AI integration services that connect your business systems to Claude and ChatGPT
AI integration services connect a language model to the systems where a company's data lives, so staff can ask questions and get answers from real records. Goldenroz builds custom MCP servers and installs an on-prem MCP gateway that links ERP, CRM, SQL and Microsoft 365 systems to Claude or ChatGPT. Access is read-only, limited per user and logged. Projects start at $25,000, fixed price per phase after a scoping call.
From $25,000, fixed price per phase
Updated 2026-11-13. By Shai Goldenberg

Who this is for
- Operations, finance and sales leaders whose teams wait on analysts for answers that already sit in the ERP or CRM.
- CIOs and CTOs who must give staff AI access to company data without sending the database to a vendor.
- Companies with an on-prem ERP or SQL Server that cloud AI tools cannot reach.
- Product teams that want their own product to be usable from Claude or ChatGPT through an MCP server.
Who it is not for
- Budgets under $25,000.
- Letting a model write to production systems on day one. We start read-only.
- Giving the model a database password and free-form SQL. We do not build that.
- Teams with no clear owner for the data. Start with AI consulting.
What is included
- System and data mappingA list of the systems in scope, the questions people need answered and the tables or endpoints that hold the answers.
- MCP gateway or custom MCP serverOur on-prem gateway for standard systems, or a custom MCP server written for an internal system or your own product.
- Approved queriesA reviewed catalog of parameterized queries. The model picks a query and fills in values. It never writes SQL.
- Per-user permissionsEach person sees through the model only what they may already see in the source system.
- PII maskingPersonal fields are masked or removed before results reach the model, by rules you approve.
- Audit logA record of who asked what, which query ran, and what was returned.
- Prompt-injection defensesData returned from your systems is treated as untrusted text and cannot change what the model is allowed to do.
- Deployment and handoverInstalled with Docker Compose inside your network, with documentation for your IT team.
How the work runs
Scoping call
Before the projectWe list the systems, the users and the first questions worth answering. You get a fixed price for phase one.
First connector
Phase 1One system connected read-only in a test environment, with a small catalog of approved queries and real user roles.
Security review
Before productionYour IT or security lead reviews permissions, masking rules and the audit log before any production data is connected.
Production install
Phase 2The gateway runs on your infrastructure. A pilot group uses it from Claude or ChatGPT while we watch the logs.
More systems
Later phasesFurther connectors and queries are added one system at a time, each as a fixed-price phase.
MCP integration: what a custom MCP server does
MCP stands for Model Context Protocol. It is an open standard that describes how an AI application discovers and calls external tools and data sources. An MCP server exposes a set of tools, such as a customer lookup or an open-invoice report, and the model calls them during a conversation.
The value of MCP server development is that one integration works across clients. A server built for your ERP can be used from Claude, from ChatGPT and from other MCP clients without writing a separate plugin for each.
A custom MCP server is the right choice when the system is internal, when your product needs its own MCP endpoint for customers, or when the tool list has to match a specific workflow. For common business systems, our gateway already covers the connection and the work is configuration and query design.
Enterprise MCP: connecting ERP, CRM and SQL safely
The risk in LLM integration services is the access path, more than the model. A model that can run any SQL can read any table, and a model that reads a poisoned record can be told to do something else. Both risks are handled in the gateway, outside the model.
Our gateway is read-only. Queries are written and approved in advance, so the model chooses from a fixed list and supplies parameters. Permissions are checked per user on every call. Personal data is masked before it leaves your network, and every request is written to the audit log.
It installs on-prem with Docker Compose. Your database stays inside your network, and credentials for source systems are never sent to the model provider. Only the results of approved queries, after masking, are sent as context for the answer.
Beyond read-only: LLM development for actions
Read-only access answers a large share of everyday questions: order status, stock, receivables, pipeline. It is also the safe place to learn how your staff use the tool.
When a team is ready for the model to create or change records, that is an agent project with approval steps and its own scope. See AI agent development. Customer-facing conversations are covered under AI chatbot development.
What backs this up
- An MCP gateway we ownShai built and owns an on-prem MCP gateway product: about 27,000 lines of code and 13 connector types, built in about two weeks.
- Security controls already builtThe gateway ships with per-user permissions, PII masking, an audit log, approved queries only and prompt-injection defenses.
- Database security in productionOn a separate production platform he maintains about 145 tables under Row Level Security with 155 policies, and tenant isolation enforced in database triggers.
- A named engineerShai, our CTO, does the integration. You own the configuration, the query catalog and any custom code.
What it costs
Each phase has a fixed price agreed before work starts. The price depends on the number of systems, the number of user roles and whether a custom MCP server is needed.
- AI integration projectFixed price per phaseFrom $25,000
- Larger multi-system platformsScoped per project$100,000 and up
- Scoping callEnds with a written scope and a price$0
- Model subscriptions and usageYour own Claude or ChatGPT accountsPaid to the provider
Who does the work
Shai leads the technology side: business software, SaaS and mobile apps, payments and escrow systems, Postgres security, AI agents built with Claude Code and MCP, and IT and infrastructure projects with senior infrastructure engineers.
Questions buyers ask
What are AI integration services?
They are the engineering work of connecting a language model to business systems such as an ERP, CRM, database or document store. The result is that staff ask a question in Claude or ChatGPT and get an answer from live company records, within their own permissions.
What is an MCP server?
An MCP server is a small service that exposes tools and data to AI applications through the Model Context Protocol, an open standard. One server can be used from several AI clients, including Claude and ChatGPT.
Does our data leave our network?
The gateway runs on your infrastructure and your databases stay there. Only the results of approved queries, after PII masking, are sent to the model as context. Source system credentials are never sent to the model provider.
Can the model change or delete our data?
No. The gateway is read-only and runs only queries that were written and approved in advance. Write actions are a separate agent project with human approval steps.
Which systems can you connect?
ERP, CRM, SQL databases and Microsoft 365 are covered by the gateway, which has 13 connector types. Internal or unusual systems get a custom MCP server. The scoping call confirms what your systems expose.
Does it work with both Claude and ChatGPT?
Yes. Both support MCP connections, so the same gateway serves either. Which plan and client you need depends on your provider agreement, and we check that during scoping.
How much does AI integration cost?
Projects start at $25,000 and are fixed price per phase after a scoping call. Larger multi-system platforms run $100,000 and up. Model subscriptions are paid to the provider from your own account.
How do you handle prompt injection?
Text returned from your systems is treated as data and never as instructions. The model can call only the approved read-only queries allowed for the current user, so a malicious record cannot widen its access. Every call is logged for review.
AI integration: by industry and use case
Guides on this topic
- An AI agent takes actions across your systems. A chatbot answers questions. Compare the two, see examples and use cases, and use the guide to pick one.
- Compare Claude and ChatGPT for company use: connectors, MCP, coding tools, admin controls and data policies. Use the checklist to run your own evaluation.
- What an AI agent costs to build and to run: the five cost drivers, project shapes by effort, and a worked running-cost example. See where your project lands.
Related services
- Custom AI agents that take actions in your systems, with permissions, approvals and logs.
- An assessment that ranks your AI use cases and ends with a build-vs-buy call and a pilot plan.
- Support and ordering bots for WhatsApp and web, with fallbacks and handoff to your team.
- A code and database security audit for apps built fast on Supabase and Postgres, with fixes as PRs.
Talk to the person who will do the work
One hour, no sales team. You talk to Shai Goldenberg, the CTO, about your systems or your product, and you leave with a clear next step and a price range.