Two-Week Discovery Sprint
AI discovery sprint: is it worth building, and what will it cost?
You are considering an LLM, RAG or agent system and someone has to decide whether to fund the build. Before that decision, it is worth two weeks to find out whether the idea works on your real data, how it should be built and what the build will cost.
That is the discovery sprint. It is my fixed-scope first engagement for a new AI build: 2 weeks, from £3,000, with the fee fixed before any work starts.
Or describe your project in the chat, or email nic.chin@nicchin.com
What you get at the end of the sprint
Four deliverables, the same on every sprint. They are yours whether or not you go ahead with the build.
- 1
A working prototype on your real data
Not a demo on tidy sample files. The core of the system - the retrieval, the model calls, the agent steps that decide whether it works - running against a representative sample of your own documents or records, so you see where it holds up and where it struggles before anyone commits to a build.
- 2
An architecture document
How the production system should be built: the components, where your data lives and who can reach it, which models and why, what has to be logged, and the options I considered and rejected. Written so your engineers, or another builder, can work from it.
- 3
An evaluation set with a pass threshold
Real questions or cases from your business with agreed correct answers, built with the people who know the work, and a pass mark agreed before any results are seen. The prototype is scored against it during the sprint, and the build is scored against the same set.
- 4
A fixed build estimate
A fixed price and timeline for the production build, set from what the sprint measured on your data rather than guessed from a brief. If the honest answer is not to build, the estimate says that instead.
Who the sprint is for
Funded or revenue-stage companies in the UK, US and Europe that have a specific job in mind for an LLM, a RAG system or an agent - answering questions from internal documents, reviewing contracts, triaging cases, drafting from records - and want evidence before they commit a build budget to it. It works best when there is one workflow to start with and someone in the business who knows what a correct answer looks like.
It is the wrong purchase if:
- The system already exists. A review is the right first step for a working system; see already have a working system below.
- There is no workflow yet, only an intention to “use AI”. Pick the one job that costs your team the most time first. The consulting guide has the reading for that stage.
How the two weeks run
Week one: the data and the definition of correct. I sample the real data the system would work on - the messy version, not the cleaned-up one - and map the systems it has to connect to. With the people who know the work, I build the evaluation set: real cases with agreed correct answers, and the pass threshold the system has to meet, set before any results are seen.
Week two: the prototype and the numbers. I build the prototype on that data and score it against the evaluation set. The architecture document and the build estimate are written from what the prototype showed, not from the brief, and I walk you through all four deliverables at the end.
The evaluation set is the easiest deliverable to skip and the one that matters most. It is how you know whether the prototype worked, and later whether the build did. I hold my own products to the same standard: SureCiteAI, which I built as sole architect, publishes a 94% retrieval hit-rate across six benchmark suites and 0 hallucinated citations across 297 benchmark cases, with the runs on the proof page. For the method, see how to build an LLM evaluation set.
The price is fixed before work starts
The sprint is from £3,000. On the first call we agree what the sprint covers, and the fee is fixed in writing before any work starts. It does not change once the sprint is under way, whatever it finds, and it is not billed by the hour.
The build is a different matter: its cost is set by your data, the systems it has to integrate with, how it is evaluated and the security and compliance work around it, which is why it is quoted from the sprint rather than guessed before it. What drives the cost of an AI build explains each of those drivers.
What happens after the sprint
The sprint ends in one of two places.
- Build it. The build is quoted from the sprint’s fixed estimate, measured against the same evaluation set, and run in your own accounts with weekly demos. You can also take the architecture document and evaluation set to another builder; they are written so that works.
- Don’t build it. This is a valid outcome, not a failed sprint. Sometimes the data is not ready, an off-the-shelf tool already does the job, or the build would cost more than it saves. The architecture document says why, and you have paid for a sprint rather than a build to learn it.
After a build, some teams want someone to keep owning the architecture as the system grows: model changes, new use cases, the next hire. That is the fractional AI CTO engagement, and it is optional.
Already have a working system?
The sprint is for something not yet built. If yours already runs, start with a review of what exists instead.
Frequently asked questions
What is an AI discovery sprint?
A fixed-fee piece of work, 2 weeks long, that answers two questions before you commit to building an LLM, RAG or agent system: is it worth building, and what will it cost? It ends with four things: a working prototype on your real data, an architecture document, an evaluation set with a pass threshold, and a fixed build estimate.
How much does the discovery sprint cost?
From £3,000. The fee is agreed in writing before the sprint starts and does not change once it has, whatever the sprint finds. It is not billed by the hour. The build that may follow is quoted separately, from the sprint's own estimate.
Is the sprint a proof of concept?
It includes one, and then goes further. A proof of concept shows something can work. The sprint also measures how well it works on your data, against an evaluation set and a pass threshold agreed in advance, and turns that into an architecture and a fixed price for the build. A proof of concept on its own tells you little about cost.
Is the prototype production-ready?
No, and it is not meant to be. It is built to answer one question quickly: does this approach work on your data, measured against the evaluation set. Production adds what a prototype skips - access control, tenant isolation, monitoring, audit logging, failure handling and behaviour under load. That work is what the build estimate prices.
What if the answer is not to build it?
Then that is the result, and it is a valid one. Sometimes the data is not good enough yet, an off-the-shelf tool already does the job, or the build would cost more than it saves. The architecture document says why, and you have spent the price of the sprint rather than the price of a build to find out.
Do we have to build with you afterwards?
No. The deliverables are yours whether or not you go ahead, and the architecture document and evaluation set are written so another team could build from them. The sprint is priced to stand on its own.
What do you need from us?
A sample of the real data the system would work on - the messy version, not the cleaned-up one - access to the systems it would need to connect to, and time from someone who knows the work well enough to say what a correct answer looks like, because they help build the evaluation set. I will sign your NDA before seeing anything.
We already have a working AI system. Is this for us?
Probably not. The sprint is for something not yet built. For a working LLM, RAG or agent system, the AI security and architecture review (from £3,500) checks it before it meets real customers. For an app built with AI coding tools such as Lovable or Cursor, the AI-built app review (from $1,300, or £1,000) is the lighter fit.
Further reading
- How much it costs to build a production AI system - the cost drivers behind the build estimate.
- How to vet an AI consultant - the questions to ask whoever runs your sprint, including me.
- Build vs buy AI - when an off-the-shelf tool is the better answer.
- The consulting guide and AI consulting services - everything else on scoping and running an AI engagement.
Deciding whether to fund an AI build?
Tell me what you want the system to do and what data it would work on. If a sprint is not the right next step, I will say so and tell you what is.
Or describe your project in the chat, or email nic.chin@nicchin.com
Last updated