How We Actually Quote AI Projects, and Why Most Estimates Are Wrong

Almost every founder we talk to has been quoted for at least one AI project before we get on the call. The quotes are all over the map. $8k for a full AI product. $200k for a chatbot. The reason isn't that agencies are dishonest — it's that AI project quotes are structurally harder than software project quotes, and most teams pretend they aren't. This is how we quote AI work, honestly.

Why AI project quotes are different

In a typical software project, the specification defines the work. Login screen, dashboard, three CRUD endpoints, a Stripe integration. An experienced team can estimate that in a few hours because the shape is known.

AI projects break that model in three ways.

Quality is discovered, not defined

You can't specify "the assistant should give correct answers." You can specify what "correct" looks like on your dataset, but you won't know how hard it is to hit that quality bar until you try. Some problems are easy for current models. Some look easy and aren't.

Cost is a function of unknowns

Token usage, inference latency, and hosting cost all depend on choices you haven't made yet — model, prompt structure, retrieval architecture. A quote given before those decisions is a guess.

Vendor risk changes weekly

Between quote and delivery, OpenAI can raise prices, Anthropic can ship a new model that changes your architecture, or an open-source release can make a bought API redundant. A quote that doesn't account for this is a quote that will be wrong.

The three questions we ask before quoting

Every AI project scoping call, we ask the same three things. If we can't answer them, we don't quote a fixed price — we quote a discovery phase.

1. What does "working" look like, measurably?

Not "the assistant answers questions." What percentage of questions? On what dataset? Judged by whom? "80% of tier-1 support tickets resolved without a human" is measurable. "The AI works well" isn't.

If the client can't answer this, that's not a red flag — it's the first phase of work. Getting to a measurable target is often 20% of the project.

2. Where does the data actually live?

The single biggest cost driver. Data in a well-organized Postgres database is one project. Data spread across Notion, Google Drive, five SaaS tools, and a fileshare is another project entirely — usually 2-4x more work.

We ask for the data map before pricing. If nobody can produce one, that itself is signal.

3. What's the constraint — accuracy, cost, or latency?

You rarely get all three. A high-accuracy RAG system on top of GPT-5-class models is expensive per query. A cheap system built on open-source models is faster to run but usually less accurate. A low-latency system may need caching that reduces accuracy at the edges.

Knowing which two matter most shapes every architectural decision downstream.

The shape of an AI project quote

We don't quote AI projects as a single number. We quote them as three phases with different risk profiles.

Phase 1 — Discovery and prototype

Fixed price. 1-3 weeks. Deliverable is a working prototype on a representative subset of data, plus a quality measurement showing what we can achieve. This phase is priceable because the scope is bounded.

Typical range: $3-15k depending on data complexity.

Phase 2 — Production build

Fixed price if Phase 1 established the architecture. Time and materials if it didn't. Deliverable is the shipped feature, integrated into the product, with monitoring and evaluation infrastructure.

Typical range: $15-80k depending on scope. Wider range than typical software work because AI projects have more variance.

Phase 3 — Iteration and hardening

Retainer or per-iteration pricing. AI features are never done at launch — quality metrics drift, models change, edge cases surface. Budget for 3-6 months of iteration after launch.

Typical range: $2-10k/month or per-milestone pricing.

What we won't fixed-bid

Some categories of AI work resist fixed pricing entirely. When a client insists on a fixed number, we walk from these.

• Open-ended agents — "an agent that runs our operations" has no bounded scope. Time and materials or per-workflow pricing only.
• Autonomous decision-making — anything where the AI acts without human review. The evaluation cost dwarfs the build cost, and evaluation is hard to bound.
• Novel research — problems where the answer to "is this possible" is genuinely unknown. Fixed-bidding research is how agencies go bankrupt.
• Compliance-heavy verticals without clear guidance — the regulatory scope changes the shape of the work fundamentally.

The cost drivers most quotes underestimate

Evaluation infrastructure

Every serious AI project needs an evaluation harness — a way to measure whether the model is doing what you want, on real data, over time. This is often 20-30% of the build cost and never in the marketing pitch.

Data preparation

Cleaning, chunking, deduplicating, versioning the data your model uses. On a messy corpus, this can be more than half the project. Our note on data pipelines for AI SaaS covers what this actually looks like.

Guardrails and safety

Prompt injection defenses, PII scrubbing, output validation. Optional if the AI is internal-only. Required if it's customer-facing. Adds 10-30% to the build.

Observability

Logging inputs, outputs, latencies, and costs at every step of the AI pipeline. Standard for a mature build, missing from most quotes.

Model change hardening

Every model deprecation costs a week of engineering. Building for model portability upfront saves that later. Usually skipped in low-cost quotes.

Why cheap quotes usually cost more

The $8k full AI product quotes we mentioned earlier aren't wrong on the surface work. They're wrong because the surface work is the smaller part.

Delivered projects at that price typically ship one of three things:

• A wrapper around an OpenAI API call with no evaluation and no observability.
• A demo that works on carefully chosen examples and falls over on real user input.
• A functioning prototype the client then has to spend another $40-80k to make production-ready.

The cheap quote is real. It just quotes the fun 20% and leaves the difficult 80% for later — either as change orders, a follow-on project, or a different vendor.

Fair signals in an AI quote

What to look for when comparing quotes.

• The quote names the evaluation approach. What does "working" mean, and how will you measure it?
• The quote itemizes data preparation separately from model integration.
• The quote includes ongoing iteration or retainer as a line item, not as an afterthought.
• The vendor asks about data before pricing, not after.
• The vendor is willing to phase the work — a discovery phase before a production commitment.

A quote that's a single number covering everything is almost always missing something.

Common mistakes buyers make

Comparing on price alone. AI project quotes vary by 5-10x for the same nominal work. The variance is usually about which parts are actually being quoted.

Buying the demo. Every AI vendor can build a compelling demo. The gap between demo and production is where the real work lives. Ask what percentage of the timeline is post-demo.

Skipping the discovery phase. "We already know what we want" is almost never true for AI projects. The discovery phase is where the client learns what's actually possible on their data.

Treating fixed-price as risk transfer. Fixed-price contracts on unbounded scope don't reduce buyer risk — they just push the fight into change orders. A time-and-materials contract with a discovery phase usually costs less in total.

Common misconceptions

"AI projects are just software projects with an API call." The API call is the least interesting part. Evaluation, data, guardrails, and iteration are the project.

"A senior AI engineer will get it right the first time." Even the best AI engineers don't know the answer for your specific data until they've measured. Iteration isn't a skill gap; it's the shape of the work.

"We should wait until models get better." Sometimes true, but rarely. Model progress compounds slowly for most business use cases. The data, evaluation, and guardrail work you do now transfers to future models.

Frequently asked questions

Final take

Quoting AI work honestly requires admitting what you don't know. A discovery phase to measure feasibility, a fixed-price production phase with clear evaluation criteria, and an iteration retainer for the months after launch is the pattern that survives real projects. Any quote that skips those phases is quoting the fun part and hiding the rest.

If you have quotes on the table and want a second opinion, the AI Audit is designed for exactly this — we look at the scope, the data, and the vendor's approach, and tell you what's missing. Or book a call and we'll walk through your options.