What an AI nutrition coach does, and how to tell a good one

The four jobs

1. Set a starting target. Every formula for calorie needs is an estimate with a wide error band; the starting number only has to be close enough to begin. A coach should set it from your weight, height, age, activity and goal, keep protein high enough to protect muscle (see how much protein per day), and treat the number as a hypothesis to test, not a verdict. The method is in how many calories to eat to lose weight.

2. Make logging survivable. The research on self-monitoring is consistent: people who keep logging lose more weight, and the hard part is keeping going. Every tap between you and a logged meal costs adherence. The question to ask is "what does logging dinner take on a Tuesday night when I am tired?"

3. Report against the target, daily. Not a dashboard you have to go and look at; a sentence at the end of the day: where you landed, what is left, what to do about tomorrow.

4. Adjust from results, carefully. The scale is the ground truth. If your logged deficit says you should be losing and the trend says you are not, the target is wrong (or the log is), and the coach should say so and propose a specific change. Carefully means: small steps, based on a trend rather than one weigh-in, with floors so it never pushes you into an aggressive deficit.

Eight questions to ask before trusting one

Question	Why it matters	What a good answer looks like
Where do the numbers come from?	A model guessing calories from a photo can be 20 percent off	Verified label data and official databases where they exist; estimates labelled as estimates
Can I log in a sentence?	Friction kills logs	"Chicken burrito bowl, no rice" is enough
What about restaurants?	Most people eat out several times a week	Published chain menus built in; honest estimates for the rest (how to eat out and still lose weight)
Does it learn my habits?	Repeat meals are most meals	Your usual portion of a repeat meal takes one line
How does it adjust my target?	This is the coaching	From a multi-week weight trend, in small bounded steps, with floors
How often does it talk to me?	More is not better	A rhythm you set; one nudge a day at most
Will it tell me when it is unsure?	Confident guesses cost you weeks	Estimates are shown for you to confirm or correct before they count
Does it know what it is not?	Software is not a clinician	It declines to advise on medical conditions and says to see a professional

Where software stops

An AI nutrition coach is not a dietitian. It cannot diagnose, manage a medical condition (diabetes, kidney disease, an eating disorder, pregnancy), or take responsibility for an outcome. If any of those apply to you, a registered dietitian or your doctor sets the plan and software can at most help you follow it. It also cannot see your plate: if you log "a bowl of pasta", it is estimating, and it should say so.

How Bay handles each part

Bay is a coach you text over iMessage, RCS or SMS. On the four jobs:

• Targets are set from your goal in the first conversation and confirmed by you. Protein is held fixed when calories change later.
• Logging is a sentence, a photo or a voice note. Numbers come from Open Food Facts and USDA FoodData Central rather than a model's guess where a match exists; a barcode in a photo resolves to the exact label; published menus for 160+ US chains are built in; and Bay learns your usual portions from what you confirm, so a repeat meal takes one line. Every log is proposed first and confirmed with a tapback, so a wrong estimate is fixed before it counts, and a correction ("that was two scoops") replaces what Bay had learned. When a barcode decodes but is in no database, it asks for the product name or a photo of the label instead of inventing numbers. See how to track calories by texting.
• Reporting is the evening recap: calories and protein against target, in a few lines. One nudge a day at most, timed to when you usually eat.
• Adjustment is the Sunday review. Bay fits a trend to your weigh-ins, estimates your real expenditure from 14 days of logged intake against that trend, and proposes a change only when the observed rate disagrees with your goal: at most 150 calories, never below 1,500 (men) or 1,200 (women), protein unchanged. It needs four weigh-ins over a week and ten logged days first; before that it tells you what data would help. A tapback applies the change; a thumbs-down holds for two weeks. "Am I on track?" gives you the same analysis any day.

Bay also keeps the facts you mention (an allergy, a food you will not eat, that you cook on Sundays) and uses them, and it covers training and daily accountability in the same thread, so the nutrition coaching does not live in a separate app from everything else. It is free right now. On the limits above, Bay is explicit: it is not a doctor or dietitian, and it says so.

What to expect in the first month

Week one is calibration: you log, Bay learns your meals and portions, and the targets stay put. By week two the weigh-ins give it a trend. By the end of week two or three it has enough to run a real review, and from then on the target moves in small steps toward your goal rate. If you log honestly, the scale and the log agree within a few weeks; if they do not, the review tells you which one to look at.