AI Training Week 4: Agents + Hackathon (Without Code)

If you're an owner reading this and wondering whether your team can really build "agents" without engineers — yes, they can, but only if week 4 is structured as a hackathon, not a tutorial.

Why agents belong in week 4 — not week 1, not week 6

Earlier than week 4, the team doesn't have the prompt judgment to design a multi-step flow without it collapsing. Later than week 4, momentum bleeds out and the program becomes performative.

Stanford's 51-deployment study found that escalation-routing AI (where the agent does the easy 80% and routes the rest) yielded around 71% productivity gain, while approval-routing (where every step waits for human approval) yielded only around 30%. Agents are the artifact that lets you cross from prompt-level uplift to workflow-level uplift — but only if the team has the judgment to draw the right human-vs-agent line. That judgment is what weeks 1-3 built.

Definition: Agent — a multi-step AI workflow that takes an input, runs through 2+ tool calls (read email, look up CRM, draft response), and produces an output without a human in every step.

What "no-code" actually means in week 4

The platforms that work for an SMB hackathon:

• Zapier AI / Make — best for cross-tool flows (Slack→CRM→Gmail→Sheets).
• Microsoft Copilot Studio — best when you're Office-anchored and the agent needs to live inside Teams/Outlook.
• n8n (cloud) — best when you want self-hosted control without writing code.
• Custom GPTs / Claude Projects — best for single-user agents (a sales rep's personal "outbound assistant").

Pick one platform per role-track, not one for the company. Each track gets the platform that fits the use case they committed to in week 2.

Definition: No-code agent platform — a visual builder where employees connect AI models, business apps, and triggers without writing programming code. Logic is configured in dropdowns and prompts.

How to structure week 4

The format that consistently produces shipped agents:

1. Monday — 60-minute kickoff. Champions demo two agents they built in advance. Not slides — live runs.
2. Tuesday — 90-minute platform onboarding per role-track. Champion walks the team through the platform. One platform per track.
3. Wednesday/Thursday — async build time (3-4 hours total per person). Pairs build their agent. Champion is on call in cohort Slack.
4. Friday — 90-minute hackathon demo. Each team shows their shipped agent, end-to-end, on real data.

Pairs, not individuals. The agent that survives the demo is almost always the one that was rubber-ducked with a partner during the build.

The hackathon brief (copy/paste)

Pick ONE use case from your role-track's week-2 backlog.
The agent must:

1. Have a real trigger (an email arrives, a form is submitted, a row is added).
2. Read at least one source of context (CRM, Sheet, calendar, knowledge base).
3. Use AI to draft, classify, or decide ONE thing.
4. Produce a real output that goes somewhere a human will see (Slack, draft email, ticket).
5. Have a human-in-the-loop step at the end. NO autonomous send to customers in week 4.

Build with your assigned platform. Pair with one teammate.
Friday demo: 5 minutes. End-to-end. Real data. No slides.

The "no autonomous send to customers" rule is non-negotiable in week 4. That's a week-5 conversation (Responsible AI), not a week-4 conversation.

Tool tip (Course for Business): The reason a no-code hackathon works in week 4 is that Augment, don't replace is now a built habit, not a slogan. By week 4 the team understands which steps need human judgment and which don't — they design the human-in-the-loop step naturally. The 6-week program at https://course.aiadvisoryboard.me/business runs the agents hackathon as paired builds with mandatory human-in-the-loop, exactly because un-paired builds collapse 70% of the time. (Course for Business)

What a week-4 agent actually looks like

Real examples from cohorts of 30-500-employee companies (anonymized):

Sales — Lead-research agent (built on Zapier). New lead enters HubSpot → agent reads website + LinkedIn → drafts a 3-bullet briefing into Slack DM to the AE → AE reviews and sends. Time-saved: ~25 minutes per lead.

Operations — Vendor-email triage (built on Make). New email to ops@ inbox → agent classifies (invoice / inquiry / urgent / spam) → drafts reply for invoice/inquiry → posts to a Slack channel for human approval. Time-saved: ~3 hours/week per ops person.

Finance — Variance explainer (built on Copilot Studio). Monthly close uploads to SharePoint → agent reads variance lines vs prior month → drafts 1-paragraph explanation per line → finance lead reviews and approves. Time-saved: ~5 hours per close cycle.

CS — First-response drafter (built on Custom GPT + Zendesk). New tier-1 ticket → agent looks up customer history → drafts a first reply → CS rep reviews and sends. Deflection improvement seen at peer companies has been around 30% in 4-6 weeks (Intercom Fin pattern).

Good vs bad week-4 agents

Bad: "An agent that sends customer emails autonomously." Good: "An agent that drafts the email, posts the draft to Slack, waits for thumbs-up, then sends."

Bad: "An agent that does everything in our sales process." Good: "An agent that does step 3 of our sales process — research the lead — and hands off to the AE."

The good versions are scoped to one step, with a clean human handoff.

Team scan (what AI champions report after week 4)

• Most cohorts ship 60-80% of the agents they start. The other 20-40% die on data-access issues, not AI issues.
• The agents that survive Friday demo and stay running by week 6 are almost always the ones with a thumbs-up-in-Slack step.
• Engineering teams want to skip no-code and write Python. Hold them to no-code in week 4 — week 6 is for code volunteers if they want.
• Champions report that pairs ship roughly 2x more often than solo builders.
• The most common failure pattern is over-scoped ambition: "let's automate the whole sales cycle." Champion's job is to scope it down on Tuesday.
• Around 1 in 4 agents reveals a data-pipeline problem (CRM hygiene, missing fields) that has to be fixed before the agent ships.
• Sales and CS agents typically ship fastest; finance agents take longer because of data-access reviews.
• Leaders report that Friday demos build more cross-team momentum than any all-hands.
• The cohort Slack peaks in volume during week 4 — peer-to-peer help is what makes the platform learning curve survivable.
• Cost surprise: most week-4 agents run for under $30/month per role-track at SMB volumes.

Micro-case (what changes after 7-14 days)

A 160-person B2B SaaS company I advised ran week 4 across four role-tracks, paired-build format. By Friday, 11 of 13 paired teams had a working end-to-end agent demoed on real data. By day 14, six of those eleven were running in production with daily use. The CS first-response agent was the standout: 84% of tier-1 tickets received a draft within 30 seconds of arrival, the rep reviewed and sent — saving an estimated 70 person-hours per month across the team. The CFO, who had blocked a similar effort in 2024 ("we'll wait for vendors to build it"), approved an expansion budget by day 17. Compare to a peer firm that ran week 4 as a vendor demo session: zero agents shipped, $80k spent on consultant decks.

Note on this case: This example is illustrative — based on typical patterns we observe with companies of 30-500 employees, not a single named client. Specific numbers are rounded approximations of common ranges, not guarantees.

Tool tip (Course for Business): The single biggest week-4 anti-pattern is letting one engineer-on-the-team try to build agents in code while the rest of the cohort uses no-code. It splits the cohort into two classes — the "real builders" and the rest — and silently kills AI Champions (1:15-20) dynamics. The 6-week program at https://course.aiadvisoryboard.me/business keeps everyone on no-code through week 4 by design — code volunteers can fork into a track in week 6. (Course for Business)

FAQ

Do we need to buy enterprise licenses for the agent platforms? Most platforms have team plans under $300/month for week-4 build volume. Don't enterprise-procure in week 4 — use team plans, validate, then negotiate enterprise in week 6 or later.

What about data privacy in agent builds? Use the same data-handling rules from week 3's tool deep-dive. If a tool isn't sanctioned for your data class, the agent doesn't see that data. Champions enforce in the build review.

Should agents be allowed to send messages to customers? Not in week 4. Human-in-the-loop is mandatory. Autonomous customer-facing comes after the Responsible-AI conversation in week 5. (We separately have an advisory product on the daily-management side, but that's a different scope.)

What if a team can't get their agent to work by Friday? That's a valid demo. They show the half-built flow, name the blocker, and the cohort helps debug. Half-built agents that get unblocked over the weekend ship more often than "perfect" agents demoed on Friday.

How do we keep the agents running after week 4? Each role-track names an agent owner — usually the champion. Week 5 includes a 30-minute "agent health" check-in. Don't skip it.

Conclusion

Week 4 is where the program leaves prompt-land and enters automation-land. No-code platforms, paired builds, scoped use cases, mandatory human-in-the-loop. By Friday, the company has shipped real agents on real data — usually for the cost of a few coffees. That changes the conversation about AI internally for good.

Next step: assign one no-code platform per role-track and have champions run the Tuesday onboarding before midweek.

If you want every employee to ship their first AI automation in five days — book a 30-min call and we'll map your team's first week: https://course.aiadvisoryboard.me/business