Business workflow guide
How to choose your first AI agent workflow
Choose work with a clear finish line and a result somebody can check. A bounded comparison or preparation task makes a more useful first pilot than an instruction to manage an entire department.
Use these four selection tests
- Repeatable: the task happens often enough that improving it would matter. A one-off curiosity may teach you about the tool, but it is not yet a business case.
- Checkable: a reviewer can compare the output with an authoritative source. If nobody can recognise a plausible but wrong answer, choose a different first task.
- Reversible: begin with preparation rather than an action that commits money, changes records or reaches customers.
- Bounded: list the permitted inputs and the output format. 'Help with operations' gives neither the agent nor its reviewer a useful definition of done.
Example: a supplier comparison, not supplier selection
Suppose your team receives three proposals for the same piece of work. A useful pilot could organise the quoted scope, exclusions, dates and unanswered questions into a source-linked comparison. The person responsible for procurement still decides what to buy.
Use fictional or approved, sanitised proposals for the first run. Include one deliberate gap, such as an unspecified delivery charge. Check whether the output flags that gap or silently fills it. This is an illustrative test design, not a client case study or a claim that we have measured a saving.
Copy this pilot brief
Adapt this example to your approved task
Task: compare the three approved proposal documents. Output: one table showing scope, exclusions, dates and unanswered questions, with a source location for every factual entry. Boundary: do not contact suppliers, change files or recommend a purchase. Do not infer a missing price or delivery term. Missing evidence: write 'not stated' and add a question for the reviewer. Reviewer: [named person]. Finish: return the comparison and an uncertainty list for review. Stop: ask if a document is missing, unreadable or outside the approved source set.
Measure the whole job
Record the time taken to prepare the inputs, run the task, check the output and correct it. Compare that total with your usual process on a comparable example. Keep tool and implementation costs separate rather than treating every minute saved as cash saved.
Use a simple error log: unsupported statement, missing item, incorrect comparison, broken source link, or action outside scope. A fast draft that requires extensive rechecking may not improve the process. Decide your acceptable error and review thresholds before inspecting the results.
Work out what a pilot would need to earn back
Use the workload you actually have. If a task takes 40 minutes and an assisted version needs 15 minutes of total staff time, including preparation, running the task, checking and corrections, the potential saving is 25 minutes per completed task. At eight tasks a month, that is three hours and 20 minutes. These are example assumptions, not measured Blue Canvas results.
Decide what those hours would be used for. Staff time only becomes a cash saving when expenditure falls; otherwise it is capacity for other work. Compare the likely benefit with the pilot fee, subscriptions, delegated-task charges and maintenance. If the case depends on removing every review step, the estimate is too optimistic.
What makes this relevant to dots?
OpenAI describes dots as agents that can continue assigned work between conversations, bring back decisions and use connected sources within their permissions. That makes an explicit responsibility useful; it does not remove the need for a reviewer.
Product source: OpenAI's official documentation, checked 8 November 2026.
Decide whether to expand
- Continue the same small test if the result is promising but inconsistent. Do not widen source access to compensate for an unclear brief.
- Expand only the part that worked: perhaps a larger approved sample or a second output format, not simultaneous access to every business system.
- Stop if a reliable review would take more effort than doing the task, or if the necessary data cannot be used appropriately.
- Write down the decision, its evidence and the next review date. A successful demonstration is a reason to investigate further, not proof of a production-ready workflow.
What to include in a paid workflow pilot
A paid pilot needs an agreed deliverable. Ask for a written scope covering one task, permitted systems, a test sample and the person who accepts the result. Keep the setup cost separate from subscriptions and ongoing support.
For a proposal comparison, the deliverable could be a configured workflow, a reusable brief, a source-linked comparison and a short record of errors, review time and limitations. Agree these outputs before commissioning the work. An attractive demonstration alone does not show whether the process will hold up on your documents.
Blue Canvas can help with workflow planning, implementation and staff training. Start with a free 15-minute call to check the fit. If a pilot makes sense, we agree the work and fee in writing before starting. We cannot grant product access or promise a particular saving.
- Bring the task frequency, current time spent and systems involved, plus a fictional or sanitised example.
- Agree who owns accounts, checks the output and approves changes or external actions.
- Ask what is handed over, what support costs, and how you stop using the workflow if the test does not justify continuing.