A common worry about AI projects is that they never end. Months of meetings, a long specification, and still nothing your team can use. A sprint is built the other way round: one clearly defined tool, a fixed price, and a working result in two to three weeks.
This article walks through what happens, step by step, and what you need to bring. The exact rhythm varies with the task, but the shape stays the same.
Before the sprint: a call and a written proposal
The free 20-minute call. You describe your company and the work that takes up time. The aim is simple: find out honestly whether AI can help, and if so, with what. Sometimes the answer is that a simpler tool you already own would do the job. That is a fine outcome too.
The written proposal. If there is a good candidate, you receive a short written proposal covering:
- what the tool will do, and what it will not do
- which systems and data it needs
- what success looks like
- the fixed price and the timetable
Nothing starts until you have agreed to this in writing. If you are not sure which task to choose, an AI Audit can come first.
Days 1–3: scoping the details and getting access
The first days are about getting the details right.
- Walk through the task with the person who does it today. Not the manager's description of the process, but how it actually happens, including the awkward cases.
- Collect real examples. Real emails, real invoices, real questions, with personal data removed or handled as agreed. These become the test material.
- Set up access. The agent needs access to the specific inbox, folder or system it works with, and nothing more. Access is set up with the least rights that will do the job.
- Agree on the checks. Which actions need a person's approval? What should happen when the agent is unsure?
Week 1–2: building the tool
With the details in place, the tool is built:
- Connections to your systems, such as the inbox, accounting software, shared drive or calendar.
- Instructions for the agent, written from your rules, documents and examples.
- A simple way for your team to use it, ideally inside tools they already use, so there is nothing new to learn.
- Logging, so you can see what the agent did and why.
You will see work in progress during this phase. Early feedback is cheaper than late changes.
Week 2–3: testing on real cases
This is where many projects cut corners, and where a sprint should not.
The agent is run against the real examples collected at the start, and then on live work with a person checking every output. The questions are practical:
- Does it get the common cases right?
- Does it recognise the cases it should not handle, and pass them on?
- Are the drafts good enough that people actually use them?
Testing also records which cases the agent passed to a person, and why. That shows where the instructions still need work. The instructions are adjusted based on what goes wrong. The success criteria agreed in the proposal decide when it is ready, not a feeling.
Handover and training
At the end of the sprint:
- The tool is in production use by your team, not sitting in a demo environment.
- Your team knows how to use it, including what to do when the agent gets something wrong and how to report problems.
- You receive documentation: what the tool does, which systems it touches, where it runs, and what it costs to run.
- You decide on upkeep. You can look after it yourselves or continue with a monthly Care Plan covering hosting, monitoring and improvements.
What you need to provide
A sprint is short because both sides are ready. From your side, that means:
- A contact person who knows the work. Someone who can answer questions quickly and check the agent's output during testing. Expect to need a few hours of their time each week.
- Access to the relevant systems and data, granted promptly at the start.
- Real examples of the task, as described above.
- Someone who can make decisions, so questions about scope or rules do not wait for a week.
What is out of scope
To keep the price fixed and the timetable short, some things are clearly outside a sprint:
- More than one tool. A sprint delivers one tool. A second task is a second sprint.
- Replacing your core systems. The agent works alongside your existing software. It does not replace your ERP or accounting system.
- Changes after the scope is agreed, unless agreed together. New ideas are welcome, but they are priced and scheduled separately so the original delivery is not put at risk.
- Legal or compliance sign-off. You remain responsible for decisions about how personal data is used. See is your company data safe with AI? for the questions to ask.
Why a sprint works
A short, fixed project forces focus. Everyone knows what is being built, by when, and for how much. You get a working tool and real evidence of whether it helps, before you decide on anything bigger.
If the first tool proves useful, the next step is easy to judge the same way: one task, a written scope and a fixed price. If it helps less than you hoped, you find that out in a few weeks rather than months, and you have paid only the agreed amount.
If you want to see how the price is set, read what an AI agent costs.