Free resources | Lang Horizon

Use this resource to define requirements. It is not a tested case study. Adapt the criteria to your task and risk; no result is guaranteed.

Download Chinese text Download English text

01 · Task and operator

Write down one recurring task, its operator, reviewer and stop decision owner. Avoid automating the entire business at once.
Record the current steps, frequency, time spent, rework and common failures. Mark unknown values as unmeasured; do not present estimates as a baseline.

02 · Inputs and data boundaries

List required inputs, sources, formats and permissions. Start with synthetic or anonymized examples. Keep customer data, credentials and trade secrets out of public demos.
Confirm external model calls, processing location, storage, retention and deletion. Pause integration if these boundaries are unresolved.

03 · Outputs and human judgment

Define one inspectable output: fields, format, evidence references, missing values and uncertainty. Specify which actions only produce drafts.
Assign human review points. An authorized person confirms sending, publishing, payments or changes to business records. The confirmation step should be cancellable.

04 · Acceptance and examples

Prepare fixed examples: normal input, missing fields, conflicting or stale sources, insufficient permissions and tool failures. Write expected outcomes before running.
Check factual accuracy, format completeness, traceable evidence and human takeover separately. Agree on pass criteria and cost limits in advance; generation success alone is insufficient.

05 · Baseline and comparison

Compare the current method and AI workflow on the same tasks, inputs and criteria. Record sample count, successes, time and cost, including failures.
Report results as successes / total, with model, tool and version details. Small samples are only signals; one smoke run cannot establish efficiency or training gains.

06 · Failure, retry and takeover

Set timeout, retry, cost and stop limits. After an external write fails, inspect its status before retrying to avoid duplicate messages or records.
On failure, show the reason, completed steps, sources and next owner. Keep a path to finish the task manually.

07 · Scope and pilot plan

Confirm deliverables, exclusions, timeline, fees, data use and acceptance owner in writing before starting a prototype or pilot.
Start with one workflow and controlled examples. Check saving, retrieval, refresh, recovery and source visibility. Passing tests does not establish live-store or production validation.

08 · Review and next step

Assign a review date and owner. Use the same criteria to continue, revise or stop. Expanding scope requires new evidence and agreement.
Keep a short record of changes, examples and outcomes, failure causes, cost, unverified points and next steps. Retain unfavorable results.

A task card to complete

Task:
Operator / reviewer:
Current process and unmeasured baseline:
Inputs / sources / processing location:
Output format and evidence:
Human confirmation points:
Fixed examples and expected outcomes:
Pass criteria / sample count:
Time / cost limits:
Failure, retries and takeover:
Deliverables / timeline / fees:
Review date / owner:
Continue, revise or stop criteria:

Example: organize meeting notes

Synthetic example: take meeting notes without personal information and return decisions, actions, owners and unresolved items, with a source quote for each entry. Mark missing owners as unresolved; do not guess or email automatically. Verify evidence for each item, retained conflicts and marked omissions, and compare against a human-produced result. This example has no actual run results.

Discuss a related workflow

AI workflow requirements and acceptance checklist