What is MCP and what does it give us?

MCP is an open standard for connecting AI clients to external systems. Instead of writing a separate integration per tool, it exposes your systems through one standard interface, and authorisation boundaries are defined at that interface. In practice it means every operation the model can reach is named individually and can be audited.

Does our data go into training the model?

Not in what we build. Enterprise provider agreements exclude data from training, and we put that in writing at the start of the project. Where data must not leave your premises at all, we set up a self-hosted model instead.

What if the model produces something wrong?

No method removes that entirely; a structure that reduces and catches it does exist. Answers are produced only from supplied data, sources are cited, missing information is stated as missing, output passes shape and business-rule validation, and irreversible actions require human approval. The evaluation set catches regressions before a version ships.

Do we have to replace our ERP?

No. This layer sits on top of existing systems and connects to them. Its entire purpose is to make working systems reachable without replacing them.

Two lines: development, and per-use model cost. A pilot at your real volume produces a monthly estimate. In a self-hosted setup the second line becomes hardware and operations instead.

Where should we start?

With a single process. An AI readiness assessment sets out in writing which process is suitable, whether the data you hold is sufficient, and what you should not attempt.

Solution

AI Layer

This is the layer that takes AI out of the demo and into production.

48 functions8 module groups6 integrations

Most AI projects die in the demo. The model is not the reason. Where nobody has defined how it connects to company data, what it may not say, how its output is measured and how access is recorded, there is nothing left to turn a working demonstration into a system somebody signs for. We build this layer as the answer to those questions. The model provider can change; the connection, the limits, the measurement and the record stay with you.

Scope and modules48

Connection and access

MCP server: company systems exposed to the model through one standard interface
Tool definitions: every operation the model may call, named individually
Authorisation boundary: which role reaches which data and which action
Read-only and write-capable tools kept apart
Authentication and session-scoped access
Rate limiting and abuse protection

Grounding in your own data

Ingestion: contracts, procedures, technical documents and record sources
Chunking and embedding scheme
Semantic and keyword search working together
Citation: every answer states the document it came from
Refresh flow: the index updates when a document changes
Access rights enforced in the retrieval layer as well

Agents and execution

Multi-step task execution with an explicit plan
Tool use: opening a record, creating an order, producing a report
Steps that require human approval, named upfront
Rollback and compensating actions
Queueing, retries and failure handling
Progress reporting on long-running work

Policy and limits

Forbidden topics and forbidden forms of steering
An obligation to speak only from the data supplied
Missing information written as missing rather than invented
Output validation: shape, range and business rules
Personal data masking and confidential field filtering
Fallback behaviour when a limit is hit

Evaluation and quality

An eval set built from real questions, held fixed
Regression checks so a prompt or model change cannot lower quality
Human review and labelling flow
A/B comparison and per-version scoring
An acceptance threshold below which nothing ships
Failed outputs collected and folded back into the test set

Observability and cost

Traces: which question, which context, which tool, which answer
Audit record and retention period
Token and cost tracking, cost per unit of work
Latency measurement and identification of slow steps
Usage reporting by team, user and process
Alert thresholds on cost, error rate and latency

Control surface

The system prompt edited and versioned from the admin panel
Version history and rollback
Model routing: a different model per task, balancing cost and quality
Failover to a secondary model when a provider degrades
Knowledge base maintained by editors
A staging environment for trying a change before it ships

Data protection and compliance

Data excluded from model training
Data residency: where a request goes and where it is stored
Processing inventory under KVKK and GDPR
A self-hosted model option
Personal data retention and deletion periods
Confidentiality terms and sub-processor transparency

What it connects to06

ERP and accounting systems
CRM and service desk
Document management and file stores
Data warehouse and reporting tools
Email, calendar and messaging
Identity providers and authorisation systems

Signs you need this system07

01The AI pilot worked and never reached production
02The model does not know your own documents and answers generically
03Nobody measures whether the output is correct
04There is no record of who reached what
05Monthly model cost cannot be predicted
06Legal and IT are asking where the data goes
07Changing a prompt means waiting for a developer

Decisions to make upfront05

01Where will the model run?: Through a provider or on your own servers. That one decision sets cost, latency and data residency at the same time, and it is usually the first question an enterprise buyer asks.
02Will the model only read, or also act?: Reading is comparatively simple. The moment the model is to open a record or start a transaction, authorisation, approval and rollback design become most of the work.
03Which source counts as the truth?: In a company where the same fact sits differently in two places, the model will produce two different answers. The hierarchy of sources has to be settled before grounding is built.
04How will accuracy be measured?: Without measurement, a quality discussion becomes a matter of opinion. An evaluation set is built at the start and every version is scored against it.
05Which steps need a human?: Every irreversible action — money moving, something sent outside, a deletion — needs approval. That boundary is written at the start of the project.

Sectors we have delivered in02

Finance and Investment B2B and Dealer Systems

Related work04

Frequently asked06

What is MCP and what does it give us?: MCP is an open standard for connecting AI clients to external systems. Instead of writing a separate integration per tool, it exposes your systems through one standard interface, and authorisation boundaries are defined at that interface. In practice it means every operation the model can reach is named individually and can be audited.
Does our data go into training the model?: Not in what we build. Enterprise provider agreements exclude data from training, and we put that in writing at the start of the project. Where data must not leave your premises at all, we set up a self-hosted model instead.
What if the model produces something wrong?: No method removes that entirely; a structure that reduces and catches it does exist. Answers are produced only from supplied data, sources are cited, missing information is stated as missing, output passes shape and business-rule validation, and irreversible actions require human approval. The evaluation set catches regressions before a version ships.
Do we have to replace our ERP?: No. This layer sits on top of existing systems and connects to them. Its entire purpose is to make working systems reachable without replacing them.
How is it costed?: Two lines: development, and per-use model cost. A pilot at your real volume produces a monthly estimate. In a self-hosted setup the second line becomes hardware and operations instead.
Where should we start?: With a single process. An AI readiness assessment sets out in writing which process is suitable, whether the data you hold is sufficient, and what you should not attempt.

Other systems08

ERP — Enterprise Resource Planning CRM — Customer Relationship Management B2B Dealer and Ordering Portal WMS — Warehouse and Stock Management Production Tracking and MES Field Service Management Integration and Middleware Document Processing and Automation

AI glossary Readiness assessment Why pilots stall

Let’s scope it together.

We do not quote a fixed price for undefined scope. We review your existing system and processes on site, then deliver the roadmap and a fixed-price proposal.

Get in touch Technical assessment How we work