. A panel that can put inputs into your model and watch it respond engages with the work differently than one being shown a notebook. If something has to be cut, cut scope elsewhere first.

What is the cheapest kind of AI thesis project?

A prediction or classification model on tabular data you already have in a clean, consistent spreadsheet, with a single well-defined outcome variable and no computer vision. If you are choosing your topic with budget in mind, that is the shape to aim for.

Can I pay for just part of it and do the rest myself?

Often yes, and it can be a sensible way to work to a budget. Common splits are having the data preparation and model built while you handle the write-up, or having the modeling done and deploying it yourself. Be clear about the boundary at the start so nothing falls between the two of you.

Keep reading

AI for thesis & capstone projects

What a research engagement actually includes, end to end.

How to incorporate AI into a thesis

Work out whether your data can support a model before you budget for one.

AI thesis ideas by course

Project shapes and their data requirements, by field.

Still have a question this didn't answer?

Bring it to a free discovery call. No pitch, straight to the developer, and an honest answer even when it is that we're not the right fit.

Book a free discovery call