Health|May 8, 2026|Francis

Can an AI coach really help with type 2 diabetes?

Can an AI coach really help with type 2 diabetes?

Can an AI coach really help with type 2 diabetes?
Most "AI for diabetes" marketing right now is hand-wavy. Companies use the phrase to mean a dozen different things, half of them medical devices, half of them chatbots, and almost none of them tell you which kind they actually are. We talked to someone at the American Diabetes Association recently who said this is the most confusing part of the whole space for patients. So this article tries to be specific. It splits up what AI in diabetes actually means, what the evidence says, and where a tool like ours honestly fits, and doesn't.
Quick disclosure: BodyBuddy is our app. It's an AI accountability coach, not a medical device. I'll tell you exactly what it can and can't do. None of this is medical advice. Talk to your endocrinologist, certified diabetes educator, or pharmacist about anything that touches glucose, insulin, or medications.

Three different things the word "AI" is doing in the diabetes world

If you don't separate these, you'll be sold the wrong thing.
Insulin-titration AI (medical devices, FDA-cleared). These are tools that read your glucose data and recommend or automate insulin dosing. Hygieia's d-Nav is FDA 510(k) cleared and adjusts basal/bolus doses based on glucose data, with about 90% of users lowering their A1C in three months without raising hypo risk. Bigfoot Unity (now Abbott) is an FDA-cleared smart insulin pen-cap system that pairs with the FreeStyle Libre 2 CGM. Tidepool Loop is FDA-cleared automated insulin delivery, originally an open-source patient-built project, now mostly used by people with type 1.
These are real medical devices. They require a prescription and a clinician relationship. They are not chatbots, and they are not what BodyBuddy does.
AI-augmented coaching programs. Vida Health, Omada, Welldoc BlueStar, Twin Health. These wrap AI features (personalization, triage, nudges) around a clinical program with human coaches, dietitians, or in some cases physicians. Welldoc's BlueStar is FDA-cleared as a digital therapeutic. RCTs showed 1.7 to 2.0 percentage-point A1C reductions vs ~0.7 in usual-care controls. Twin Health's RCT (published in NEJM Catalyst, August 2025) found 71% of patients on their bundled AI-plus-sensor program reached A1C below 6.5% with fewer medications, vs 2.4% in standard care. The "AI" here is a layer on top of real care, not the entire product.
Habit and accountability AI (consumer wellness). This is where BodyBuddy lives, alongside more general AI coaches. The product is daily texting nudges around food, walking, sleep, medication reminders, and habit follow-through. No glucose data, no insulin recommendations, no clinical claims. The mechanism is behavioral, not pharmacological.
Generic chatbots (ChatGPT, Bard, Perplexity). These are not diabetes products at all. People use them like ones, which is a problem. A 2024 Mount Sinai study found AI chatbots produced problematic medical responses about half the time and would repeat false claims with confident elaboration. A separate Duke researcher reported that Bard told them to start insulin as the first-line medication for type 2 diabetes, with no mention of metformin, no warning about hypoglycemia. Don't use a general-purpose chatbot for medical decisions.
If a product won't tell you which of these four things it is, that's a flag.

What the evidence actually shows

The honest picture: AI in this space has real, modest evidence behind it. Not magic. Not nothing.
The Stanford GSB study from 2024. Sridhar Narayanan and colleagues analyzed roughly 65,000 HealthifyMe users over three years. AI alone produced about 2.7 lb of weight loss over three months. AI plus a human coach produced about 4.7 lb, 74% more. The behavioral mechanism is interesting: human coaches don't help by giving better advice. They help by getting users to log meals roughly twice as often. Frequency of contact was the lever, not the wisdom of the contact.
Welldoc BlueStar RCTs. 1.7 to 2.0 A1C-point reductions vs roughly 0.7 with standard care, in published RCTs from 2018 onward. This is the strongest A1C-specific evidence in the AI-coaching category.
Omada Health PREDICTS. Modest but durable. About 0.23 percentage-point A1C drop vs 0.15 in control, with 5.4% body weight loss vs 2% in control, in a randomized prediabetes population. Two-year peer-reviewed durability data published in JMIR.
Twin Health / Cleveland Clinic NEJM Catalyst (August 2025). The most striking AI-forward outcome to date: 71% of T2D patients on their bundled program reached A1C below 6.5% with fewer medications. But Twin Health is sensor-heavy: CGM, smart scale, smart watch. It's not a habit chatbot. It's a tightly integrated digital therapeutic.
Texting meta-analysis (J Telemed Telecare, 2019). Not strictly AI, but relevant: text-message coaching interventions in diabetes lowered A1C by an average of 0.62 points; 12-month interventions hit 1.6 points. The mechanism BodyBuddy uses is rooted in this same body of research.
Johns Hopkins (2025). A trial comparing AI-led DPP to human-led DPP found similar prevention outcomes. AI alone, when designed well, can match human DPP coaches on at least the prediabetes endpoint.
The pattern across all of this: AI works, AI plus humans works better, and any product claiming a 2-point A1C drop without published RCT evidence is selling you a story.
An AI coach can show up daily without fatigue, but it can't replace your care team.
An AI coach can show up daily without fatigue, but it can't replace your care team.

What real users are actually worried about

Worth listing because the concerns are mostly legitimate.
Privacy and HIPAA. Most consumer AI apps are not HIPAA-covered entities. What you tell a chatbot isn't protected the way what you tell your endo is. A PBS NewsHour story put it bluntly: read the privacy policy before you tell an app anything sensitive.
Hallucination. That ~50% problematic-response rate from the Mount Sinai study isn't a marketing exaggeration. It's a real failure mode. The Duke "start insulin first" example wasn't an edge case. AI will produce a confident wrong answer about your medication if you give it the chance.
Generic advice. Klinio defaults to 1,150 to 1,300 calories a day for many users, including active adults. That's a crash diet and not safe for most type 2 patients on medications. An AI without a real clinical layer will give you an average answer that's wrong for you specifically.
Replacing the relationship. A certified diabetes educator who knows your name and your story is more valuable than any AI. Some patients I've talked to worry that their employer will route them to an AI coach because it's cheaper. That's a real concern. The right framing for AI is augmentation, not replacement.
AI as the safety net for hypoglycemia. It can't be. Severe hypoglycemia is a medical emergency, associated with about 4x cardiovascular mortality risk. That's a CGM with alarms, a glucagon kit, and trained humans, not a chatbot.

What an AI coach genuinely can and can't do

Plain version, in two columns of honesty.
Can do well:
  • Daily reminders without judgment, fatigue, or time-zone problems
  • Walking nudges, sleep accountability, post-meal walk follow-ups
  • Medication adherence prompts at fixed daily anchors
  • Food logging support, especially the conversational kind that beats spreadsheet apps for adherence
  • Showing up every day for years, in a way human coaches can't because they sleep and have other clients
Cannot do, and shouldn't be marketed as doing:
  • Read your glucose meter or CGM. No real-time glycemic awareness.
  • Titrate insulin. That requires FDA clearance and live glucose data.
  • Diagnose anything.
  • Replace your endocrinologist, certified diabetes educator, registered dietitian, or pharmacist.
  • Be the safety net for severe hypoglycemia. Call your care team, not a chatbot.
  • Make medication changes. Even adjusting metformin timing should go through your prescriber.
If a product is sold to you as "AI that manages your diabetes," and it doesn't fall in the medical-device category I described in section 1, ask hard questions about what they actually mean.

Where BodyBuddy honestly fits

BodyBuddy is a daily AI coach that texts you over iMessage. It checks in on the boring 80% of type 2 management: food, walks, sleep, medication adherence. The AI is fully transparent about being AI, and we say that on the homepage because we don't think it's right to dress an LLM as a human.
For type 2 diabetes specifically, the honest framing is: BodyBuddy is the daily-habits layer that sits next to your meter, your CGM, your medications, and your care team. It's not a substitute for any of them. The texting coaching evidence (the 0.62-point A1C average reduction across 14 trials, 1.6 points for the 12-month interventions) is the underlying mechanism we lean on. We don't claim to lower your A1C. We don't have the RCT to back that claim, and we'd rather be honest than aspirational.
If you have a meter or CGM, an endocrinologist or PCP managing your medications, and a care plan you basically know, but you can't keep the daily lifestyle work going alone, that's the gap BodyBuddy is built for. If you don't have any of that infrastructure yet, start there first. An AI coach without a real clinical foundation is putting drywall up without a frame.

Frequently asked questions

Is AI better than a human coach for type 2 diabetes?

The 2024 Stanford GSB study found AI plus a human coach produced about 74% more weight loss than AI alone, mostly because human contact roughly doubled meal-logging frequency. So: human coaches help, when accessible. AI alone produces real but modest results. The 2025 Johns Hopkins data showed AI alone matching human-led DPP for diabetes prevention specifically. The right answer for you depends on what you have access to.

Is it safe to ask a chatbot for diabetes advice?

No, not for medical decisions. The 2024 Mount Sinai study found roughly 50% of chatbot responses on medical questions were problematic. Bard reportedly recommended insulin as a first-line type 2 medication in one researcher's test, with no mention of metformin or hypoglycemia warnings. For lifestyle questions ("what foods are higher in fiber"), generic chatbots are okay. For anything touching glucose, insulin, or medications, talk to a real clinician.

Are AI diabetes apps HIPAA protected?

Usually no, unless explicitly built as a covered entity. Most consumer wellness apps are not HIPAA-covered. Read the privacy policy. If the app is part of an employer benefits program through Teladoc, Vida, Omada, or similar, those typically are HIPAA-compliant. Standalone consumer apps frequently aren't.

Can an AI coach lower my A1C?

The published evidence on text-based coaching averages around 0.62 percentage-point A1C reductions across multiple studies, with longer interventions reaching 1.6 points. But "an AI coach" lowering YOUR A1C depends on the specific product, the specific user, and a lot of other factors. Honest products will tell you the mechanism (daily contact, habit reinforcement) rather than promising you a number. Talk to your endocrinologist about what A1C target makes sense for you.

What's the difference between an AI coach and a digital therapeutic?

A digital therapeutic (Welldoc BlueStar is the example) is FDA-cleared as a software-based medical product with published clinical evidence and a regulatory pathway. An AI coach (BodyBuddy and most others) is a wellness tool. They're different categories. If a product matters to your medication management, you probably want a digital therapeutic. If you're filling the daily-habits gap, an AI coach is the better fit and the right price point.

The takeaway

AI can help with type 2 diabetes. The evidence is real but modest, and very specific to which kind of AI you're talking about. The order I'd recommend if you're building your toolkit: a meter or CGM, a relationship with a primary care physician or endocrinologist plus a CDE, and then a daily-habits layer (an AI coach, a paid human-coaching app, or an employer benefits program if you have access).
BodyBuddy is the third thing: the daily habits layer. It's not the first or second. If you're missing those, go get them first. If you have them and the lifestyle work is the gap, an AI that texts you every day is one of the cheaper, more sustainable ways to fill it.
Just don't ask it to titrate your insulin. That's not what it's for.

Want daily accountability?

BodyBuddy texts you every day.

Build a healthier relationship with food and movement — one text at a time.

Join 500+ usersstaying healthy