A few months into a machine-learning project, a team lead on a small data group once described the strange feeling that comes with working in this field: everyone was busy, the backlog kept growing, and nobody could say with confidence what had actually been finished. That memory is the reason this article exists. A done list for an AI team, kept as a plain column of shipped and checked work, answers more questions than most dashboards ever will. The questions that follow are the ones newcomers raise in their first week of trying it.
Why does an AI team need a done list at all?
Because the work in this field is unusually easy to lose track of. A model gets retrained, a feature pipeline gets patched, a prompt template is rewritten, an evaluation set is refreshed. Each change is small, and each one lives in a different tool. Weeks later someone asks, “Did we already fix the labelling drift?” and the honest reply is a shrug. A done list is the shared memory that stops that shrug. It records what has shipped, what was checked afterwards and when.
The contrast with a backlog matters here. A backlog describes what you hope to do. A done list describes what you did. When a stakeholder wants to know whether the team has been productive, only one of those two documents can answer truthfully.
What counts as “done” for an AI task?
Be strict. A task joins the list only when it has shipped and somebody has confirmed that it behaves as intended in the place where it matters. For a retrained model, that might mean the new version is serving traffic and a comparison against the previous version has been reviewed. For a data-cleaning job, it means the corrected data is in the downstream table and a spot check has been recorded. “Code merged” alone is not enough, because plenty of merged changes never reach anyone.
A simple rule of thumb helps: if you would hesitate to tell a colleague “that is finished, you can build on it,” the item does not belong in the done column yet.
How detailed should each entry be?
One or two lines. Record the date, what changed, where it now lives and how it was checked. Skip the narrative; the commit history and the experiment tracker hold the depth. The done list is an index, and it only works if scanning it takes a minute. Teams that let entries sprawl into paragraphs stop reading them, and then stop writing them.
Is this just a changelog under a different name?
Close, but not identical. A changelog is usually written for outsiders, such as users of a library or a product, and it tends to describe features. A team done list is written for insiders and includes the unglamorous items: the access request closed, the stale feature flag removed, the monitoring threshold adjusted, the documentation page corrected. Those chores are what stop an AI service from slowly rotting, and they vanish from a feature changelog.
Who should write the entries?
The person who finished the work, at the moment they finish it. Delegating the record to a coordinator at the end of the week produces vague summaries and forgotten items. If the habit feels heavy, attach it to something that already happens, like moving a card to the finished column or merging a reviewed change.
How does a done list help when something breaks?
This is where it earns its keep. When a model’s output quality drops on a Tuesday, the first useful question is “what changed recently?” A dated record of completed work lets you scan backwards in seconds instead of interviewing everyone. The team from the opening story found that three of its first four investigations ended at an entry that was less than a week old. Nobody had been careless; the changes simply had not been connected in anyone’s head until the list put them side by side.
What about decisions, not just tasks?
Include them when they close a question. “Decided to keep the existing embedding approach after comparing it with an alternative” is a finished piece of work. It also prevents the same debate from being reopened by a newcomer three months on. If you want the reasoning behind how organisations act on this kind of evidence, the piece on the way machine learning feeds into choices made by companies picks up that thread.
Does it help non-technical stakeholders?
Greatly. Managers and project sponsors rarely want to learn the vocabulary of training runs and inference latency. They want to know what is complete. A done column written in plain language gives them that without a meeting. It also dampens the pressure to add more work, since visible progress makes it easier to say that a new request should wait for the next cycle.
Where do the entries live?
Anywhere the whole team can open without friction: a shared document, a wiki page, a dedicated board column. Tool choice matters far less than consistency. If you already run a tracker, a filtered view of closed items may be all you need. Teams working on heavy graphics or visual models sometimes pair the list with a note on infrastructure, much like the engineering discussion in this write-up on building visual systems without bottlenecks, so that finished infrastructure work sits beside finished model work.
Can a done list be useful outside work?
Yes, and the idea is older than any software team. Travellers keep trip logs of bookings made and documents sorted. A family might keep a note of which forms have already gone out. Even a website such as MotoShare, a rental service in Bahasa Indonesia covering motorbikes and cars, presents its journey as steps a renter works through and completes. Closing steps in sequence and seeing them closed is satisfying for a reason: it turns vague effort into visible progress.
What are the usual traps?
- Counting unverified work. If nobody checked it, it is not done. Park it in a separate “shipped, awaiting check” spot.
- Using the list as a scoreboard. The moment people compete over entry counts, they begin splitting tasks artificially. Keep it a record, not a ranking.
- Skipping the dull items. Small maintenance entries are exactly the ones that explain later incidents.
- Never pruning the pending side. A done list sits beside a to-do list; trim stale to-dos regularly so the contrast stays meaningful.
How do we start this week?
Create one shared page titled with the current month. At the end of each working day, spend two minutes adding whatever was completed and checked. At the end of the month, read it aloud in a team meeting and note which items surprised you. Most groups discover they accomplished far more than the backlog led them to believe, and that discovery alone tends to improve morale and planning at once.