AI and automated tools can help people manage diabetes and other chronic conditions. The strongest evidence is for hybrid closed-loop systems, which adjust insulin automatically using glucose sensor readings, in type 1 diabetes. Coaching and tracking apps have more mixed evidence. All of them work best when a clinical team can step in, so do not change treatment based on an app without advice.
People living with type 1 diabetes make treatment decisions all day, every day, with no weekends off. How much insulin for this meal. Whether that walk will drop blood sugar. What to do at 3am when the alarm goes off. It is relentless mental work, performed by patients and parents, most of it without a clinician in the room. A typical specialist appointment might happen a few times a year. The disease happens every hour in between, which is why so much of its management has always been self-management, and why tools that share that load matter so much.
That is why diabetes has become the proving ground for AI in chronic disease self-management. Continuous glucose monitors generate a reading every few minutes. Insulin pumps can act on those readings. Algorithms can connect the two. The result, in the best systems, is a form of automation that has improved outcomes in randomized trials.
Here is the thesis. Automation in chronic disease works, sometimes remarkably well, but the systems that deliver the best results are the ones with a human in the escalation path: a clinician who reviews data, a patient who knows when to override the machine, and a clear route to help when technology fails. The goal is not to remove people from the loop. It is to move them to where they matter most.
What is a closed-loop system for diabetes?
A hybrid closed-loop system, sometimes called automated insulin delivery or an “artificial pancreas,” has three parts: a continuous glucose monitor that measures glucose under the skin, an insulin pump, and an algorithm that decides how much insulin to deliver based on the glucose readings. It is “hybrid” because the user still has to tell the system about meals and deliver meal-time insulin, while the algorithm handles background adjustments.
These algorithms are not usually large language models or the kind of AI in chatbots. Many use control theory and predictive models refined over decades. But they do what patients want from AI: take a relentless, data-heavy task and do much of it automatically.
Do closed-loop systems actually improve diabetes control?
In type 1 diabetes, the evidence is strong enough that health technology assessors now recommend them.
In December 2023 the National Institute for Health and Care Excellence in England published guidance, TA943, recommending hybrid closed-loop systems for adults with type 1 diabetes whose HbA1c is 58 mmol/mol (7.5%) or higher, or who have disabling hypoglycaemia, despite best possible management with a pump or glucose monitoring. It also recommended them for children and young people, and for people who are pregnant or planning pregnancy. Because of cost and capacity, NHS England planned a five-year phased rollout.
One of the most striking trials involved pregnancy, when tight glucose control matters for both mother and baby and is especially hard to achieve. The AiDAPT trial, published in the New England Journal of Medicine in 2023, randomized pregnant women with type 1 diabetes to a hybrid closed-loop system or standard care. Time spent in the target glucose range for pregnancy rose to 68.2% in the closed-loop group, compared with 55.6% with standard care, an adjusted difference of 10.5 percentage points. That is roughly two and a half extra hours a day in the target range.
What about AI apps for type 2 diabetes?
The picture is less clear, largely because type 2 diabetes is managed differently and the tools vary more.
Many people with type 2 diabetes do not use insulin, so there is no pump to automate. Instead, apps focus on coaching: logging food and activity, reminders to take medicines, feedback from glucose readings and messages from coaches, some human and some automated. In March 2024 the US Food and Drug Administration cleared the first over-the-counter continuous glucose monitor, intended for adults who do not use insulin, which has widened access to glucose data for people with type 2 diabetes and prediabetes.
Studies of coaching apps show a range of results. Some report modest improvements in HbA1c or weight; others show little difference once engagement fades. A consistent theme across digital health is that benefit depends on continued use and on whether someone acts on the data. We have written about the same pattern in remote patient monitoring: data does not help anyone until a person responds to it.
What happens when the technology fails?
This is where the human escalation path stops being a nice idea and becomes a safety requirement.
In 2024 the FDA classified a recall of a mobile app used with an insulin pump as Class I, the most serious category. A software problem caused the phone app to crash and relaunch repeatedly, generating excessive Bluetooth communication that could drain the pump’s battery and shut it down sooner than expected. When a pump stops, insulin delivery stops, which can lead to high blood sugar and diabetic ketoacidosis, a life-threatening condition. By mid-April 2024 the FDA had received reports of 224 injuries. A corrected version of the app was released.
The incident is a useful reminder of what automation changes. It does not remove risk; it moves it. The failure was not in the insulin or the algorithm’s judgment but in the everyday software layer connecting phone and pump. Patients who knew the warning signs, checked their pump and had a plan for switching to injections were protected. Those who trusted the system to handle everything were more exposed.
A worked example: automation with a safety net
This example is hypothetical.
Leila, 29, has type 1 diabetes and starts using a hybrid closed-loop system. For the first time in years, her overnight glucose stays stable, and she stops waking to alarms. Her time in range rises substantially.
Three months in, she gets a stomach bug. She is vomiting and cannot eat. Her glucose readings are not especially high, and the system is reducing insulin because she is not eating. But she feels increasingly unwell. Her diabetes team’s sick-day guidance says to check ketones in this situation. Her ketones are raised.
She calls the diabetes team’s helpline. They advise on insulin and fluids and tell her when to go to hospital if things do not improve. She recovers at home.
The system did its job, adjusting insulin to her glucose. But it could not know she was ill, and glucose alone did not show the danger. What protected her was a person who knew to check ketones and a team she could call. That is the human escalation path.
Now imagine the same scenario with no sick-day plan and no helpline. The technology is identical. The outcome could be very different.
The strongest counterargument: why not trust the machine more?
There is a respectable argument that “human in the loop” can become an excuse for keeping outdated processes.
Algorithms do not get tired at 3am. Closed-loop systems make many small adjustments a day that no person could match. Some patients have built their own open-source closed-loop systems, sharing designs online, because they felt official technology and health services were moving too slowly. For many people, the burden of diabetes is precisely the constant human decision-making, and handing more of it to machines is the point.
That argument is largely right about routine control. The case for human involvement is not that people should second-guess every insulin adjustment. It is about the exceptions: illness, device failures, pregnancy, changing needs, and the gradual drift that data review can catch. Automation should handle the routine so that clinicians and patients have more attention for the exceptions. A system designed that way is both more automated and safer.
What this means for patients, clinicians and health systems
For patients. If you have type 1 diabetes, ask your diabetes team whether a closed-loop system is appropriate and available to you. If you use any automated system, keep your sick-day rules, know how to switch to injections if a device fails, and keep backup supplies. For coaching apps, use them as support, not as a guide to changing medicines without advice.
For clinicians. Closed-loop systems shift work from constant dose adjustment to data review, education and managing exceptions. That work needs time in clinic schedules. Patients also need clear guidance on when to call and what to do if the technology fails.
For health systems. The NICE rollout shows that access depends on training and staffing as much as on device prices. Buying devices without the diabetes educators to support them risks disappointing results. It is the same lesson we have seen in diabetes prevention at scale: the programme around the tool determines the outcome. And as glucose data from consumer devices grows, as we noted in wearables as an undesigned screening programme, systems will need to decide who looks at it.
Frequently asked questions
What is the best AI tool for diabetes management?
There is no single best tool, and the right choice depends on your type of diabetes, treatment and needs. For type 1 diabetes, hybrid closed-loop systems have the strongest evidence and are recommended by NICE for many people. For type 2, coaching and glucose monitoring tools have more mixed evidence. Discuss options with your diabetes team.
Are closed-loop insulin systems safe?
They have improved glucose control in randomized trials, including in pregnancy, and are recommended by health technology bodies. Like any device, they can fail. A 2024 FDA Class I recall involved a pump app that could drain pump batteries, with 224 injuries reported. Users need training, sick-day rules and a backup plan such as insulin pens.
Can a diabetes app replace my doctor or diabetes nurse?
No. Apps and automated systems can reduce the daily burden and improve control, but they cannot recognise illness, adjust treatment plans safely in every situation or respond to emergencies. Your clinical team remains essential, particularly when you are unwell, pregnant or your needs change. Do not change medicines based on an app without advice.
Do diabetes coaching apps lower HbA1c?
Some studies show modest improvements, especially among people who stay engaged, but results vary and benefits can fade over time. Apps that connect data to a human coach or clinician who responds tend to be more useful than tracking alone. Ask your clinician whether a particular app suits your situation.
Should people with type 2 diabetes use a continuous glucose monitor?
Some people with type 2 diabetes benefit from continuous glucose monitoring, particularly those on insulin. The FDA cleared the first over-the-counter monitor for adults not on insulin in 2024. Whether it helps you depends on your treatment and goals. Talk to your clinician about whether it is useful and how to act on the readings.
Automation in chronic disease is no longer a promise; for many people with type 1 diabetes it is standard care. The systems that work best still have a phone number on the back.
If you live with diabetes or care for someone who does, where has automation helped most, and where did you still need a person?
