Prepare your AI agent pilot: a briefing in seven questions

Most AI agent pilots don't fail on the technology. They fail on the assignment. "The agent handles our customer questions" is not an assignment; it's a wish. A pilot that starts with a vague briefing usually ends up as a demo: nice to watch, but four weeks later nobody uses it.

This article contains a briefing in seven questions you can fill in today. They cover the task, the sources, the desired output, tools and permissions, exceptions, acceptance tests, and who owns the work after the pilot. It is written for businesses that have an AI agent built by a partner or set up in-house, and it works just as well if you build it yourself. Answer the questions with your team and bring the result to your first meeting. Use the briefing as a checklist: the sharper your answers, the more precise the quote, and the smaller the chance of an AI agent pilot without a verdict.

Large white platform with blue dots, dashes, and lines of light next to a tiny desk and chair
Seven questions that make the difference between a pilot with a verdict and a demo.

Why does a briefing work for your AI agent pilot?

A briefing works because it forces three things before your AI agent pilot starts. You pick one task instead of five. You agree on what the agent may and, above all, may not do. And your pilot gets an end date with a verdict: continue, adjust, or stop.

An AI agent is not a ready-made switch you flip. It is software that performs a task within goals and boundaries you set upfront. Those boundaries are not defined during the pilot, but before it. In a ten-step plan for agentic AI pilots, Salesforce describes starting with one problem and one repeatable task. After that, you define the agent's role and agree on guardrails and measurement points beforehand.

New to the topic? First read what an AI agent is and how it works or how an AI agent differs from a chatbot.

By some estimates, more than 80% of AI projects fail, according to a 2024 RAND report. For that report, RAND interviewed 65 experienced data scientists and engineers. It names the main cause directly: "Misunderstandings and miscommunications about the intent and purpose of the project cause more AI projects to fail than any other factor."

For agents specifically, Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027. The causes: rising costs, unclear business value, or weak risk controls. Gartner analyst Anushree Verma calls most current projects "early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied." A briefing does not remove all of those causes. It does put them on the table before you start.

QuestionWhat you defineWhat it looks like

Which task?

One recurring task with an owner

"Weekly report from project data"

Which sources?

Input and allowed data

"Project tool, read-only; no customer data"

Which output?

Concrete desired result

"One-page report in a fixed structure"

Which rights?

Tools and minimal access

"Read-only access to the sources"

When to stop?

Exceptions and boundaries

"Nothing gets sent without human review"

When good?

Acceptance tests and baseline

"Five correct reports in a row"

Who after?

Ownership and support

"Owner requests changes; partner provides support"

The seven questions explained

Below, each question explains why it matters and what input a good answer contains. The sample answers are deliberately concrete: a briefing you cannot fill in is a form, not an assignment. At Voltti, we use the same seven questions for our own Hermes installation. They line up with the Salesforce plan and with the causes in the RAND report.

  1. 01

    Which recurring task will you delegate?

    Pick one task that comes back regularly, runs more or less predictably and produces output you can check. "Something with email" is not a task; "remind about every open quote after three days" is. Write the task the way you would brief a new employee, exceptions included. Also note how often the task comes up per week. Volume tells you whether the pilot will produce enough cases to judge. How to spot suitable tasks is covered in our article on recurring tasks.

    Name the owner too: who in your company knows this task best and judges whether the agent does it well? A pilot without an owner is a pilot without a judge.

  2. 02

    Which sources and data may the agent use?

    Some data belongs in a pilot, some does not. State explicitly which systems the agent consults: project tool, CRM, email, document folder. Also name the data that stays out of scope, such as personal data you don't need, financial details, or anything under an NDA.

    Think about shape as well. Can the agent read the source as it is, or does data need cleaning first? Messy input is the most underestimated cause of disappointing pilot results.

  3. 03

    What does the desired output look like?

    Describe the result concretely enough to test it. Think of a report of at most one page in a fixed structure, a draft email in your house style, or a daily summary at 9 a.m. Add one or two examples of a good result, and optionally of a bad one.

    The sharper the output picture, the easier acceptance tests become later (question 6). Vague phrases about "better output" cannot be tested.

  4. 04

    Which tools and rights does the agent need?

    Map out which systems the agent works with and which rights come with them. Start small with minimal rights: separate accounts, read-only where possible, write access only where the task truly needs it. Write it down per system: may the agent read, propose a change, or write directly? A reporting agent needs read access, not admin rights to your entire CRM. Also agree on where the agent records every action, in a log or a ticket. That way the owner can check afterward what happened.

    Ask a partner how they handle rights and whether connections to your existing tools are possible; our article on connecting AI to your CRM and CMS goes deeper.

  5. 05

    What are the exceptions and stop conditions?

    Every real task has edge cases. What does the agent do when a source is unreachable, when data is missing or when a customer asks something unusual? Define what the agent does then: skip, flag for human review or stop.

    Set hard limits as well: nothing gets sent without human review, no access to data outside the scope, no spending above a fixed amount without approval. Stop conditions protect your business and make the pilot safe enough to test for real.

  6. 06

    When is the pilot a success?

    Define upfront when you will call the pilot a success. Good acceptance tests are small and countable: five consecutive correct reports, checked by the owner. Collect a small set of test cases before the pilot starts: ordinary ones and awkward ones, such as a duplicate record, a missing field, or a request outside the scope. Write down the expected result for each, and run the set again after every change. Record a baseline too: how long does the task take now, manually? Without a baseline you have nothing to compare against after the pilot.

    Agree on a cost ceiling as well, so the pilot can also fail on cost instead of slowly getting more expensive. Set an end date and what happens next: continue, adjust, or stop. One month with a verdict says more than a pilot that quietly keeps running.

  7. 07

    Who is responsible after the pilot?

    Who judges the results, who requests changes, who acts when the agent does something unexpected? And what support do you expect from your partner after delivery: updates, monitoring, availability when things break? How Voltti fills that in is described in our approach to a pilot.

    This question feels administrative, but it prevents a classic trap: a working agent nobody maintains anymore. Naming responsibility is the difference between an experiment and a service.

What does a filled-in briefing look like?

Here is a fictional example: the briefing for a weekly project report, filled in question by question. The numbers are invented, not measured at a real customer.

  1. 01

    Task and owner

    "Every Friday, summarize the project status into a one-page report. Owner: the project lead."

  2. 02

    Sources and data

    "Project tool (read-only) and the shared document folder. No customer data, no financial data."

  3. 03

    Desired output

    "Report with fixed sections: status, risks, decisions pending. Sample report available."

  4. 04

    Tools and rights

    "Separate account for the agent, read-only access to the sources, no access to email. The agent writes every run to a log file."

  5. 05

    Exceptions and boundaries

    "The agent flags missing data; it fills in nothing. Nothing gets sent without human review."

  6. 06

    Acceptance tests and baseline

    "Five consecutive correct reports. Baseline: it currently takes 45 minutes a week manually."

  7. 07

    Ownership and support

    "The owner reviews weekly. The partner delivers updates and support as agreed in the quote."

Copy the briefing as a checklist

Copy the template below into a document and fill it in with your team. Don't know an answer yet? Leave the line open: that is input for your first meeting.

AI agent pilot briefing
1. Task and owner (how often per week?):
2. Sources and data (allowed / out of scope):
3. Desired output (format and one example):
4. Tools and rights per system (read / propose / write):
5. Exceptions and stop conditions:
6. Acceptance tests, test cases, and baseline:
7. Ownership and support after the pilot:
Log (where does the agent record its actions?):
Cost ceiling:
End date and decision: continue / adjust / stop

An example from our own practice

At Voltti, a Hermes installation runs for our own work. For its SEO tasks, we go through the same seven questions. One recurring task: analyzing search data and proposing priorities. Sources: read-only access to our own Search Console data, nothing more. Output: a priority list with reasoning. Boundary: Search Console stays read-only. The agent makes changes in the CMS and publishes some of them itself; a person reviews that work afterward and adjusts where needed. Log: every change starts as a ticket and ends with a report in that same ticket. How that works is described in our article on SEO with an AI agent.

What that yields: a briefing of half a page that everyone involved can read. The agent knows its boundaries, the owner knows what to check, and there is a measurable verdict: do the proposed priorities match what we would choose ourselves?

How do you get from briefing to pilot?

With a filled-in briefing, you are ready to discuss your AI agent pilot with a partner. They translate your answers into a proposal: which steps the agent performs, which connections are needed, what the pilot covers and what support includes afterward. Which questions to ask a vendor is covered in our article on what an AI agent costs.

Not sure whether to build an agent yourself or have it built? Our article on building or outsourcing your own AI assistant compares both routes. On the AI agent development page you can read what Voltti takes care of and which first tasks are suitable. Voltti guides pilots like this for businesses in Belgium and the Netherlands. Ready to start? Discuss your pilot and bring your filled-in briefing: the first conversation is immediately substantive.

Portrait of Seppe Gadeyne

Seppe Gadeyne is a developer and the owner of Voltti. He sets up AI agents for companies in Belgium and the Netherlands.

Updated on

Frequently asked questions

01How long does filling in the briefing take?

Expect an hour, with the people who know the task. Most time goes not into filling it in, but into agreeing on the task and the boundaries. That conversation is exactly what the briefing is for. If you cannot answer a question yet, write down the open question: that is useful input for the first meeting too.

02Should I share the briefing with my vendor?

Yes. The briefing is the starting document of your pilot: your vendor or partner bases the approach, the rights and the quote on it. Without a briefing, your partner has to guess, and guessing costs time and therefore money. Also note what you deliberately keep out of the pilot.

03Can a pilot work without a briefing?

Technically anything is possible, but a pilot without a briefing measures nothing. Afterward you don't know whether the agent fell short or the assignment was unclear. The briefing is the difference between an experiment with a verdict and a demo. Without a measurement point agreed upfront, the evaluation comes down to gut feeling.

04What if I don't know which task to pick?

Pick the recurring task that happens most often and is least enjoyable to do manually. Still unsure? Read how to spot recurring tasks or request an intro call. In a short call, we look for a suitable first task together.

05How long does the pilot itself take?

At Voltti, a pilot usually takes 2 to 4 weeks, from the first conversation to the evaluation. A filled-in briefing shortens that first conversation, because the task, the boundaries, and the measurement point are already on paper. Read how those weeks run in our approach.

Request a conversation

Discuss your pilot