Engagement
AI development cost and engagement models
What actually drives the price of an enterprise AI project, how we scope work, and how long each stage takes. Written so you can budget before you talk to anyone.
There is no list price for AI development, and any agency quoting one before seeing your systems is guessing. What can be stated up front is the structure: what makes a project expensive, what makes it cheap, and what you get at each stage.
We scope every engagement as a fixed-outcome proposal with a written success metric, a timeline and a price. If discovery shows an off-the-shelf tool already solves the problem, that is the recommendation, and it is a far cheaper answer than finding out six months into a build.
Six things that move the number
Integration surface
Each system the solution touches (CRM, ERP, ticketing, warehouse) adds auth, data mapping and failure handling. Two integrations is a different project from eight.
Data readiness
Clean, accessible, well-labelled data shortens a build dramatically. Where labels do not exist we can often generate them from historical records instead of manual annotation.
Accuracy bar
Getting to a useful answer is fast. Getting from useful to audit-grade takes evaluation suites, retrieval tuning and review cycles, and that is where budget concentrates.
Deployment environment
A hosted endpoint is the cheapest path. Your own cloud tenancy, a specific region, on-premise or the edge each add infrastructure sizing and operational work.
Governance requirements
Audit logs, human approval consoles, redaction and access controls are not optional in regulated settings, and they are real engineering line items.
Ownership after launch
Handover with documentation and pipelines costs less over time than a retained support model. We can do either; the choice changes the shape of the number.
How we structure engagements
2-4 weeks
Discovery sprint
One workflow assessed, success metric agreed, implementation estimate written. Cheapest way to find out whether the project is worth doing.
6-12 weeks
Fixed-outcome build
A scoped production build against a written success metric, delivered into your environment with dashboards, documentation and handover.
Quarterly
Ongoing partnership
Monitoring, retraining, evaluation and the next use case, run alongside your team rather than instead of it.
Typical timelines by service
- Discovery and strategy on one workflow: 2-4 weeks
- Grounded generative AI assistant in front of internal users: 4-6 weeks
- Production AI agent on a single workflow: 8-12 weeks
- Decision intelligence supervised pilot: 10-12 weeks
- Custom or fine-tuned model, including evaluation and deployment: 10-16 weeks
Common questions
What drives the cost of an AI development project?
Four things: how many systems the solution has to integrate with, the state of your data, the accuracy bar the use case demands, and whether the model can be hosted or has to run in your own environment. Model licensing is usually the smallest line item; integration and evaluation are the largest.
How does RoseTech price engagements?
Every engagement is scoped as a fixed-outcome proposal with a written success metric, timeline and price. We do not bill open-ended time and materials, so the commercial risk of a scope surprise sits with us rather than with you.
What is the smallest sensible first project?
A two to four week discovery on one workflow, ending in a written recommendation, a measurable success metric and an implementation estimate. If the evidence says a hosted tool already solves it, we say so and you stop there.
How long does an enterprise AI project take?
Discovery and strategy typically take 2-4 weeks. A grounded generative AI pilot reaches internal users in 4-6 weeks. A production AI agent on one workflow lands in 8-12 weeks. Decision intelligence pilots usually run 10-12 weeks before measurable impact.
What does it cost to run AI once it is live?
Running cost is driven by usage volume, retrieved context size and model choice, not by seat count. We model it before the build and instrument the deployment so you can see cost per query and per team, and trade quality against spend deliberately.
Is it cheaper to fine-tune a model or use a hosted API?
A hosted API is cheaper for most workloads. Owning a model pays off when per-call pricing at your volume exceeds hosting, when accuracy on your specific task plateaus, or when data residency rules out a third party. We benchmark both against your task before recommending either.
Let’s solve your data challenge