Agentic AI / Pilot acceptance sheet
Your first agentic AI workflow: a controlled SME pilot
Agree the task and permissions before connecting tools.
Download worksheetEditable CSV · Open in your spreadsheet app
Download article in MarkdownStart an agentic AI pilot with one task, approved information, and a narrow set of possible actions. For an SME, a useful first pilot might prepare an order-exception summary or a purchasing recommendation. The business should be able to inspect the result and continue working when the AI step fails.
Separate advice from permission to act
A system that drafts a response has a different responsibility from one that sends it. A system that recommends a refund has a different responsibility from one that changes a payment record. Write these boundaries into the scope before selecting a model or connecting tools.
For an illustrative order-exception pilot, the agent can read the relevant order, stock, and shipment records. It summarizes the problem and proposes a next step. An authorized operator reviews the evidence and decides what happens. Expanding the pilot to write records is a separate acceptance decision.
Cloudflare Workers AI can provide the inference component. Cloudflare's human-in-the-loop patterns describe approval mechanisms at different layers, including paused workflows. Choosing such a mechanism still requires application-level authentication, authorization, and a clear definition of what is being approved.
Design the approval around the actual action
Show the affected record, current values, proposed changes, and relevant evidence to the reviewer. Approval should apply to that particular action. If the order, amount, recipient, or stock state changes afterward, recheck the action and request renewed approval where required.
Keep permissions limited to the pilot. Instructions in a document, customer message, or retrieved page are data to interpret; they should not grant new tool access or override the workflow's rules. Enforce consequential restrictions in application logic as well as model instructions.
Build a pilot acceptance sheet
| Boundary | Question to settle | Acceptance test |
|---|---|---|
| Task | Which inputs and outputs belong in the pilot? | Ordinary, ambiguous, and out-of-scope requests |
| Information | Which records may this user and agent access? | Attempt to retrieve an unauthorized record |
| Action | Which tools can read or change data? | A proposed action outside the permitted scope |
| Approval | Who can approve which exact change? | Changed data or expired approval before execution |
| Repetition | How is a repeated request recognized? | Retry without a duplicate business action |
| Failure | What happens after a timeout or bad result? | Human handoff with useful context |
| Cost | Which limits stop an expensive loop? | Maximum steps, retries, and usage reached |
Evaluate the hard cases before using live authority
Prepare examples with expected behavior before development is finished. Include missing data, contradictory records, stale stock, and a message containing instructions that conflict with the workflow. Hold back some examples for a later check so the final evaluation is not only a replay of development cases.
Review language quality and action correctness separately. A fluent explanation does not establish that a refund amount is right. A refusal or escalation can be the correct result when evidence or permission is missing.
Begin with historical, synthetic, or appropriately controlled records. Move into a monitored pilot only after the business has accepted the relevant checks and access boundaries. Decide who can disable the workflow and how unfinished work returns to the human queue.
Measure the complete cost of a finished task
Record accepted, edited, rejected, and escalated outputs. Include human review time, retries, model use, supporting services, and maintenance. Measure cost per correctly completed task alongside the percentage of tasks completed without an intervention.
For example, an invented workload of 300 drafts requiring two minutes of review each consumes ten staff hours. That time belongs in the business case even if inference is inexpensive. Use current provider rates and measured usage when pricing a real pilot.
Connect the pilot to the internal-tool design that people will use to review it. Devekko's Enterprise Modernization service scopes agentic work around the operating task, permissions, evidence, and handover.
Your next step
Discuss your modernization
Share the task, your current systems, and the decision you need to make. We’ll discuss a useful first scope.
Discuss your modernization