AI Systems Workshop for 9 Startups

I ran a 90-minute workshop with 9 startups from the Scale SEE program on scaling their organisation with AI.

I went in expecting to talk mostly about automation.

But the most useful part was seeing what people wrote down when asked: “What internal loop still depends on one person every week?”

The answers were simple, but real: invoices, customer support, onboarding, outbound, sales follow-up, lead follow-up, WhatsApp feedback becoming Jira tickets, code review requests becoming merge decisions, and content review needing product and technical context.

Different startups had different products, but a similar problem appeared.

Important context gets stuck: in someone’s head, in Google Drive, in Confluence, in emails, in ChatGPT or Claude threads, in WhatsApp groups, and in tools that not everyone reads.

Then one person has to translate it for everyone else.

Sales to technical. Customer feedback to product. Founder idea to execution. Technical limitation to client explanation.

During the session, teams mapped internal loops that still depended on one person every week.

• Invoices and accounting
• Customer support
• Onboarding
• Outbound
• Sales follow-up
• Lead follow-up
• WhatsApp feedback becoming Jira tickets
• Code review requests becoming merge decisions
• Content review needing product and technical context

The main framework was: Trigger > Source > Judgment > Output > Check.

This helped separate vague “we should use AI here” ideas from actual system design.

A useful AI system needs to know what starts the loop, where the source material lives, what judgment has to be made, what output should be produced, and who or what checks the result.

AI can help when context needs to move between people, tools, and decisions with less friction.

Not by magically replacing people.

But by helping the right context reach the right person, tool, or decision point at the right time.

A useful pattern emerged: before you automate, make the context clear enough that a team can use it.

One thing I liked from the workshop was that when teams were asked for the next move, nobody chose “automate now.”

Most chose either custom build or fix the loop first.

That felt like the right answer.

AI systems fail when teams skip the source layer and the human approval layer.

Before something can be automated or assisted, the team needs to understand the sources, owners, decisions, tools, and checks behind the work.

Sometimes the right answer is not “automate now.”

• Fix the loop first
• Buy a simple tool
• Custom-build only the part that needs custom logic
• Leave the human in the loop

This workshop is not a technical build case study.

It shows the discovery layer before a build.

• Mapping messy internal work
• Identifying AI-ready loops
• Separating automation hype from system design
• Translating business problems into possible AI architecture
• Helping teams decide what should be fixed, bought, built, or left alone

Teams were trying to understand where AI could help, but their internal work was scattered across people, documents, tools, messages, and approval steps.

I facilitated a workshop that mapped roles, internal loops, knowledge sources, judgment points, expected outputs, and human checks.

The session used Trigger > Source > Judgment > Output > Check to turn vague AI opportunities into concrete system-design questions.

The output was a clearer view of which loops were AI-ready, which needed better structure first, and where custom AI systems might make sense.