The pattern of good first scenes
First scenes are chosen to be safe, not to impress: data already exists, rules are clear, outputs are checkable, mistakes can be rolled back. Scenes that meet all four show value within weeks; the rest are demos. The three types below are the most common safe starts.
Knowledge retrieval and drafting
Internal documents and procedures turned into on-demand answering and first-draft generation. Data exists, errors are reviewable before anything ships, value lands within weeks.
Why it is safe: the agent drafts, a human decides — everything passes through a human first, so mistakes stay internal and cost one revision. Boundaries: govern the knowledge base first (see what to prepare before deploying); replies committing the company externally need approval; and if the agent reads the public website, check its machine-readable foundations — in Negency's flagship measurement of the global homepages of 30 Chinese exporter companies (a single-day sample on 30 October 2026, measuring technical website visibility only, not brand mention rates in AI answers), machine readability averaged only 69.2 (median 68.5) and 17 of 30 sites had no JSON-LD at all. Source: Flagship Report 2026, Issue 1.
Structured repetition
Lead triage and routing, report and weekly-summary generation, meeting notes into action items. Clear rules, high repetition, verifiable outputs.
Why it is safe: people already handle this work by fixed routines — the agent takes over the routine, not the exceptions, and errors are visible at a glance. Boundaries: write the rules down first — an agent cannot invent them; when output leaves the company, sending needs approval while the agent only drafts; and exceptions need a human lane — outside the rules, the agent stops and asks, not guesses.
Alerting and pre-processing
Monitoring anomalies (stock, prices, reviews, system events) and preparing material for humans to decide on. The agent prepares and flags; humans decide.
Why it is safe: the agent makes no decision — it lays the situation in front of a person, change flagged, source attached. Boundaries: the business sets thresholds, not the agent; the agent may flag "worth a look" but must not trigger external actions alone (automatic reordering, public replies); and alerting needs regular review — false alarms exhaust attention, which is the same as no alerting.
Scenes to keep off the first list
Autonomous external commitments (quotes, contracts, payments without approval), irreversible operations (bulk deletion, production config changes), and processes whose rules nobody can state.
- Why not external commitments: a commitment, once sent, has effect. A quote with one wrong digit costs far more than the labour saved. These scenes need approval points first (see governance and approval points); automate only after approval runs smoothly.
- Why not irreversible operations: bulk deletion and production config changes cannot be undone with one click, and agent errors come in chains — one misjudgement can trigger ten steps. The iron rule for first scenes is "mistakes must be roll-backable".
- Why not rule-less processes: if the business cannot say what counts as a pass, the agent has no standard to execute against and can only guess. A right guess is luck; a wrong one is an incident. Write the process down first, then automate.
A working screen: does this scene give an agent a clear rule, reviewable output, and a way to stop? Three yeses and it qualifies; any no, fix the process first.
The rollout rhythm
Pick 2–3 scenes, run them small for 2–4 weeks, and measure error rate, time saved and override rate. Expand only from that data. Scenes run; governance runs with them from day one — they are not separate projects.
Watch the override rate closest: high overrides mean the agent's judgement diverges from the business standard — tune the rules or narrow the scope. Zero overrides for weeks may mean approvers stopped looking.
The four tests for a safe first use case
Pick the first use case wrong and everything after it meets resistance. Score each candidate against four tests and start with the ones that pass all four:
- Frequent and repetitive: it happens dozens of times a week and people already handle it by a fixed routine — the agent takes over the routine, not the exceptions;
- Checkable output: a person can tell at a glance whether the output is right (summaries, classification, drafts, comparisons), not only after expert review;
- Containable downside: the worst case is rework, not an external commitment, a payment or a compliance incident; anything touching those starts in human-approved mode;
- A real user waiting: someone on the business side actually wants it, instead of IT building it and then looking for adopters.
Conversely, use cases that speak directly to customers, trigger irreversible actions, or need cross-department accountability should not be in the first batch, however impressive the demo looks.