Automated Insulin Delivery (AID):
A Loop That Helps, Not a Cure
Everything in one place. Read the plain version with Jude, earn your way into the evidence with Grace, then the full model with John. Stop wherever you have enough.
How we teach: three rules, borrowed from Taleb
You earn each level by showing you understand it, not by scrolling past it. We only teach what we would use on ourselves and the people we love.
Understanding beats memory and luck, so the checks reshuffle every time you retry. A pass means you got it, not that you guessed it. And we teach you to tell a trend (signal) from one reading (noise).
We give you the scaffolding and get out of your way. Roam where your curiosity leads, go as deep as you want, and ask Grace anything. We will not teach a bird how to fly.
Got a question about your own system or settings? Ask Grace, then take it to your care team.
One page, three depths
This guide compounds: each layer rests on the one beneath it. Read Jude’s plain version, then pass a short understanding check to open Grace, then another to open John. You can roam freely within a layer; you cannot skip ahead a layer, because the next one would not make sense and you would be standing on a gap.
The whole thing, in plain words
An automated insulin delivery system, often called an AID system or a hybrid closed loop, is three things working together: a continuous glucose monitor reading your glucose every few minutes, an insulin pump that drips insulin in, and a small computer program (an algorithm) that connects the two. The CGM tells the algorithm where your glucose is heading; the algorithm nudges the pump to give a little more or a little less background insulin to match. It is doing this all day and all night, in the background, while you get on with your life.
Here is the honest part. It is called a hybrid loop for a reason: it is a partnership, not a robot you can forget about. You still tell it about meals (this is the bolus, the dose for food), and you still change the parts that touch your skin. The loop handles the slow, steady background work brilliantly; it cannot read your mind about the sandwich you are about to eat.
What does it do for you? For most people it means more of the day spent in a good glucose range, fewer of the dangerous lows, better sleep, and a quieter mind because the system catches drifts before you would. It is a tremendous help, but it is not a cure. Type 1 diabetes does not go away; the loop just shares the load. And no system is “the best” one; the evidence says the popular systems land in much the same place when set up well, so the right choice is the one that fits your life and your skin, decided with your team.
Does this match the life of the person living it? A loop that lifts the night-time worry can change a family’s whole week, and that matters as much as any number on a screen. If the technology ever makes you feel watched rather than helped, that is worth saying out loud to your team.The Pemberton lens, lived recognisability, one of the four GNL appraisal lenses.
How the loop adjusts, and what the trials show
How an AID system adjusts insulin
Every commercial AID system works the same way at its core: it modulates background (basal) insulin, and on the more automated systems delivers small automatic correction doses, using a calibrated model of how much insulin on board (IOB) you already have. High IOB tells the algorithm to hold back; low IOB gives it permission to act.1 Systems are sometimes described by a ladder of capability: low-glucose suspend (reacts at a low), predictive low-glucose suspend (anticipates a low), hybrid closed loop (adjusts basal continuously), and advanced hybrid closed loop (adds automatic corrections).2
You still announce meals. An AID system anticipates a meal far better when you tell it food is coming than when it has to chase the rise afterwards. That single behaviour, meal announcement, is one of the largest things in your control on a loop, and we return to it below.
The evidence on time-in-range gains
Across pivotal randomised trials and large real-world registries, AID raises time in range (glucose between 3.9 and 10.0 mmol/L, 70 to 180 mg/dL) by roughly 10 to 15 percentage points compared with a sensor-augmented pump or injections with CGM.3 The gain is largest for people starting from a high baseline: in the NHS England real-world pilot, time in range rose from about 34% to about 62% (a gain near 28 percentage points) in a group with a mean starting HbA1c near 9.4% (79 mmol/mol), and that gain held at twelve months.4
| Setting | Comparator | Time-in-range change | Grade |
|---|---|---|---|
| Pooled meta-analysis (Ferreria 2024) | SAP / MDI+CGM | about +11 to +13 points | A |
| Control-IQ pivotal (Breton 2020) | Sensor-augmented pump | about +11 points | A |
| NHS real-world pilot (Crabtree 2023; Liarakos 2025) | Pre-loop, high baseline | about +28 points, held at 12 months | B |
| From diagnosis at 48 months (CLOuD) | Multiple daily injections | about +12 points, sustained | A |
One quieter finding is the most human. In the same NHS pilot, the share of people with high diabetes distress fell roughly three-fold, and around 95% said the loop had a positive impact on their life.4 For many people the relief of the night-time burden matters as much as the glucose curve.
What an AID system can and cannot do
It can: lift time in range, reduce the most dangerous lows (one large youth registry found hypoglycaemic coma was about a third less likely on a loop than on a pump without automation), and ease the mental load.5
It cannot: replace your meal boluses (removing meal announcement cost about 9 percentage points of time in range in a controlled trial), and it does not preserve beta-cell function. The CLOuD trial put a loop on from the moment of diagnosis and still found no protection of the body’s own insulin production beyond what good glucose control would give; an important null result honestly reported.6 7 A loop shares the work; it does not switch the condition off.
A loop lowers severe-hypo risk but raises the risk of diabetic ketoacidosis (DKA) compared with a pump without automation, because when an infusion site silently fails, the system keeps showing insulin on screen while none is reaching the body. In one youth registry the DKA rate was higher on the loop (incidence rate ratio about 1.8), and the excess was concentrated in those with the highest HbA1c.5 The practical guard: check ketones for any unexplained high that auto-corrections do not fix within a couple of hours, and discuss site-change habits with your team.
A headline TIR gain is an average across a trial population, not a promise to you. Whenever you are handed a benefit figure, ask the two questions that keep everyone honest: out of how many, and compared with what? A 28-point gain from a high baseline is a different animal from an 11-point gain in a well-controlled trial cohort.The Goldacre lens, evidence-grade discipline, one of the four GNL appraisal lenses.
The IOB trade-off, the algorithms, and the Optimiser
Where the real risk lives: insulin on board
Every AID system uses insulin on board as a brake. The displayed IOB is a calibrated model output, not a measurement of how much insulin is actually in your body.1 The settings that shape that model (active insulin time, glucose target, correction factor) are therefore the primary algorithm-strength lever on every system. Turn the strength up and you buy time in range; you also concentrate more active insulin into each moment, and in the real world the hypoglycaemia that matters tends to come from stacked IOB sitting on top of higher-strength settings. That trade-off is the dominant safety driver in day-to-day use, and it is under-represented in stock manufacturer guidance.8
Algorithm differences, at a conceptual level
The systems differ in philosophy, rule-based versus model-predictive, and in which single setting does the most work. Each has one primary lever:
| System | Primary lever | What turning it does |
|---|---|---|
| Tandem Control-IQ (and Mobi, same family) | Correction factor (CF) | The algorithm leans on CF for its automatic corrections |
| MiniMed 780G | Active insulin time (AIT) | A shorter AIT tells it insulin clears faster, prompting more frequent corrections |
| CamAPS FX | Insulin-to-carb ratio (ICR) | A stronger ICR pushes more delivery into the algorithm-controlled component |
| Omnipod 5 | Glucose target | The lowest target delivers roughly 4 to 5 points more time in range than higher targets |
The headline is counterintuitive and well evidenced: at the population level the major systems land in much the same place on time in range when groups are matched on baseline control. Three independent studies converge here, an adult matched comparison, a youth matched comparison, and a UK three-system real-world series, none finding a clinically meaningful between-system difference.9 Within any one system, the gap between best-configured and worst-configured users is larger than the gap between systems: in one registry of over 100,000 users on the 780G, only about 6.4% were on the optimal-outcome setting combination.10 What happens after initiation matters more than which badge is on the pump.
The AID Optimiser, and how we position it
GNL builds an educational tool called the AID Optimiser: a five-level ladder of algorithm strength across CamAPS FX, MiniMed 780G, Tandem Control-IQ, Tandem Mobi and Omnipod 5, showing which levers an experienced professional would consider at each level. Its positioning is locked and I will not soften it.
The five-level ladder, and the drivers and settings adjusted across each level, have been reviewed and refined with input from CamAPS, MiniMed, Tandem and Insulet global medical leads. Their input shaped which levers are exposed at each level and how they are described.
However, the levels themselves have not been validated against any manufacturer’s internal simulator or proprietary dataset, so this is not a manufacturer endorsement of the GNL ladder. The Optimiser is a Grade D suggestion layer built on a Grade A and B evidence base (peer-reviewed trials, registry data and pharmacokinetic studies).11
It carries a deliberate, declared bias toward IOB visibility, because that trade-off is the dominant real-world safety driver. Any deviation from manufacturer-recommended starting settings is a clinical decision made with your diabetes care team. Reviewed by, not endorsed by, not validated by, not co-developed with.
It makes no claim that any one system is superior; the evidence does not support that, and the tool says so. It does not give you a personal dose, and it is not a recommendation for your settings; it shows the shape of the conversation to have in clinic.
The limits, named honestly
Three to keep in view. First, the between-system equivalence is a population finding; individuals genuinely do better on one system than another, and those experiences are real even though they do not aggregate into superiority. Second, the strongest real-world numbers come from expert centres with mostly White cohorts and no control group; read the 28-point gain as the ceiling for a high-baseline, well-supported start, not a universal promise. Third, the loop is only as good as what reaches the body: the DKA-via-silent-site-failure mechanism is the price of automating delivery, and ketone vigilance is the mitigation, not an optional extra.
The loop is a brilliant manager of the average Tuesday; it is the rare, asymmetric event that should hold your attention. A silent site failure on a loop is exactly the fat-tail risk worth engineering against, because its cost (DKA) is out of all proportion to its frequency. Optimise for surviving the bad day, not for a prettier good one.The Taleb lens, robustness to outliers, one of the four GNL appraisal lenses.
The Optimiser is a model, and a model is only as honest as its declared assumptions. Ours is a Grade D synthesis with an openly stated IOB bias, sitting on a Grade A and B substrate, reviewed but never validated against a manufacturer simulator. That declaration is the integrity of the tool. Name what would strengthen it (a simulator validation), and never sell the ladder as the territory.The Hayes lens, technical and methodological rigour, one of the four GNL appraisal lenses.
The whole guide, summarised
Glucose never lies; a loop just listens to it minute by minute and shares the steering. The destination is set with your team.
This page is the taster. The full journey, three modules and their 30 questions, with your progress saved, lives in Learn with Grace. Glucose never lies; come and learn to read the loop.
References
Evidence grades A (strongest) to D (editorial or working analysis).
- GNL Grace concept pages: Insulin On Board and The IOB Trade-Off, synthesising pump and AID IOB-model behaviour. B
- GNL Grace teaching frame, predictive CGM and AID capability ladder (LGS, PLGS, HCL, AHCL, FCL). D (educational synthesis)
- Ferreria et al. 2024, Cambridge HCL meta-analysis, Lancet Digital Health; pooled mean TIR gain about +11 to +13 points vs control. Breton et al. 2020, Control-IQ pivotal, NEJM 383:836-845 (TIR about +11 vs sensor-augmented pump). A
- Crabtree et al. 2023, ABCD DTN-UK NHS England HCL audit, Diabetes Care 46(10):1831-1838 (n=570; TIR 34.2% to 61.9%); Liarakos et al. 2025, 12-month durability extension, Diabetes Technology and Therapeutics (TIR +26.7 points sustained). B
- Karges et al. 2024, DPV registry, Lancet Diabetes and Endocrinology (n=13,922 youth aged 2-20; vs sensor-augmented pump): DKA IRR 1.81 (95% CI 1.37 to 2.40); hypoglycaemic coma IRR 0.68 (95% CI 0.48 to 0.97). DKA excess concentrated at higher HbA1c. B
- Garcia-Tirada et al. 2023, Diabetes Technology and Therapeutics (n=102, 780G cross-over): meal announcement vs algorithm-only, 86% vs 77% TIR, a 9-point difference. A
- CLOuD 48-month RCT, Ware et al. 2024, Lancet (n=97, CamAPS FX from diagnosis vs MDI): TIR +12 points sustained; C-peptide NOT preserved (null result). A
- GNL Grace, The IOB Trade-Off; IOB-vs-algorithm-strength as the dominant real-world safety driver; under-represented in stock manufacturer guidance. D (educational synthesis on a Grade A/B base)
- Beato-Vibora et al. 2024 (adults, matched, 780G vs Control-IQ, no significant TIR difference); Gera et al. 2025 (youth, matched, OP5 vs CIQ, p=0.08); Khan et al. 2026 (UK three-system real-world, no clinically significant difference). B
- Choudhary et al. 2024, 780G real-world registry (n=101,629): only 6.4% on the optimal-outcome setting combination (AIT 2h plus target 5.5 mmol/L). B
- GNL AID Optimiser Positioning Policy (locked 1 May 2026), and NICE TA943 Hybrid Closed-Loop Systems for T1D (2023), which evaluates HCL as a class and endorses no specific device. Optimiser is a Grade D synthesis on a Grade A/B base; reviewed by, not endorsed by, the four manufacturers. D synthesis; NICE TA943 A
One page, three voices: Jude, Grace, John. Population-average, not personalised.
Keep learning: HbA1c and Time in Range · CGM · Mealtime insulin · Hypoglycaemia · Exercise · Type 1 diabetes