Every machine learning project begins with a handful of beliefs that never make it into a slide: the data will look next month like it looks today, the labels mean what we think they mean, and somebody will notice when predictions drift. Writing down assumptions at the start is the cheapest way to turn those beliefs into things you can test. Teams who try this tend to raise the same handful of worries, so each one gets a short, jargon-free reply below.
What counts as an assumption in an AI project?
It is anything your plan relies on that you have not yet confirmed. If you could be wrong about it and the project would feel the damage, it belongs on the list. A sentence like “the camera images in production are lit the same way as the training photos” is an assumption. “We have a training set” is a fact, so it does not need a line.
A handy test: finish the phrase “this works as long as…” and see what comes out. Whatever follows is your list.
Why bother writing them down if the team already talks about them?
Conversations evaporate. A spoken assumption gets re-interpreted by each listener, and after a few weeks two people will remember two different versions. A written line has one wording. It can also be reread by the person who joins the project later and wonders why the model rejects half the inputs from one source.
There is a second reason. Writing forces precision. “The data is clean” feels fine when said aloud and looks hollow on paper. On paper you are pushed to say which columns, which period and clean by whose standard.
What should a good entry look like?
Keep each entry to three parts so it stays quick.
- The belief: one sentence, no hedging words inside it.
- How sure you are: a rough label such as solid, probable or a hunch.
- How you would find out: a small action, like sampling fifty rows by hand or asking the data owner a direct question.
That last part is what makes the list useful rather than decorative. A hunch with a ten-minute check is worth far more than a polished paragraph about risk.
Can you give an example from a real-looking project?
Suppose a team wants a classifier that sorts product photos into categories. Their short list might say:
- Photos uploaded by sellers resemble the curated photos we trained on. Probable. Check: pull a random batch from this week’s uploads and eyeball it.
- Category names mean the same thing to every labeller. Hunch. Check: have two people label the same small sample and compare.
- A wrong guess costs a seller a few minutes, not money. Solid for now. Check: revisit if the feature is wired into pricing.
- Someone will look at the confusion between two similar categories each week. Hunch. Check: name an owner and put it on a calendar.
If the topic of how a model reads photos is new to you, a friendly walkthrough such as this simple tour of how image classification works makes the example above easier to follow.
How often should the list be revisited?
Whenever something moves: new data source, new user group, a change in the upstream system, a different threshold. Many teams attach a quick review to the same meeting where they already look at results. If an assumption turns out wrong, do not delete it. Strike it through, add what you learned and the date. The record of what changed your mind is as valuable as the original line.
Does this slow down the experiment?
Expect to spend a quarter of an hour up front; the payback arrives the first time something looks off. When a model underperforms, the first debugging hour is often spent rediscovering what everybody had taken for granted. With a list, you walk down it and cross off suspects. Teams that make decisions with models also tend to find that stakeholders trust numbers more when the limits are laid out; the piece on how models feed into everyday business decisions touches on that human side.
Who should own the list?
One person keeps the document, but everybody adds to it. Treat it like a shared glossary rather than a report card. The owner’s only duties are to keep it readable and to nudge people when a low-confidence line has sat untested for too long.
Is there a place to learn how technical teams document this?
Plenty of places. Design notes, tickets and review templates in engineering teams often have an assumptions section. One example of an outside resource is NoOpsSchool, an India-based training website that describes itself as offering courses toward certification, along with consulting and other services, for people working in IT. Browsing a site like that can give you a feel for the vocabulary used when practitioners write about operations, though your own team’s template will matter more than any outside one.
What if my assumption is about people, not data?
Those are often the most expensive ones, so write them with the same care. “The support team will review flagged cases within a day” is a belief about a human process. So is “the business owner agrees that a false alarm is cheaper than a miss.” Check them the same way: ask the person, in writing if possible, and record the answer next to the line. Teams are sometimes surprised that the people-shaped entries change more often than the technical ones.
How is this different from a risk register?
Registers of risk collect bad outcomes that might occur. The assumptions list sits one step earlier and lists the conditions under which your plan makes sense at all. Many risks are simply assumptions turned upside down. If you keep both, link them: each risk can point to the belief that, if false, would trigger it. If you keep only one, the assumptions list is the lighter and friendlier starting point.
What if the team disagrees about a line?
Good. A disagreement on paper is cheap, and a disagreement discovered after launch is not. Record both views, give each a confidence label, and agree on the quick check that would settle it. Often the argument dissolves once the check is named, because both sides discover they were describing different parts of the same system.
What mistakes do beginners make?
- Listing wishes (“the model will be accurate”) instead of conditions.
- Writing so many entries that nobody reads them. Start with five.
- Skipping the confidence label, which makes everything look equally certain.
- Never revisiting the list after the first week.
Where do I begin?
Tonight, open a blank page, write “This works as long as…” and finish the sentence five times. Give each line a confidence word and one quick check. Share it with a teammate and ask which line surprises them most. That surprise is exactly the piece of information the exercise exists to find, and it costs you a coffee break rather than a post-mortem.