Turn an AI prototype into a system you can rely on
Many AI projects work in a demo and stall before real customers use them. We take an AI prototype, pilot or idea and build it into a production system: connected to your data, with guardrails, testing against real cases, human handover, monitoring, data protection in place and documentation your team can run.
Service 09
No 17040048
Major build
From £25,000
Fixed quote after scoping
Scoped in writing with milestones. Larger systems are quoted on their own scope. Model, hosting and third-party costs are yours and estimated in the quote.
Who it is for
- Businesses with an AI pilot or prototype that works in a demo but is not ready for real customers or real data.
- Founders building an AI product who need it production-ready, not just impressive.
- Teams in sectors with sensitive data that need the data protection chain right from the start.
What the work covers
Architecture and scope before code
A demo proves an idea can work once. Production means it works on the thousandth conversation, on bad data, when a supplier is slow. We start by writing down what the system must do, what it must never do, which data it touches and where each part runs. If you have a prototype, we review it and say plainly what survives and what gets rebuilt. The written scope sets milestones you can check progress against, so the fixed quote means something and changes are visible rather than absorbed quietly into the build.
Grounding and guardrails
Most production failures come from an AI answering beyond what it knows or doing more than it was meant to. We ground answers in your approved sources and make the system cite them, so a wrong answer can be traced to a wrong source. Guardrails are built as code and configuration, not just instructions in a prompt: topics it will not discuss, commitments it cannot make, such as prices or refunds, and actions that need a person to approve them. Each rule is tested, so you know it holds rather than hoping it does.
Evaluation sets from your real cases
We collect real examples from your business: typical questions, awkward ones, edge cases and the requests the system should refuse. Each case has an agreed expected outcome. The set runs before every release, so a change to a model, a prompt or the code is checked against the same cases each time. This is how we can say what the system does on your work without quoting accuracy figures we have not measured. You keep the evaluation set, and it becomes the basis for monitoring once the system is live.
Human handover designed in
Every customer-facing system we build can pass a conversation to a person, with the context attached so nobody has to ask the customer to repeat themselves. We agree the triggers with you: a complaint, a clinical or financial question, low confidence, or a customer simply asking for a human. We also design what happens when nobody is available, because a handover that leads nowhere is worse than none. ReplyOS hands over on every conversation for this reason, and we bring the same thinking to client builds.
Security, access and many customers at once
An AI tool that reads your documents must respect who is allowed to see what. We connect it through proper authentication and apply your existing permissions, so the AI cannot show a member of staff something they could not open themselves. Where the system serves many customers, as a feature inside your product, each customer's data is kept separate at the data layer and tested, not just filtered in the interface. Keys, logs and audit trails are set up so you can see who did what, and when.
The data protection chain, then handover
We map what personal data the system sees, where it is processed, which suppliers and subprocessors are involved and how long anything is kept, in line with UK GDPR as amended by the Data (Use and Access) Act 2025. We work alongside your advisers and do not give legal advice. At launch, monitoring and logging are already running, documentation is written for the people who will operate the system, and we hand over to your team in working sessions. A review at 30 days checks how it behaves under real use.
What we typically build
These are illustrations of the kind of work, not client results.
A customer-facing assistant
Answers only from your approved sources, hands over to a person, never commits to prices or refunds, and records every conversation. Example only.
An internal AI tool used by staff
Reads your documents, cites where each answer came from, and keeps access in line with who may see what. Example only.
An AI feature inside your product
Built for many customers at once, with each customer's data kept separate and every change tested before release. Example only.
What you get
We choose models and infrastructure to suit your data rules and budget, and build so they can be changed later. Bespoke code written for you is yours, as set out in the contract.
- 01Architecture and a written scope with milestones, agreed before the build starts.
- 02The system built and connected to your data and tools, with authentication and access control.
- 03Guardrails: what the AI may and may not do, approval steps, and human handover.
- 04An evaluation set built from your real cases, run before every release.
- 05Monitoring, logging and an audit trail of what the AI did.
- 06Data protection in place: processing records, supplier and subprocessor chain, retention.
- 07Documentation, handover to your team, and a review after 30 days.
What changes
- An AI system real customers or staff use, not a demo that only works when someone is watching.
- Written rules for what the AI may and may not do, each one tested.
- An evaluation set from your own cases that shows what the system does before every release.
- A clear record of the data involved, the suppliers in the chain and how long data is kept.
- A team that can run the system, with documentation and an audit trail they understand.
How it runs
- A 20-minute fit call, then a scoping step agreed according to size.
- A fixed quote from £25,000, with milestones in writing.
- Built in milestones against your real cases, with a named owner on your side.
- Monitored launch, handover sessions and a review at 30 days.
- Ongoing support after the review is a separate retainer.
What you supply
- The prototype, pilot or brief, and the people who built or own it.
- Access to the data and systems involved, starting with test environments.
- A named owner who can make decisions quickly during the build.
Limits
- A fixed quote covers the agreed scope. New features are scoped and quoted separately rather than absorbed.
- We do not promise accuracy figures we have not measured. The evaluation set shows what the system does on your cases.
- Ongoing development and support after the 30-day review are a separate retainer.
How it works
- 01
Scope
A 20-minute fit call, then a paid or free scoping step depending on size. You receive a fixed quote with milestones.
- 02
Build
Built in milestones against your real cases, with guardrails and handover designed in from the start.
- 03
Test
The evaluation set run, failures fixed, and the system tested with real users before launch.
- 04
Launch and hand over
Monitored launch, documentation, handover to your team and a review at 30 days.
We built ReplyOS
Evidence from our own work
ReplyOS is our live AI platform for UK dental and aesthetics practices: enquiries across channels, AI-assisted replies under operator controls, human handover on every conversation, and the data protection chain a health-adjacent sector needs. We designed, built and run it. It is our own product, not a client project.
See ReplyOSWhere it fits
A diagnostic or control review often identifies the system worth building, and monitoring keeps it behaving as approved once it is live.
Questions before you start
How much does it cost to take AI to production?
From £25,000 as a fixed quote with milestones. Larger systems are quoted on their own scope, and model, hosting and third-party costs are estimated separately.
Can you work with a prototype someone else built?
Yes. We review it first and tell you what to keep, what to rebuild and why, before you commit to the build.
Who owns the system?
Ownership of bespoke work is set out in the contract, and normally sits with you. Third-party tools and models stay under their own licences.
How do you handle sensitive data?
Data protection is designed in from the start: what data the AI can see, where it is processed, which suppliers are involved and how long it is kept, agreed in writing.
What is the next step?
Tell us about the prototype or idea in the contact form. We arrange a 20-minute call and agree how to scope it.
How long does a production AI build take?
It depends on the scope, so we set the timeline in the written quote with milestones rather than guess before scoping. The main factors are how much of an existing prototype can be kept, how many systems it connects to, and how quickly your side can give access and make decisions.
Which AI model will you use?
We choose the model to suit your data rules, the task and your budget, and we build so it can be swapped later. Because the evaluation set runs before every release, a change of model is tested against your real cases rather than taken on trust.
Can our own developers work alongside you?
Yes, and it usually makes handover easier. We agree who owns which parts in the scope, review changes together, and run handover sessions so your developers can maintain the system, its guardrails and its evaluation set once we step back.
Tell us what needs to work better.
Describe the process and the tools involved. If it looks like a fit we will arrange a 20-minute call.