A robot found the vein, cleaned the skin, put the needle in and swapped the tubes — 1,633 times, and beat trained humans on the hardest arms. The FDA authorised it on 19 August. Its own press release contains six words almost nobody quoted.

Sponsored by

🩸 The machine that decides whether to try

There is a moment in every blood draw that nobody talks about, because until last week only humans had it.

A phlebotomist takes your arm, presses two fingers into the crook of your elbow, and feels around. In that second or two, they are answering a question: can I do this one? Most of the time the answer is yes and you feel a scratch. Sometimes it’s “let me try the other arm.” Occasionally it’s “I’m going to get someone else.”

On Wednesday 19 August 2026, the US Food and Drug Administration authorised a machine that asks itself the same question — and, when the answer is no, refuses to pick up the needle.

The device is called Aletta. It’s made by Vitestro, headquartered in Utrecht in the Netherlands. The FDA calls it the first standalone robotic device that can draw blood from a patient’s arm “without hands-on operator intervention.”

Here is the sentence from the FDA’s own press release that this entire issue is about:

Six words: when it proceeds with a stick.

They are doing an enormous amount of work, they are entirely honest, and they are the first thing that falls off when the headline number goes out into the world.

🤖 What it actually does, step by step

Start with what’s genuinely impressive, because a lot of it is.

You sit down and put your arm in the machine. A trained phlebotomist starts the session. Then either you or the supervisor presses a button, and the robot takes over:

1. Near-infrared light picks the starting position for the ultrasound probe. It doesn’t choose the vein — it tells the ultrasound where to look.

2. The machine sprays 70% alcohol on the skin and scans the crook of your elbow with ultrasound, hunting for a vein deep enough, wide enough and straight enough to take a needle.

3. Doppler ultrasound listens for flow. This is the step that tells a vein from an artery — arteries move blood with a pulse, and hitting one is the mistake you least want a machine to make.

4. If nothing suitable turns up, it stops. No needle comes out. You get sent to a human.

5. If it finds a vein, it runs the rest on its own: tourniquet, needle in, tubes swapped as they fill, needle out and into the sharps bin, bandage on.

(The FDA’s summary and the published paper list the tourniquet and skin-prep steps in slightly different orders. Nothing turns on it.)

A supervising phlebotomist stays available throughout, and afterwards checks two specific things: that the collection tubes were filled in the correct order, and that each one is adequately full. Both matter more than they sound — additives carry over between tubes, and an underfilled coagulation tube produces a wrong answer rather than no answer.

The safety layers are sensible and, unusually, mostly mechanical rather than algorithmic:

The FDA notes that one phlebotomist can oversee up to three Aletta units at the same time. That ratio is the commercial argument, and we’ll come back to it, because it’s where the six words start to bite.

📈 The numbers are good. Genuinely good.

The evidence behind this isn’t a press release. It’s a peer-reviewed multicentre trial — Performance, Safety, and Patient Experience of an Autonomous Robotic Phlebotomy Device — published in Clinical Chemistry (DOI 10.1093/clinchem/hvag029), part of the ADOPT trial (NCT05878483). It ran between 1 May and 30 September 2025 across one academic hospital, two regional teaching hospitals and their outpatient phlebotomy departments in the Netherlands — Amsterdam UMC, St Antonius Hospital and OLVG Lab among them.

1,633 patients had blood drawn by the robot.

Read the difficult-veins row again. 92.7% against a manual literature range of 75%–85%. If you’re one of the people who gets described, apologetically, as “a hard stick” — who has been dug around for, who has left a phlebotomy chair with three plaster marks and no blood taken — the machine is better at you than people are, provided it agrees to try. Hold that caveat. It’s the whole of the next section.

The paper’s explanation for the hard-veins result is one sentence long:

Ultrasound sees veins that fingers can’t feel. The device uses a 21G single-use collection set and, in the authors’ words, prevents “blind sticks” and needle probing.

And of the roughly 940 patients who answered the survey:

That’s the story most outlets ran, and they weren’t wrong to run it. It’s a good story.

One quick word from today’s sponsor — then the turn.

Build. Break. Fix. Learn.

KodeKloud gives you 1,280+ hands-on labs where you provision Kubernetes clusters, write Terraform configs, build CI/CD pipelines, configure Linux systems, containerize apps with Docker, automate with Ansible, and manage Git workflows. 

78+ playgrounds let you experiment freely in sandbox AWS environments, Kubernetes clusters, and CI/CD systems without risk. 

190+ courses across DevOps, Cloud, and AI pair theory with hands-on labs at every step. 

KodeKloud Engineer and 100 Day Challenges provide real-world job scenarios with automated grading that confirms your solutions work.

Stuck? The 55,000+ member Discord community connects you with peers and instructors ready to help. 

Every lab runs in a live environment. You deploy, you troubleshoot, you learn. No videos without context. No simulations. The kind of practice that actually builds confidence because you've done real work, not watched someone else do it.

Start Learning For Free

🎯 Now the six words

Go back to the FDA’s phrasing: comparable to or better than trained human phlebotomists when it proceeds with a stick.

The obvious question is: how often does it not proceed?

The paper answers directly, in plain language, without being asked. Credit where it’s due — this is the manufacturer’s own study volunteering the number that undercuts its best headline:

And then, in the discussion, the sentence that should have travelled:

So there are two true numbers:

Neither is spin. The 94.5% is the correct answer to “how good is the robot at hitting a vein it has found.” The 88.5% is the correct answer to “if I send a hundred patients to this machine, how many walk out with the blood taken.”

They differ by six percentage points, and those six points land on somebody. The paper reports the screened-out rate broken down by subgroup in Supplemental Table 6 — which is exactly where anyone buying one of these should look first, and exactly where a journalist won’t. What the paper doesn’t do is put that breakdown in the main text, next to the 92.7%.

The machine doesn’t fail on the patients it can’t read. It declines them. That’s a far better design than failing on them. But an exclusion is not a success, and the difference walks back out to the waiting room.

🧮 Where this bites: the 1:3 ratio

The commercial case is one phlebotomist supervising three devices. The FDA explicitly frames the authorisation as a response to workforce pressure, and the pressure is real: US Bureau of Labor Statistics data puts phlebotomists at about 139,700 jobs in 2024, growing 6% to 2034, with roughly 18,400 openings a year — and by BLS’s own arithmetic, almost all of those are replacing people who leave rather than new posts.

Michelle Tarver, M.D., Ph.D., who directs the FDA’s Center for Devices and Radiological Health:

All true. Here’s the modelling error waiting to happen.

If you staff on the assumption that one person covers three machines and every patient is a machine patient, you have quietly assumed a 100% eligibility rate. The trial says 94%. Six in every hundred patients get handed back — and handed back to the one human you were counting on to be supervising, not sticking.

That’s not an argument against the device. It’s an argument for a specific operational number:

⚠️ Four more things worth knowing before the robots arrive

1. The performance cohort was single-arm. Cohort 1 (n=153) did run a randomised, within-subject comparison against a manual draw from the other arm — but that was to show the robot’s tubes give the same lab results, not that it’s better at finding veins. Every performance and safety number above comes from cohort 2, which the paper describes as a “single-arm study with ARPD phlebotomy replacing manual phlebotomy.” There was no randomised head-to-head against human phlebotomists working the same chairs on the same day. So “better than humans” means better than the literature’s humans — a legitimate and common design, but not the same claim.

2. The company wrote the paper. The co-first author’s affiliation is the Department of Clinical Affairs, Vitestro, and he is Vitestro’s Chief Medical Officer; the other co-first author is also Vitestro. Four of the fifteen authors are Vitestro staff. Independent clinical chemists from the participating hospitals are co-authors, the study is peer-reviewed and open access, and the awkward numbers are in the text — that’s the good version of this. Worth adding, though: Mayo Clinic appears three times over — as a co-author’s institution, as the site of the US acceptance test, and as a strategic investor in Vitestro’s $70m Series B, announced 10 March 2026 alongside Labcorp Venture Fund and Sutter Health. None of that makes the data wrong. All of it is context a reader deserves.

3. The robot is not the whole chain — and the trial proves it. Overall haemolysis was 0.3%. Split by site, it’s a different picture:

Sites 2 and 3 recorded zero haemolysed robot samples. Site 1 recorded 5.2% — but that’s five samples out of ninety-six, with a 95% confidence interval of 1.7% to 11.7%, which comfortably contains that site’s 6.9% manual rate. You cannot say the robot beat the humans there. What you can say is that site 1’s manual rate was five times site 2’s, across tens of thousands of samples. Whatever was wrecking those tubes was local — handling, transport, timing. The machine neither caused it nor fixed it. Automating the needle does not automate the twenty minutes after the needle.

4. The skin-tone question is open, and the two documents don’t line up. The FDA states its review covered patients “with varying skin tones.” The published multicentre trial says, flatly: “No ancestry information was collected.” Not necessarily a contradiction — the FDA reviewed a submission dossier that isn’t public and goes beyond this paper. And the strongest reason not to panic is in the engineering: vein selection here is ultrasound and Doppler. Sound, not light. Pigment-independent. But near-infrared still sits upstream, aiming the probe, and medicine has been burned by light-through-skin assumptions before — pulse oximetry was in ubiquitous clinical use for decades before the size of its skin-pigmentation error was properly characterised. The right posture isn’t alarm. It’s: the public evidence doesn’t answer this yet, and post-market data should.

🛡️ What good looks like from here

This is a well-designed device that fails safe, and the FDA has fenced it carefully. The De Novo authorisation creates a brand-new device classification, which means the special controls published alongside it — covering labelling, performance testing and clinical testing — become the bar every competitor must clear. That’s the regulatory system working as intended: the first device through the door sets the standard for everyone behind it.

The limits are written into the authorisation itself:

(A small wrinkle: the trial enrolled from age 16, median 58, range 17–93. The authorisation is adults-only.)

A nine-point version of what to actually do:

1. Labs: publish both numbers. First-stick success and overall collection success including screened-out patients. Reporting only the first is technically true and practically misleading.

2. Treat the screened-out rate as a staffing input, and measure it on your own population rather than importing 6% from a Dutch cohort.

3. Keep experienced human phlebotomists on the hard cases. The machine concentrates difficulty rather than removing it — the people it hands back are the ones who most need a good stick.

4. Don’t buy it as a sample-quality fix until you’ve audited your own preanalytical chain. Site 1 is the warning.

5. Enforce the supervisor’s two checks — tube order, tube fill. They’re the human step the design depends on, and exactly the sort of check that erodes when one person watches three machines and nothing has gone wrong for six months.

6. Watch for automation complacency. A supervisor who has never seen a fault in 4,000 draws is a different supervisor from the one on day one. That’s a training and rota problem, not a device problem.

7. Ask for post-market performance broken down by skin tone, and treat its absence as an open question rather than a settled one.

8. Patients: you can ask. If you’re told the machine wouldn’t attempt your draw, that’s the device working correctly — not a verdict on your veins. Ask to be sent to a person.

9. It’s an authorisation, not an approval. Fierce Biotech ran “Meet Aletta: The first FDA approved robotic blood draw device” on 20 August; iTechPost did the same. De Novo authorisation and premarket approval are different regulatory acts with different evidentiary bars. That’s not pedantry — it’s the difference between “novel, low-to-moderate risk, special controls apply” and something considerably heavier.

🧷 The thing I keep coming back to

The best engineering decision in this device is a refusal.

Almost everything impressive about Aletta — the infrared, the Doppler, the tourniquet, the tube changer, the needle that lets go when you flinch — is a machine doing something. The single most important thing it does is not do something: when it can’t find a vein it’s confident about, it puts nothing in your arm and calls a human.

Six percent of the time, that’s the answer. And the number that tells you how well this technology actually serves patients isn’t the one measuring how good it is when it’s sure.

It’s the one that counts the times it wasn’t.

📚 Sources

This issue went through an adversarial fact-check before publication. It changed nine things, including a piece of arithmetic I’d got wrong about site 1 and a claim about the needle set that no source actually supported. Where the FDA’s summary and the published paper describe something differently, both are quoted rather than reconciled.