The first 90 days of agentic work.
Most teams don't fail at AI because the models are weak. They fail because they pick the wrong first workflow, skip instrumentation, and design for the exception instead of the rule. This is the playbook we run with every new engagement — free, because a well-prepared client is a fast client.
The one rule that decides everything
Pick the workflow where the cost of a mistake is low and the volume is high. Not the most painful workflow. Not the most impressive one. The one where a bad Tuesday is an annoyance, not a news story.
This single rule explains most "AI transformation" failures. Teams pick the crown-jewel process, the one with the most stakeholders and the least tolerance for error, and then wonder why six months later nothing shipped. The agent that routes support tickets succeeds quietly; the agent that "runs your pricing strategy" dies in committee.
If you remember nothing else: low stakes, high volume, clear finish line.
Nine workflows that pay first
These nine show up in nearly every engagement we run, across industries. Each has a defined trigger, a bounded action, and a human-escalation path. Score yourself against them: which one is yours?
- Weekly performance digests. Pull data from the systems you already have, cross-check, publish. The first digest an agent ships buys trust for every one after.
- Inbound triage and routing. Classify every inbound message — email, form, ticket — and route it with a draft reply attached. Humans stop being switchboards.
- Outreach follow-up sequences. Draft-personalised follow-ups on a cadence, hand off the instant a human replies. Volume work, judgment at the edges.
- Onboarding checklists. Account setup, data provisioning, welcome sequences — the checklist-heavy 90% with humans on the exception path.
- Data entry and migration. The least glamorous, fastest-paying agent you'll ever ship. It never fatigues, and every correction teaches the next run.
- Reconciliation and matching. Invoices to POs, leads to accounts, imports to records. Agents flag mismatches; humans adjudicate.
- Report assembly. Not the analysis — the assembly. Pulling numbers, formatting decks, chasing owners for inputs. Hours vanish here every single week.
- Meeting-to-action pipelines. Transcribe, extract commitments, publish owners and dates, chase the stragglers before the next meeting.
- Content repurposing. One long asset into the six short ones your channels need. Brand-safe drafts, human-approved publishing.
Scope it like a hire, not a project
You wouldn't onboard a new employee with a 40-page spec. You'd give them a desk, three tasks, and a supervisor for the first month. Agents deserve the same: a narrow job description, a named escalation human, and a 30-day review.
Concretely: one trigger, one action, one output channel. "When a lead fills the form, enrich the record, draft the reply, and post it to the sales channel for one-click approve." That's a scope. "Automate sales" is not.
The 90-day shape that works: days 1-14 map the workflow as-is (no automating yet), days 15-45 ship the agent on the 90% case with humans on exceptions, days 46-90 tighten the exception rate and prove the hours saved. If day 45 arrives with nothing in production, the scope was wrong — shrink it.
Instrument before you automate
A week of logging what the workflow actually does beats a month of assumptions. Before an agent touches anything, you should know: how many items run per week, where they stall, who touches them, and what the exception rate really is. Teams routinely discover 30% of "the process" is dead weight nobody needed.
Then instrument the agent itself: every decision logged, every exception reviewable, every override reversible. Not because the agent will be perfect — because your team will only trust what it can see.
Keep humans in the loop — cheaply
The question is never "human or agent." It's where the human sits. The patterns that work, in order of how long teams stay comfortable with them:
- Human-in-the-loop approval: agent drafts, human one-click approves. Start here for anything customer-facing.
- Human-on-exception: agent runs the rule; a human gets pinged only for the 10% it's unsure about.
- Human-on-audit: agent runs everything; a human reviews a sample weekly and can reverse anything.
Most workflows can be run at 90%+ autonomy inside 90 days. The last 10% — judgment calls, edge cases, relationships — stays human for years, and that's fine. That 10% is where the interesting work went anyway.
When to kill the agent
Honest automation includes an exit ramp. Kill or re-scope the agent when: the exception rate climbs instead of falling; the human reviewers start rubber-stamping (they've stopped adding judgment — remove them or fix the agent); or the workflow itself changed and nobody updated the map. An agent supervising a dead process is worse than no agent at all.
The teams that win at agentic work aren't the ones with the best models. They're the ones with honest kill criteria.
Ready to find your first workflow?
Two minutes, eight questions, an honest score — plus which of the nine you should automate first.