Choose Your First AI Workflow From A Real Bottleneck

Describe The Job Before The Technology

Write the job as input, action and output. For example: take an invented enquiry, identify which service it asks about, and suggest an owner for review. That is testable. 'Use an agent to improve sales' has no clear finish line.

Use the response audit to find the failure. If enquiries have no owner, begin with assignment. If the right person is assigned but lacks a necessary detail, try a checklist. If the task involves varied language and a reviewer can judge the answer, drafting or classification may be worth testing.

NIST's framework describes four risk-management functions: Govern, Map, Measure and Manage. Our worksheet uses that broad approach to define a task and check its result. The selection gates are our own; NIST has not validated them.

Context: NIST: AI Risk Management Framework 1.0.

Compare The Smallest Workable Options

For each job, write down what you can already do without a new tool. A shared task list, a saved reply or a fixed condition may be enough. The useful question is whether an AI-assisted version handles cases your simpler option cannot handle acceptably.

The fictional table uses 20 invented cases per task. 'Needs judgement' means an authorised person must resolve something the sample rule cannot decide. It is not an error rate and does not mean the remaining cases are safe to run unattended. Counts describe only this made-up worksheet.

judgement_cases = {
    "routing": 2, "missing_detail": 4, "reply_draft": 7,
    "bespoke_price": 12, "complaint": 15,
}
for task, count in judgement_cases.items():
    print(task, f"{100 * count / 20:.0f}%")
# routing 10%
# missing_detail 20%
# reply_draft 35%
# bespoke_price 60%
# complaint 75%

Apply Gates Before Scoring Anything

A high convenience score cannot resolve a privacy or permission problem. Keep the test offline until a named person can check its replies and you have permission to use the data. Start with invented messages and no sending connection.

The reply-draft row is a reasonable offline candidate because a person can inspect the output before it leaves the system. That is a hypothesis about testability, not proof that AI is better than the manual baseline. The price and complaint rows stay with a person because this exercise has not defined the authority or acceptance rules needed to automate them.

1. Permission: can the test use this data and tool without new accounts, unapproved spend or exposing personal information? If not, stop that path.
2. Acceptance: can two reviewers describe a correct output and a failure using the same examples? If not, clarify the task.
3. Control: can a person approve every external action and stop the workflow? If not, keep it offline.
4. Reversibility: can you restore the prior process and inspect what happened? If not, reduce the scope.

Context: NIST: AI Risk Management Framework 1.0; NIST: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile.

Write A Test Brief That Can Fail

A useful brief names one job, one reviewer, a fixed sample and a stop rule. For our fictional draft assistant: use the six enquiries in the next guide, require the necessary question or escalation in every answer, and reject any invented price or booking. Do not connect email, messaging, a CRM or payment software for this first pass.

Keep a human-written baseline and, where useful, a fixed-template baseline. Run the same cases through each. Count corrections and failed requirements. If you later time the work, include preparation, checking and rework; do not turn unmeasured convenience into a time-saving claim.

Before a later authorised vendor test, set a cost limit and check the plan's charges. Confirm how it stores data, where it is available and how to export your work. We have not checked paid plans or country availability for this guide.

Context: NIST: Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile.

Keep, Change Or Stop The Trial

Keep the candidate only if it meets the predefined requirements in the test and there is a realistic next validation step. Change it when a specific failure can be repaired. Stop when the task needs authority or reliable information the workflow does not have.

An offline pass is permission to investigate further, not permission to send customer replies automatically. A later operational trial needs an approved data path, tested recovery, an owner and a fresh review against more representative cases.

Worldwide English readers can use the same worksheet. US, UK, Australian and Canadian teams still need to check local data handling and a vendor's actual availability before using real enquiries. We have not established regional demand or willingness to pay from this exercise.