Workshop · For systems, data, finance and IT
The Technical & Data Architecture Workshop
Not to turn your team into software engineers, but to give them the confidence to build safely themselves: a target architecture for your systems, live worked examples and a hands-on build on a practice stack.
- Length
- One morning, about 5 hours
- Who
- Systems and finance analysts, operations, sales, the CFO, your IT partner
- Led by
- Martin Kolar
- Best after
- AI Roadmap workshop
The recommended architecture
Five layers. Your core systems stay exactly as they are.
AI reads your data through documented, read-only views that a data owner has approved. Every request carries the user's own permissions. Shared build standards mean staff-built tools all follow the same rules. Writing back to core systems is "prepare and approve": the tool drafts the change and a named person approves it.
- 5 · Staff build and use toolsteam AI plan, shared projects
- 4 · Shared skill filebuild, naming, storage and deployment standards
- 3 · Permissioned MCP serverevery call carries the user's own credentials
- 2 · Read-only SQL viewsdocumented, approved by a data owner
- 1 · Existing systems, unchangedERP · CRM · configurator · files
The agenda
The morning, session by session.
A typical 08:00 start. Slides and demos are tailored to your systems, and we run the worked examples on a practice copy of an ERP, never on your live data.
- 08:00
Opening and recap
What we heard in discovery, and the top use cases from the leadership workshop. The use-case board becomes a living tracker of what is built, wanted and blocked.
- 08:30
The recommended architecture
The five layers, applied to your systems, and the "prepare and approve" rule for writing back to the ERP.
- 09:15
Three live worked examples
Built live on a practice ERP: an orders-by-customer page; a stock and cost lookup that masks fields by role; and a dashboard request correctly refused by permissions, which shows why instructions in a prompt are not enough.
- 10:00
Terminology for non-engineers
ODBC, APIs, command-line tools, MCP, skills, projects, scheduled tasks, agents, SQL views and SQLite, explained in plain English, and what can safely be shared in chats, code, artifacts and projects.
- 10:30
Security and access options
Team plan versus cloud-hosted versus self-hosted models; what auditors need (a written policy, reused access controls, central logging); four ways to connect AI to your ERP, compared; and connecting email, spreadsheets and export-only systems.
- 10:50
The data storage decision
Buy a data warehouse, build a custom store, or use what you already have. We work through query load, refresh cadence and how live the data really needs to be.
- 11:15
Build recipes for your use cases
"Could you build this today?" for each top use case; a live buy-or-build decision; and how to run AI building without an IT department: a stack owner, builders, reviewers, your IT partner for infrastructure and a fractional AI lead. Your IT partner joins for the hosting discussion.
- 11:45
Hands-on: build it yourselves
With step-by-step instructions: create a shared team project; connect the practice ERP connector, sign in under a role and see permissions in action; set up a scheduled task; then download the skill file and build a tool.
- 12:50
Close: the governance model
Read-only views; every write owned by a person; every tool with an owner, a backup, a register entry and a README for handover.
Six principles
What we have learned building AI in operational businesses.
- 01Permissions belong in the infrastructure, not in the prompt.
- 02Every tool has a named owner.
- 03Start with drafts, not sends.
- 04Don't buy a data warehouse until you need one.
- 05The real cost is people's time, not the software.
- 06You can run this without an IT department, but not without an owner.