AI systems engineered around your operation.
Not a model, and not an agent. The foundation underneath: knowledge, memory, model routing, boundaries and a record of everything that happened.
What is an AI system?
An AI system is the engineered foundation everything else runs on. Not the model, and not the agent. The layer underneath that holds your knowledge, decides which model handles what, remembers context, enforces boundaries and records what happened.
- Knowledge
- Your documents, policies and records made retrievable at the moment of the question, so answers come from what your business holds rather than what a model guessed.
- Memory
- State that persists across a case and across sessions, so the system does not start from nothing every time somebody speaks to it.
- Model access
- A routing layer that sends each task to the appropriate model on cost, latency and capability, instead of paying premium rates for work a smaller model handles fine.
- Boundaries
- What may happen automatically, what needs a person, and what must never happen. Enforced in code, not in a prompt. How we handle credentials, least privilege and permission boundaries is set out in our security approach.
- Observability
- Every decision logged and traceable, with a frozen evaluation set that catches drift before your customers do.
The distinction that matters commercially: a model is something you rent. An agent is something that acts. A system is the thing that makes the agent safe to run in your business.
What goes into a real one?
These are the parts we build. Most deployments need four of them. Nobody needs all of them on day one.
RAG knowledge bases
Retrieval over your own material, chunked and indexed so the right passage surfaces at the right moment. The difference between a system that knows your business and one that improvises confidently.
Model routing
Cost and capability routing across models. A classification task does not need the same model as a legal summary, and paying as though it does is the most common source of runaway AI spend.
Guardrails and evaluations
Input and output checks, permission enforcement, and a frozen test set run on every change so regressions are caught before deployment rather than by a customer.
Self-improving loops
Systems that accumulate memory and improve from their own operation, within boundaries you set and with changes you approve, rather than ones that shift silently.
Integration layer
Connections to the CRM, inbox, calendar, phone and document systems you already run, so output lands where work actually happens.
Monitoring and audit
What the system did, when, on whose authority, and what a person approved. Built in from the start because retrofitting it is painful and sometimes impossible.
How is an AI system built?
Architecture before code. The expensive rescues we get called into almost always began with a tool choice rather than a process map.
- Map the operation. What actually happens, including the exceptions everyone has quietly worked around for years.
- Decide the boundary. Autonomous, assisted, or human only. Written down and agreed before anything is built.
- Build the foundation. Knowledge, memory, routing and guardrails first. Agents and workflows come after, because they depend on it.
- Instrument everything. Logging, tracing and an evaluation set, so behaviour can be proven rather than assumed.
- Deploy narrow. One process, live, measured. Scope widens once the thing has earned it.
Bespoke system, platform, or off-the-shelf?
Three genuinely different purchases that get sold with the same vocabulary. Choosing wrong is the most common reason an AI programme quietly stops.
| Off-the-shelf SaaS | Platform build (Copilot Studio, low-code) | Bespoke AI system | |
|---|---|---|---|
| Fits your process | You fit its process | Partly, within the platform | Built to it |
| Connects to legacy tools | Rarely | Sometimes | Yes, that is the job |
| Where the data sits | Vendor | Vendor | Yours, self-hosting available |
| Audit trail depth | Whatever is offered | Platform level | Designed to your obligation |
| Cost shape | Per seat, forever | Per seat plus consumption | Build cost, then low running cost |
| You can leave | Lose everything | Lose the build | System stays with you |
| Best when | Process is standard | Microsoft estate, moderate complexity | Process is specific or consequential |
When you should not build one.
We turn work away on this basis and would rather say it here than on a call.
If your process is standard enough that software already does it, buy the software. It will cost a fraction and it will work on Monday. If you already run everything in Microsoft and your requirements are moderate, Copilot Studio is a reasonable place to start and we will tell you so. If nobody in the business can describe the process the same way twice, no system will hold, and the honest first step is fixing that rather than buying anything. On the other hand, where the process is specific to you, touches systems that were never designed to talk to each other, or carries consequences a regulator would ask about, off-the-shelf will not reach it and platform builds tend to run out of room about six months in.
Questions we get every time.
What is an AI system, and how is it different from an AI model?
A model is a general capability you rent from a provider. An AI system is the engineered layer around it: your knowledge made retrievable, memory that persists, routing between models, permission boundaries enforced in code, integrations with your existing tools, and logging that records what happened. The model is one component of the system, not the system itself.
What is the difference between an AI system and an AI agent?
The system is the foundation. The agent is what runs on it. An agent pursues a goal and takes actions, while the system provides the knowledge, memory, model access, guardrails and audit trail that make those actions safe and repeatable inside a business.
Is a bespoke AI system better than Microsoft Copilot Studio?
Not automatically. If your organisation already runs on Microsoft and the requirements are moderate, Copilot Studio is a sensible starting point and cheaper to reach. Bespoke engineering earns its cost when the process is specific to your business, when it must reach systems without modern APIs, when data cannot leave your estate, or when a regulator will one day ask exactly what happened and when.
Where does our data live?
On our managed infrastructure by default, or on your own infrastructure where the engagement covers self-hosting. Your data remains yours in every case, and it is never used to train anything for anyone else.
How long does an AI system take to build?
The foundation for a single operational area is typically a matter of weeks after the architecture and diagnosis phase. Multi-system deployments and anything requiring regulatory evidence take longer, because audit trails and approval gates are engineering rather than configuration.
How much does a custom AI system cost in the UK?
FlowNest builds start at £3,500 plus VAT for a single system with up to three integrations, £7,000 for connected multi-step work, and £15,000 where multi-agent architecture or regulatory evidence is involved. Model, infrastructure and third-party licence costs are passed through at cost and shown separately before approval.
Can we run the system ourselves afterwards?
Self-hosting is available on the Enterprise package. Complete source-code and IP transfer is available through Full Handover. On Starter and Premium, FlowNest hosts and operates the system while you hold a permanent licence to use it within your business.
Bring us the process, not the brief
Thirty minutes. Describe what happens today and we will tell you whether this is the right shape of solution. You leave with a written diagnosis whether or not we work together.