What Is Automated Insulin Delivery and How Does It Work?

Automated insulin delivery is a system that continuously reads a person’s glucose level, runs those readings through a software algorithm, and adjusts insulin from a wearable pump with little or no manual input. Often called an “artificial pancreas,” AID ties together three devices that already existed independently and makes them talk to each other in a closed feedback loop. The result is tighter blood sugar control, fewer dangerous highs and lows, and less of the round-the-clock decision-making that traditionally falls on people with diabetes and their families.

The Three Parts of the Loop

Every AID system has the same basic architecture. A continuous glucose monitor (CGM) sits just under the skin, usually on the abdomen or arm, and measures glucose in the fluid between cells every few minutes. That stream of numbers feeds wirelessly into a control algorithm, which can live on the insulin pump itself or on a separate device like a smartphone. The algorithm compares the current glucose level against a target range and predicts where glucose is heading. Based on that prediction, it tells an insulin pump to increase, decrease, or pause insulin delivery. The pump pushes rapid-acting insulin through a thin cannula inserted under the skin.

What makes the system “closed loop” is that human judgment is largely removed from the minute-to-minute dosing. In conventional pump therapy, you or your clinician program basal rates and you manually bolus for meals. In an AID system, the algorithm handles basal adjustments automatically and, depending on the system, can also deliver correction doses on its own. The person still wears and maintains the hardware, but the constant mental math of diabetes management shifts to software.

How the Algorithm Decides

Three families of algorithms power most AID systems. Model predictive control uses a mathematical model of how insulin and glucose interact in your body, running that model forward in time to choose the dose that keeps predicted glucose closest to target. Proportional-integral-derivative control reacts to the current glucose error, the accumulated error over time, and the rate of change. Fuzzy-logic systems encode expert clinical rules in software rather than relying on a single equation.¹ Model predictive control has become the most common approach in commercial devices because it can be tuned to individual patients and runs efficiently on small processors.²

All of these algorithms share a fundamental constraint: insulin delivered under the skin takes time to absorb and act. That lag, typically on the order of tens of minutes for rapid-acting analogs, means the algorithm is always working with outdated information and a tool that cannot be un-delivered once infused. The delays and variability of subcutaneous insulin absorption create both effectiveness and safety challenges that the algorithm must constantly navigate.³ This is why, despite impressive results, no current system perfectly mimics a healthy pancreas.

Hybrid Closed-Loop Versus Fully Closed-Loop

Most commercial AID systems today are hybrid closed-loop, meaning they automate basal insulin adjustments and sometimes correction boluses, but they still ask you to announce meals and enter an estimated carbohydrate count so the system can deliver a meal bolus. The algorithm handles the fine-tuning before and after, but it relies on you for the big upfront dose that covers food.

Fully closed-loop systems aim to eliminate meal announcements altogether. Pilot studies using open-source software have tested this scenario. In one randomized pilot, a fully closed-loop mode controlled glucose for about 95% of the study period, with time spent below the safe threshold staying under 1%. The percentage of time in the target range did not differ significantly between the hybrid and fully closed-loop modes in that small trial, suggesting that skipping meal boluses did not dramatically worsen control in a supervised setting.⁴ These are early results, and most clinicians still recommend meal announcements for the best outcomes, but fully closed-loop operation is where the technology is heading.

What the Clinical Evidence Shows

The metric that matters most in AID trials is “time in range,” meaning the percentage of the day spent with glucose between 70 and 180 mg/dL. International guidelines consider 70% or higher the goal for most adults with type 1 diabetes. Across randomized trials and real-world studies, AID systems consistently push time in range upward while reducing both highs and lows.⁵

In one randomized controlled trial of adults with type 1 diabetes who had suboptimal blood sugar control despite already using a pump and CGM, switching to AID increased time in range by about 19 percentage points over 14 weeks, while HbA1c dropped by roughly 0.9 percentage points. The control group, continuing with their usual pump setup, saw no change.⁶ Those are substantial improvements for people who were already using advanced technology; the gains tend to be even larger for those coming from manual injections or older pumps.

Where AID Shines Most: Overnight

Nighttime has always been the most dangerous period for insulin-dependent diabetes. You cannot feel a low while you sleep, and you cannot adjust a dose at 3 a.m. without waking up. AID changes this dramatically. In a randomized trial testing an overnight closed-loop system at home, hypoglycemia episodes with CGM values at or below 60 mg/dL occurred on about 21% of nights with the system running, compared with 33% of control nights. The total burden of hypoglycemia, measured by area under the curve, dropped by 81%, and prolonged lows lasting more than two hours fell by 74%.⁷

A separate crossover trial in children found that the percentage of overnight time spent in near-normal glucose (63–140 mg/dL) jumped from about 29% on standard pump therapy to 76% with closed-loop control, and mean overnight glucose fell by 36 mg/dL.⁸ The overnight period is where the algorithm has the easiest job, with no meals to deal with, and the results reflect that advantage.

The Safety Net: Predictive Low Glucose Suspend

Even systems that do not run a full closed-loop algorithm often include a simpler but life-saving feature called predictive low glucose suspend (PLGS). When the algorithm projects that glucose will fall below a user-set threshold within the next 30 minutes, it automatically stops basal insulin delivery. Once glucose starts rising again, insulin resumes on its own. If the pump and app lose connection for 30 minutes or more, insulin delivery restarts as a safety catch.⁹

A systematic review and meta-analysis of PLGS in children found high-quality evidence that the feature cut daily time spent in hypoglycemia by about 17 minutes and nocturnal hypoglycemia by about 26 minutes per night compared with standard sensor-augmented pump therapy, without increasing hyperglycemia or diabetic ketoacidosis risk.¹⁰ For families not ready for full AID, PLGS is often the entry point into automation.

AID in Very Young Children

Managing diabetes in toddlers and preschoolers is uniquely difficult. Their eating is unpredictable, their activity levels swing wildly, and they cannot communicate symptoms. AID systems have been tested and increasingly approved for this age group. A meta-analysis of randomized controlled trials in very young children found that AID improved time in range by about 9 percentage points and lowered HbA1c by roughly 0.4 percentage points compared with standard care, without increasing hypoglycemia or serious adverse events.¹¹

In a pilot of the Control-IQ system modified for children aged two to five, the proportion meeting prespecified glycemic goals rose from 33% at baseline to 83% during AID use, and time in range climbed from about 64% to 71%.¹² A larger crossover trial of the MiniMed 780G system in children aged two to six showed that the automated mode achieved mean time in range of about 68% versus 58% in manual mode, with a between-treatment difference of roughly 10 percentage points and no episodes of severe hypoglycemia.¹³ For parents of very young children, the overnight safety gains alone can be transformative.

AID for Type 2 Diabetes

Until recently, AID research focused almost exclusively on type 1 diabetes. That is changing fast. A landmark randomized trial published in the New England Journal of Medicine tested AID against standard insulin therapy in adults with type 2 diabetes. Over 13 weeks, time in target range rose from about 48% to 64% in the AID group while barely budging in the control group. HbA1c dropped by 0.9 percentage points with AID versus 0.3 in controls.¹⁴

A nonrandomized clinical trial found similar results, with HbA1c falling from 8.2% to 7.4% and time in range jumping from 45% to 66%, improvements that held regardless of age, sex, race, insurance status, or whether participants were also taking other diabetes medications like GLP-1 receptor agonists.¹⁵ In a fully closed-loop crossover trial in adults with type 2 diabetes, the contrast was even starker: time in range nearly doubled from about 32% under standard insulin therapy to 66% with closed-loop control, and HbA1c was 1.4 percentage points lower after the closed-loop phase.¹⁶ People with type 2 diabetes on intensive insulin regimens may ultimately benefit as much as, or more than, those with type 1.

Open-Source Systems Built by the Community

Before any commercial AID system reached the market, a community of people with diabetes and software developers built their own. Projects like OpenAPS, Loop, and AndroidAPS reverse-engineered older insulin pumps and wrote open-source algorithms that anyone could download. These do-it-yourself systems have been used by thousands of people worldwide, and researchers have now subjected them to formal clinical trials.

The CREATE trial, a 24-week randomized controlled trial of an open-source AID system, found that time in range increased from about 61% to 71% in the AID group while actually declining slightly in controls. Participants using AID spent over three additional hours per day in the target glucose range. No severe hypoglycemia or diabetic ketoacidosis occurred in either group.¹⁷ A 24-week extension of the trial to 48 weeks confirmed that the benefits were sustained and safe across different insulin pump types and age groups.¹⁸ Open-source systems remain officially unapproved by regulators, which creates an uncomfortable tension: the evidence is solid, but the devices are not legally marketed, and people use them at their own risk.

The Psychological Payoff

Blood sugar numbers get the headlines, but much of what AID changes is psychological. Managing diabetes requires hundreds of decisions a day, and the cumulative burden takes a real toll. A systematic review and meta-analysis found that AID use led to a small but significant reduction in diabetes distress among adults. The effect was more pronounced for caregivers of children with diabetes, who experienced a moderate reduction in distress. Interestingly, the benefit was larger for parents of younger children compared with parents of teenagers.¹⁹

Sleep is another major area. Parents of children with type 1 diabetes often wake multiple times per night to check blood sugar. A meta-analysis of AID’s effect on sleep found that caregivers reported improved sleep quality when their child used an AID system, with significant improvements seen in both randomized trials and prospective studies.²⁰ The algorithm does the overnight worrying for you, and that matters to quality of life in ways that HbA1c alone cannot capture.

Alarm Fatigue Is Real

AID systems come with a trade-off that manufacturers tend to downplay: alarms. Lots of them. A cross-sectional study of over 800 people with diabetes found that AID users reported the highest alarm frequency of any technology group, with about half experiencing alarms several times daily compared with roughly a third of those using a CGM alone. AID users also reported greater alarm disruptiveness, annoyance, and unwanted attention. They were nearly twice as likely to report frequently ignoring alarms and about 60% more likely to overcorrect low glucose readings in response to alerts.²¹

This creates a paradox. The system that is supposed to reduce your mental burden also buzzes, beeps, and vibrates throughout the day and night. Over time, people tune out. Ignoring an alert because most alerts feel unnecessary is understandable, but the one you ignore might be the one that mattered. Manufacturers are working on smarter alert logic, but for now, alarm fatigue remains one of the most common complaints among AID users.

Exercise Remains a Tough Problem

Physical activity is one of the scenarios where AID systems struggle most. Exercise drops blood sugar through mechanisms that are hard for an algorithm to predict: muscles take up glucose independently of insulin, and the effect can last for hours afterward. A joint position statement from the European Association for the Study of Diabetes and the International Society for Pediatric and Adolescent Diabetes reviewed the evidence and offered detailed practice points for managing exercise with AID, acknowledging that glucose fluctuations during and after activity continue to challenge current systems.²²

Most AID systems have an “exercise mode” or “activity target” that raises the glucose target temporarily, causing the algorithm to back off on insulin. But users often need to activate this manually before starting exercise, and the timing is tricky. Start too late and the insulin already on board pulls glucose down. Start too early and glucose runs high before you begin. Some people reduce their meal bolus before exercise or eat extra carbohydrates, layering manual strategies on top of the automation. The algorithm helps, but it does not solve exercise on its own.

Dual-Hormone Systems on the Horizon

Current AID systems deliver only insulin. A healthy pancreas, by contrast, also releases glucagon to raise blood sugar when it drops too low. Dual-hormone systems aim to replicate both sides of this balance by pairing insulin delivery with small doses of glucagon. The idea is that adding glucagon gives the algorithm a tool to actively push glucose up, not just passively wait for insulin to wear off, which could reduce hypoglycemia further and allow tighter overall control.²³

Several research groups are testing dual-hormone prototypes, and simulation studies suggest they can keep glucose in the normal range more effectively than insulin-only systems, particularly during unannounced meals or unexpected activity.²⁴ The main barriers are practical: stable liquid glucagon formulations have only recently become available, and wearing two separate infusion sets adds bulk and complexity. No dual-hormone system is commercially available yet, but it remains one of the most anticipated next steps.

Infusion Site Failures and Other Weak Links

An AID system is only as good as its weakest hardware component. One of the most common real-world failures is an infusion site problem, where the cannula kinks, dislodges, or loses absorption efficiency, and insulin that the pump records as delivered never actually reaches the bloodstream. Because the algorithm “thinks” it delivered a dose, it may not compensate quickly enough, leading to unexpected highs and, in severe cases, ketosis. Research efforts are underway to build real-time detection of infusion site failures into AID algorithms so the system can alert users before the situation becomes dangerous.²⁵

CGM sensor drift is another issue. Sensors become less accurate toward the end of their wear period, and calibration errors can mislead the algorithm. Adhesive failures, signal dropouts from Bluetooth interference, and smartphone app crashes all interrupt the loop. None of these problems are catastrophic on their own, but they mean that AID requires ongoing engagement. You cannot simply set it and forget it.

Who Gets Access

AID technology works, but it does not reach everyone equally. A study of youth with type 1 diabetes at a single center found that fewer than half of Black youth were using AID, compared with 70% of white youth. Even after statistically adjusting for socioeconomic factors like household income and insurance type, Black youth were still about 7 percentage points less likely to use AID.²⁶ The reasons are complex: referral patterns, implicit bias in clinical recommendations, trust, and the sheer logistical burden of initiating and maintaining the technology all play a role. Insurance coverage varies widely, and even when coverage exists, the out-of-pocket costs for sensors, pump supplies, and transmitters add up. Expanding access is as much a policy and equity challenge as a technological one.

Cybersecurity in a Life-Critical Device

An AID system is a networked medical device that wirelessly controls a drug that can kill you in the wrong dose. That combination raises cybersecurity concerns that go beyond privacy. Wireless communication between the CGM, the algorithm, and the pump creates potential attack surfaces. A compromised system could theoretically alter insulin delivery, leading to life-threatening hypoglycemia or sustained highs.²⁷

No real-world attack on a commercial AID system has been publicly documented, and manufacturers encrypt their communication protocols. But security researchers have demonstrated vulnerabilities in older pump models, and the open-source community, by definition, exposes its code publicly. Regulatory agencies now require cybersecurity risk assessments as part of the approval process for new devices. For users, the practical advice is straightforward: keep firmware updated, use strong passwords on connected apps, and be cautious about third-party modifications to any system that delivers insulin.

Reinforcement Learning and the Next Generation of Algorithms

Current commercial algorithms follow fixed rules or models. The next wave may learn and adapt on the fly. Researchers are exploring reinforcement learning, a branch of artificial intelligence where the algorithm improves its dosing strategy through simulated trial and error. In one recent study, reinforcement learning policies trained on virtual patients with type 1 and type 2 diabetes kept blood glucose in the normal range for a significantly higher percentage of the day and spent significantly less time in hypoglycemia compared with standard control algorithms, even when faced with unannounced meals.²⁸ These results are from simulations, not real patients, so the leap to clinical use is still ahead. But the direction is clear: algorithms that learn your specific physiology over time, rather than relying on a one-size-fits-most model, are where the field is heading.