Risk tiers
Human-in-the-loop path for “Human control in AI workflows”
Handoffs in “Human control in AI workflows”
- Low risk — internal drafts, tagging, research digests → AI can run continuously; humans sample-review.
- Medium risk — CRM updates, customer-visible drafts, scheduling → AI proposes; human approves high-impact items.
- High risk — money movement, legal commitments, public posts, irreversible deletes → human required; AI only prepares.
Gate design patterns
- Approve to send — drafts queue until a person releases them.
- Approve exceptions only — AI handles happy path; edge cases escalate.
- Dual control — two humans for high-risk actions (finance).
- Time-boxed autonomy — AI acts within limits (hours, spend thresholds).
- Rate limits — hard caps on volume and tool use per day.
Designing the approve queue
Approvals need SLAs, context, and one-click correct/reject. If correcting takes longer than doing the work manually, people bypass the system. That is a design failure, not a culture failure.
Logging as a product feature
Every material action should leave a trail: what was sensed, what was planned, what tool ran, who approved. Without logs you cannot train operators or defend decisions later.
Progressive trust
Week 1: draft only. Week 3: auto-run low-risk steps. Week 6: expand only where evaluation samples stay green. Autonomy is earned with evidence, not declared in a kickoff deck.
Anti-patterns
- No logging of what the agent did
- Approvals on everything (rubber stamp, zero speed)
- Approvals on nothing “because the model is smart”
- Gates without SLAs (queue dies unread)
- Removing gates to “go faster” without measured error rates
When to remove a gate
Only when error rate is measured, cost of remaining errors is acceptable, and a kill-switch exists. Removing gates without data is how demos become incidents.
SMB default
Start stricter than you think. Loosen gates only after measured reliability. Pair with evaluation samples weekly during Pilot and monthly during Run. See Vision for production philosophy and pilot scoping for how gates enter the charter.
Want this applied to your stack?
Fixed-scope pilots for AI agents and automations. Map first. Ship one real workflow. Then run it.
Field notes from delivery
Across real scoping conversations, the same failure modes show up: no baseline, no owner, no written metric, and a desire to automate everything before anything is proven. Guides like this exist to slow that impulse without killing momentum. Clarity is speed.
When you brief Kokasync Labs, bring the messy truth — screenshots, exception examples, “this is how it actually works on Tuesdays.” Polished process docs that nobody follows will produce systems nobody uses.
A 30-day action plan
- Days 1–3: Inventory the three workflows that burn the most hours.
- Days 4–7: Score them with the checklist dimensions (frequency, time, risk, clarity, access).
- Days 8–10: Choose one pattern per workflow (chatbot / automation / agent / hybrid).
- Days 11–14: Write a one-page pilot charter with a single primary metric.
- Days 15–30: Run Map with a builder (internal or Kokasync) and decide go / no-go on Pilot.
How this connects to the rest of the site
The Library goes deeper on individual concepts. Vision explains the firm we are building. How we work is the engagement contract. AI agents and Automations are the service doors. These Guides are the shared language between them.
Glossary (quick)
- Baseline — measured performance before the pilot.
- Human checkpoint — a required person on a material action.
- Kill-switch — how you stop the system quickly when quality or cost drifts.
- Run layer — ongoing operation after pilot proof.
- Fixed scope — one deliverable, one number, one timeline.
Closing
If this guide changed how you would brief the next AI conversation, it worked. If you want help turning it into a live pilot, write to [email protected] with the workflow, the tools, and what better looks like in 30–60 days.
Detailed walkthrough: from pain to pilot
Start with a pain that has a clock attached — something that burns hours every week. Write the current steps as if training a new hire. Circle every handoff and every tool. That map is more valuable than a feature wishlist.
Next, invent the metric before you invent the architecture. “Make us more efficient” is not a metric. “Cut average handle time from 18 minutes to under 8 on tier-1 requests, with human review on external sends” is a metric. If you cannot write it, you are not ready to buy models, tools, or agency time.
Then decide the pattern per step — not for the whole company. Rules-based steps become automations. Language-heavy multi-step work becomes agent work with gates. Questions become chat with escalation. Hybrid systems are adult systems.
Finally, pilot small and instrument everything. Keep a weekly sample of successes and failures. Name an owner. Write a kill-switch. Only then expand. This is how Kokasync Labs compounds craft without becoming a factory of demos.
Quality bar we hold ourselves to
- Written metric before build
- Human control on material risk
- Logs that reconstruct a bad run
- Owner after launch
- Honest no-go when economics do not work