A vendor promises savings for an AI pilot, but no one knows what's allowed and what's not. Today, you'll build a multi-agent workflow where every figure is traceable. Imagine this opening scene: three agents negotiating in a coffee room with a calculator. That's fiction – the numbers, however, are not.
A simple chat response tells you what ROI *could* look like. An agent, however, receives an objective, tools, and a termination condition: it retrieves figures from your documents, calculates, marks assumptions, and stops if the evidence is insufficient. Let's do this for one case. Fictional example: a medium-sized company is considering an AI pilot for handling customer service emails. All figures are illustrative, not results from any actual client engagement.
Task brief: objective, input, tools, limits, termination condition
Start with role division. The roles are researcher, implementer, and auditor – and the auditor is not part of the same conversation as the creators. Here's a copyable task brief. Objective: evaluate if an AI pilot is worthwhile in this process and define its scope. Input: process description, hourly costs excluding personal data, current processing volume, vendor's offer. Tools: company's own documents, spreadsheets, and offer PDFs. Limitations: no passwords, no personal data, no invented figures. Termination condition: three scenarios, a source for each figure, assumptions marked, a go/no-go recommendation, and a written list of what not to do in the pilot.
Researcher: retrieves, marks sources, doesn't guess
The researcher agent works in parallel. It reads the process description, extracts message volumes and processing times from the system, and reviews the vendor's offer – even scanned attachments and images are now readable content for an agent, as are dozens of files at once. Multi-stage planning is evident in the sequence: first, the agent lists what figures are needed, then retrieves them, and marks the source for each, indicating whether the figure is measured or estimated. If something is missing, it doesn't guess – it leaves a visible gap and states who can provide the information.
Implementer: scenarios, assumptions, sensitivity analysis
The implementer agent builds a model from these figures. Illustrative example: current processing of 8,000 messages per month, average hourly cost of 45 euros including employer costs, agent's share of processing in three scenarios at 30, 55, and 70 percent. Human working time is only partially reduced because review and exceptions remain with a human. Tools include a spreadsheet where each cell refers to a source, and a separate list of assumptions. The model produces a sensitivity analysis: which variable sways the result the most. In this case, it's usually the time spent on human review, not the model's cost.
One decision memo: limits, requirements, go/no-go
The consolidation produces a single decision memo, not three separate reports: one page detailing profitability thresholds, implementation requirements, and red lines. The red lines are the most crucial part of this work. Illustrative list: the pilot will not process messages containing personal data without a separate assessment, the agent will not be given permission to send replies to customers without human approval, and customer data will not be transferred to an unapproved tool. If these conditions are not met, the recommendation is: do not start yet. That is an acceptable outcome.
Auditor agent evaluates – and stops if necessary
The auditor agent evaluates the result against separate criteria: does every figure have a source, are estimates marked as such, does the material contain personal data or identifiers, and are the acceptance criteria met? A typical finding: the average hourly cost was calculated without employer costs, which artificially inflated savings. Another: one scenario completely lacked a source. The auditor does not silently correct; instead, they return the work to the creator and halt publication. This also happens when evidence is insufficient: uncertain work does not proceed embellished but returns as a question.
Three takeaways. One: divide the work into roles – researcher, implementer, auditor – and keep the auditor separate from the creators. Two: demand a source for every figure and mark estimates as such; a sensitivity analysis tells you where to focus. Three: document the pilot's forbidden areas before anyone touches production data.
Exercise
Choose one recurring process (can be fictional). Write a task brief using the video's structure: objective, input, tools, limitations, termination condition. Run it with two roles: a researcher agent lists the necessary figures and their sources, and an implementer agent creates three scenarios in a table and marks assumptions. Use only invented, clearly marked example figures – no real personal data, passwords, or customer material. Finally, check the result against the acceptance checklist and note at least one point that did not pass.
Checklist
- Every figure has a marked source or is labeled 'estimate' with a justification for the estimate
- Three scenarios (conservative, baseline, optimistic) and a sensitivity analysis of which variable impacts the result the most
- A written list of what not to do in the pilot, and a clear stop condition: if evidence, access rights, or acceptance criteria are not met, the work does not proceed
- No passwords or unnecessary personal data have been given to the agent at any point
— Aino Pilvi, AI Specialist, Cloud Computing Oy
Questions? aino@cloudcomputing.fi | WhatsApp: +358 33 119 3196
