Guide · Costs & budgeting
What AI development really costs — and why the model is the cheap part.
Ask five vendors what an AI project costs and you’ll get five numbers that can’t be compared. This guide won’t give you a magic price. It shows where the effort actually goes, which costs appear only after launch, and how to structure the budget so you never pay for a full build before you know the idea works.
By Nythrex EngineeringUpdated 5 min read
Integrations & data pipelines
Application, UX & permissions
Evaluation & quality work
Infrastructure & security
Prompting & model work
Discovery & project management
Illustrative shape of effort in a typical first production release of an AI assistant — not measured data. Your mix will differ; the point is the order of magnitude.
Why AI quotes are impossible to compare
Two estimates for “an AI assistant for our support team” can differ by an order of magnitude, and both can be honest. One vendor priced a demo on clean sample documents. The other priced connectors to four systems, permission filtering, an evaluation suite, a helpdesk integration and a rollout. They are not estimating the same thing. Before comparing numbers, compare what’s included — our brief generator helps you make vendors quote the same scope.
The four cost buckets
| Phase | What you pay for | Typical effort shape | Output |
|---|---|---|---|
| Discovery | Goal, users, data audit, risks, architecture, test set | 1–3 weeks, 1–2 senior people | Written plan and estimate you can take to any vendor |
| Proof of concept | Thin end-to-end slice on real data, measured against the test set | 3–6 weeks, 2–3 people | Go / no-go with accuracy, latency and cost per request |
| Production build | Integrations, UX, permissions, guardrails, monitoring, deployment | 2–6 months, 3–6 people | A live system your team can operate |
| Operation | Model usage, hosting, monitoring, evaluation runs, improvements | Ongoing, fraction of a team | Quality that holds up as data and models change |
What moves the price the most
- 1
Number and messiness of data sources
One clean knowledge base is cheap. Five systems with scanned PDFs, spreadsheets and inconsistent permissions are not. Each source needs a connector, parsing, cleaning and a sync strategy.
- 2
Integrations with systems of record
Reading from a CRM is easy; writing back safely, with validation and audit, is real engineering. Actions multiply testing effort.
- 3
Who the users are
Internal tools can be plain. Customer-facing AI needs polished UX, abuse protection, rate limiting, localisation and brand-safe output.
- 4
How wrong it’s allowed to be
A drafting assistant that a person reviews tolerates mistakes. Automated decisions in finance or health need far more evaluation, guardrails and documentation.
- 5
Compliance and hosting constraints
EU data residency, on-premise hosting, SOC 2 or sector rules add architecture and paperwork. Budget them up front — retrofitting is expensive.
AI project effort estimator
Time to production-ready v1
14–21 weeks
Effort
39–59
person-weeks
Complexity
Significant
How the time splits
- Discovery: 2
- Proof of concept: 4
- Build: 8
- Hardening & launch: 2
Suggested team
- AI / tech lead
- Backend engineer
- QA engineer (part-time)
- DevOps (part-time)
What drives the effort most
- The solution type sets the baseline: agents that take actions need the most safety work.
- Every extra data source adds ingestion, cleaning and access rules.
- Integrations are usually the hidden half of an AI project.
A rough orientation based on typical project structure — not a quote. Cost = effort × team rates; the Discovery sprint turns this into a fixed, itemised estimate.
Send us these answers and a few sentences about the goal — we’ll tell you what we’d build first and why.
Get a real estimateThe hidden costs vendors leave out
01Data preparation
Cleaning, de-duplicating and structuring source content often takes longer than building the pipeline. Ask who does it.
02Evaluation
Writing a test set and running it on every change is real work. Without it, you can’t tell improvements from regressions.
03Permissions
Mirroring existing access rights into an AI system is rarely in the first estimate — and is a security incident when forgotten.
04Change management
Training users, updating processes and handling the “I don’t trust it” phase. AI that isn’t adopted has infinite cost per use.
05Model and API changes
Providers deprecate models. Plan budget for re-evaluation and prompt updates at least a couple of times a year.
06Content ownership
Someone has to fix the wrong or outdated documents the AI surfaces. That’s a business cost, not an engineering one.
Running costs: what to expect after launch
Monthly costs have four parts: model usage (tokens), infrastructure (hosting, search index, storage, monitoring), quality work (reviewing failures, updating the test set, re-evaluating after changes) and support. Model usage is the one people worry about and the one easiest to control — through shorter prompts, caching, fewer retrieved passages and routing simple requests to smaller models. Estimate yours with the LLM API cost calculator, and see 11 ways to cut your LLM bill.
How to structure the budget safely
Discovery · fixed price
Clarity on goal, data, risks and a real estimate
Proof of concept · time-boxed
Numbers on your data; explicit go / no-go
Build · milestone-based
Pay for working increments, changes approved with cost
Operate · monthly
Usage, monitoring, improvements
This structure caps your downside: if the proof of concept shows accuracy isn’t good enough, you stop having spent a fraction of the full budget, with a report explaining why. Compare contract models in Fixed price vs T&M vs milestones.
Questions to ask every vendor about their estimate
Frequently asked questions
Keep reading
Want a second opinion on your project?
Tell us what you’re building and where you’re stuck. We’ll reply within one business day with the most practical next step — even if that step isn’t us.