We keep trying AI. Nothing sticks.

The plan

Write a one-page trial plan for one recurring task. This is a manual, AI-assisted workflow: you copy approved inputs into Claude, check the output, and move the accepted result into your existing process. Start with the synthetic example below.

You need access to Claude and your team’s approved document or spreadsheet workspace. Use only data your business permits in those services. Tool plans and limits vary; this walkthrough does not require an API connection.

1. Prepare a small, useful input

Choose one job you performed at least five times recently. Write its input, output, owner and current manual steps in a document. Time five examples, including checking and correction. Use synthetic inputs for the first trial and choose a reviewer who knows the job.

Use this fictional sample for your first run:

TASK: Draft replies to basic return-policy questions.
INPUT: Anonymized request plus approved policy.
OUTPUT: Unsent draft.
OWNER: Support lead.
BASELINE: Five replies, 35 total minutes including checking. This is a fictional example, not a claimed result.

2. Give Claude a precise job

Open a new chat at claude.ai. Paste the prompt below, then paste the sample after it under a heading called INPUT. For a PDF, attach the sample file to the chat instead. Keep the source and instructions clearly separated.

Create a one-page trial plan from these facts. Include exact input and output, five sample cases, a review checklist, what the AI may not do, manual fallback, and a results table. Label any proposed success threshold as a proposal for the owner. Do not invent savings, integrations or a deployment result. Limit scope to draft replies.

INPUT:
[Paste the sample or your approved source here.]

Read the response before continuing. If required columns are missing, ask: “Return the same result using every requested column. Mark unsupported fields UNKNOWN.” Do not fill gaps by asking the model to guess.

3. Check against the expected result

A proposed trial uses five normal cases plus missing-policy and conflicting-policy examples. Every draft stays unsent until checked. The result table records drafting, review, correction time and factual errors separately.

4. Put the checked result to work

Have the owner agree to the threshold before trying the workflow. For example: zero unsupported policy claims and total work below the measured baseline across the same sample. Use the support-email challenge for the actual drafting prompt. Record observed results, including failures, in a sheet.

Keep a status of draft, needs clarification, approved, or completed beside each item. A draft is not a completed action. If the service is unavailable, use the original source and your normal manual process.

5. Test the awkward cases

If a draft invents a policy, stop and fix the source or prompt. If review takes longer than the baseline, keep the manual process. If the trial passes, expand to a small supervised batch and set a weekly review date; use the existing handoff and maintenance guides below.

Run the same cases again after changing the prompt or source. Keep the input, output, correction, and reviewer decision together. Expand only when someone can reliably check the results.

If the result is off

Narrow the input to one record and recheck the source. For missing facts, ask the process owner. For an unsupported statement, remove it and add an explicit constraint to the prompt. If the same failure repeats, do the task manually until the source or instructions are fixed.

Scope: This is an original illustrative recipe with synthetic data, adapted from the field guides below. It is not a live-tested integration or a claim of time saved. Claude’s file-upload documentation explains supported attachments; the copy-and-paste path does not need file uploads.