The AI Employee. Connect an autonomous agent to your live ERP, CRM, or HR system. It reads your data, runs your checks, and writes draft records back. Sensitive data is masked locally before anything reaches the AI. Already running in production. Already catching what manual processes missed. Open source. Zero installs. Any system with an API.
MCP Enterprise Kit
Give an AI agent access to your live enterprise system. It reads your data, understands your rules, and runs your workflows autonomously.
Not a chatbot. Not a dashboard. An AI employee that operates inside your ERP, CRM, HR system, supply chain platform, or any software with an API.
What This Actually Is
When people say "AI will replace jobs" they usually mean chatbots answering questions. This is different. This is an AI agent that connects directly to your enterprise system, applies your organisations specific business rules, and executes the same workflows your team does manually. Every day. In three minutes. With zero installs on a locked-down corporate machine.
Built and tested in production against a live enterprise finance system in UK public sector. First scan found duplicates worth millions that no spreadsheet, formula, or manual review had ever flagged.
How It Works
Your Enterprise System --> MCP Server (your machine) --> AI Agent --> AI Provider
ERP / CRM / HRIS PowerShell (Windows) Claude Code (inference)
The AI agent calls tools you define. Each tool makes an API call to your system and returns formatted results. The AI reasons about the results and tells you what it found. It can also write back to create draft records for human approval.
The AI never has direct access to your system. It can only do what your tools allow. You control the boundary.
What Makes This Different
Zero dependencies. Runs on PowerShell 5.1 (pre-installed on every Windows machine). No Python. No Node.js. No Docker. No admin rights. No IT ticket. No app registration. Uses your existing corporate credentials via OAuth2 device code flow.
Domain intelligence, not generic queries. Every existing AI-to-ERP integration exposes generic CRUD: list customers, update records, show orders. This kit encodes YOUR business rules into the tools. Your exclusion logic. Your validation checks. Your approval workflows. Your reconciliation processes. Thats the difference between a data access layer and an AI employee.
It writes back. This is not a read-only reporting tool. The AI creates records, populates journals, stages transactions, and drafts entries directly inside your live system. Draft records require human approval before posting. The AI does the work. The human signs off. End to end.
Pattern detection beyond formulas. The AI catches what rule-based systems structurally cannot. Same vendor, same reference, same amount, same date, different transaction ID. No VLOOKUP, INDEX/MATCH, or SQL query finds this because there is no single field to match on. The AI scans thousands of records and flags the pattern.
Use Cases (Any System, Any Industry)
Finance / ERP
- Duplicate payment detection across staging tables
- Payment run automation (query, validate, batch, create draft journals)
- VAT/tax workings and monthly calculations
- Month-end readiness checks
- Bank reconciliation matching
- Aged debt analysis with trend commentary
HR / People Systems
- Starter and leaver checklist automation
- Absence pattern detection
- Payroll validation before submission
- Contract renewal flagging
- Training compliance gap analysis
Supply Chain / Procurement
- Purchase order duplicate detection
- Supplier spend analysis and anomaly flagging
- Goods receipt matching against orders
- Contract expiry monitoring
- Price variance detection across suppliers
CRM / Sales
- Pipeline hygiene checks (stale deals, missing fields, orphaned contacts)
- Quote-to-order reconciliation
- Customer credit limit monitoring
- SLA breach detection
- Renewal risk scoring
IT / Service Management
- Ticket pattern detection (same issue, different reporter)
- SLA compliance monitoring
- Change request impact analysis
- Asset lifecycle alerting
- Licence utilisation reporting
Any System With an API
If your system has an API (REST, OData, GraphQL, SOAP), this pattern works. The AI agent does not care what the system is. It cares about the data structure and the rules you encode into the tools.
Quick Start
git clone https://github.com/adriangziut/mcp-enterprise-kit.git
- Edit
enterprise-mcp.ps1— update Section 1 (your system URL, tenant, credentials) and Section 2 (your field names) - Save as UTF-8 with BOM
- Add to VS Code MCP settings
- Open Claude Code and type your first prompt
First run asks you to authenticate once. After that it auto-refreshes silently.
Security
| Concern | How Its Handled | |---------|-----------------| | System access | Uses YOUR existing credentials. Same permissions as your browser session. Nothing more. | | PII exposure | Masked at tool boundary. Names become Vendor-0001. Bank details never queried. | | Data safety | Reads from staging/buffer tables, not production ledgers. Writes create drafts requiring approval. | | Token storage | Plain text on disk (same risk as browser cookies). Revoke via identity provider if compromised. | | Network | No ports opened. No endpoints exposed. MCP runs as local subprocess via stdin/stdout. | | AI provider | Data transits during inference. Use zero data retention plans for sensitive environments. |
Full security documentation in docs/SECURITY.md.
File Structure
mcp-enterprise-kit/
SKILL.md Claude Code skill file
enterprise-mcp.ps1 MCP server (configure for your system)
CLAUDE.md Daily prompt template
docs/
SETUP.md Step-by-step setup guide
SECURITY.md Security architecture
examples/
tools.ps1 Finance, HR, and generic tool examples
dashboard/
README.md HTML dashboard documentation
LICENSE
README.md
The Point
This is what "AI replacing jobs" actually looks like. Not a chatbot answering questions. An AI agent that logs into your system every morning, runs the checks your team runs, catches what they miss, writes the results back as draft records, and waits for human approval. The team reviews and signs off. The AI did the work.
Three minutes instead of three hours. Every day. On any system.
License
MIT License. See LICENSE.
Author
Adrian Gziut — Software & AI Consultant
- Email: gziutadrian@gmail.com
- LinkedIn: linkedin.com/in/adrian-gziut-3631ab13b
- Website: adrian-gziut.me
DM me on LinkedIn, email gziutadrian@gmail.com, or WhatsApp +44 7897 308 141.