Artificial intelligence is arriving in allergy and immunology the same way it arrived everywhere else in medicine: fast, unevenly, and with more promises than evidence. This guide is for allergists, allergy nurses and practice managers who want a clear picture of where AI actually helps an allergy practice today, what the published evidence says, where the risks are, and which questions to ask before letting any AI tool near a patient record.
I build Medora, the AI copilot built only for allergy practices, together with a working allergy practice, Allergy Affiliates. That gives me a point of view, so I've separated three things on this page: what the literature says (cited), what Medora does (labeled), and what I think (stated as opinion).
What "AI" means in an allergy practice
"AI" is used loosely. In practice, three families of technology matter to an allergy clinic:
- Language models and speech recognition. They turn a recorded visit into a draft note, summarize prior visits, or answer a patient's question in a chat window. This is the technology behind ambient "AI scribes".
- Computer vision. Models that find and size things in an image, such as wheals on a skin prick test panel or reactions on a patch test. This is the newest and most tightly regulated area.
- Predictive and statistical models. Models trained on structured data to estimate a risk or a probability, for example of sensitization. These are mostly in research, not routine clinical use.
The American Medical Association prefers the term augmented intelligence, which it describes as a conceptualization of AI that focuses on its assistive role, enhancing human intelligence rather than replacing it.[1] That framing fits allergy well. Almost every clinically meaningful decision in an allergy visit (is this sensitization clinically relevant? is this patient a candidate for immunotherapy? is this reaction irritant or allergic?) depends on the history, which only the clinician can weigh.
What the specialty literature says
The allergy literature on AI has grown quickly, and most of it is reviews and frameworks rather than trials. Four sources are worth knowing:
- The AAAAI framework (2022). An AAAAI Health Informatics, Technology, and Education Committee work group did a scoping review of AI in health care and in allergy and immunology. It found many potential applications, from diagnosis to reducing multidimensional data in electronic health records, and concluded that specialists should be involved in the design, validation and implementation of AI in the specialty.[2]
- A systematic review (2025). Goktas and Damadoglu, in Annals of Allergy, Asthma & Immunology, screened 192 studies and selected 20. They describe AI models that predict allergen sensitivity and work on immunotherapy strategies, mostly at the research stage.[3]
- An ethics review (2025). Writing in the World Allergy Organization Journal, González-Díaz and colleagues warn of the risk of diminishing clinical reasoning skills through over-reliance on AI, and of challenges with data privacy, informed consent and algorithmic transparency.[4]
- Chatbots (2023). A JACI: In Practice commentary on large language model chatbots in allergy stresses bias, privacy and the need to verify AI-generated findings.[5]
The honest summary: the specialty is interested, the frameworks are sensible, and the allergy-specific outcome evidence is thin. The strongest evidence we have for AI in day-to-day clinical work comes from outside allergy, from randomized trials of ambient documentation.
Five places AI shows up in an allergy practice today
1. The visit note
Documentation is where the time goes. In a time-and-motion study of 57 US physicians in four specialties, physicians spent 27% of their office day on direct clinical face time and 49.2% on electronic health record and desk work, plus 1 to 2 hours of after-hours work each night, mostly on the EHR.[6] Allergy was not one of the four specialties studied, but anyone who has charted a new-patient allergy visit with a full environmental, food and drug history will recognize the pattern.
Two randomized trials published in NEJM AI in late 2025 are the best evidence yet for ambient AI scribes:
- In a three-arm trial of 238 outpatient physicians across 14 specialties, one scribe reduced time-in-note by 9.5% against usual care and the other showed no significant change. Both groups reported improvements in burnout measures, and clinicians reported clinically significant inaccuracies "occasionally".[7]
- In a 24-week stepped-wedge trial with 66 practitioners, ambient AI reduced work exhaustion and interpersonal disengagement, and time spent on notes fell by 0.36 hours per day. Diagnostic billing codes improved, and the authors report no compromise in note quality.[8]
Two points matter for allergists. First, the benefit is real but moderate, and it varies by product. Second, "occasional inaccuracies" means every draft needs review. A tool that makes review fast is worth more than one that claims it doesn't need review. We go deeper in our guide to choosing an AI scribe for allergists.
2. Skin testing images
Skin prick and intradermal testing produce a panel of wheals that someone has to measure, in millimeters, against the controls, within a timed window. The 2008 AAAAI/ACAAI diagnostic testing parameter notes that reliability depends on the tester's skill, the device, skin color, the day's skin reactivity and age, among other things.[9] That variability is what computer vision research is trying to address.
The published work so far is promising and still early. In a 2025 Nature Communications study of an automated prick-test device with an AI-assisted readout, physicians adjusted 5.8% of the AI measurements in the test cohort, and readout variability and time fell; several authors work for the device maker.[10] A company-sponsored study of a 3D-laser reader for intradermal tests found only moderate-to-low agreement with manual measurement (intraclass correlation 0.40 to 0.65).[11] We review this research in detail in AI and allergy skin testing: what's possible today.
3. Patch testing
Patch testing adds a time dimension: readings at about 48 hours and again between days 3 and 7, graded on the ICDRG scale.[12] Photos help when patients can't make every visit, but a classic study found that experienced dermatologists grading patch-test photographs disagreed widely on ICDRG grades and agreed only after collapsing them to negative, positive and irritant.[13] A 2025 review of AI in patch testing found 10 qualifying studies, mostly convolutional neural networks, and flagged small samples, variable image capture and a lack of standardized reporting on skin types as the main limitations.[14]
4. Continuity across visits
Allergy is longitudinal. Sensitization results, reaction histories, immunotherapy courses and delabeled drug allergies all need to carry forward. The 2022 drug allergy parameter, for example, says that once a patient is delabeled, the updated status should be communicated across record platforms so the label doesn't come back.[15] This is less glamorous than image analysis, but it may be where AI saves the most clinical risk: putting what was already documented in front of the clinician at the right time.
5. The front office and coding
Phones, scheduling, refill requests and testing-day preparation (for example, reminding patients about antihistamine washout before skin testing) consume a large share of staff time. AI phone and chat agents can capture and route these requests. Coding support is the other operational use: the ambient AI trial above reported improved diagnostic billing codes.[8] We cover this in AI for allergy practice operations.
Why allergy is different from general medicine
A generic medical AI tool can draft a reasonable primary-care note. Allergy breaks it in specific ways:
- The test record has its own rules. The 2008 parameter lists what a complete skin-test record contains: who tested and when, the method, each allergen's common name and concentration, the controls with their results in millimeters, and wheal and erythema sizes in millimeters rather than 0 to 4+ grades.[9]
- Sensitization is not allergy. A positive skin test shows sensitization; it is not a diagnosis on its own. The NIAID food allergy guidelines say the skin prick test alone cannot be considered diagnostic of food allergy.[16] An AI that turns a positive test into a diagnosis is unsafe.
- Criteria differ by context. The 2008 parameter treats a prick response at least 3 mm larger than the negative control, with equivalent erythema, as evidence of specific IgE, while the 2022 drug parameter requires a wheal at least 3 mm larger than the negative control plus a flare of at least 5 mm.[9][15] Software has to apply the practice's own criteria, not a hard-coded rule.
- Skin tone matters. Histamine wheals are significantly larger in darkly pigmented skin, according to the 2008 parameter,[9] and the patch testing AI literature lacks standardized reporting on skin types.[14] Any image tool needs evidence across skin tones.
- The vocabulary is specialized. Extract names, cross-reactive allergen families, venom protocols, ICDRG grades and immunotherapy build-up schedules are easy for a general-purpose model to garble.
Risks and the guardrails to insist on
The risks named in the literature are consistent: inaccuracies that look plausible, over-reliance that erodes clinical reasoning, privacy and consent, opaque algorithms, and bias.[7][4][5] In my view, five guardrails answer most of them:
- A person confirms, a clinician signs. Nothing an AI drafts should enter the record as final without a human.
- Statements are checkable. The clinician should be able to see where a statement in the draft came from.
- No silent negatives. For testing, "no reaction" should never be recorded unless a person confirmed it.
- Honest status labels. Investigational features should be labeled as investigational, and image-analysis tools should state their regulatory status plainly.
- Contracted privacy. A signed BAA, a known hosting region and encryption at rest and in transit, before any PHI flows.
A checklist for evaluating any allergy AI tool
| Question | Why it matters in allergy |
|---|---|
| Was it built for allergy, or adapted from a general tool? | Allergy notes, test records and immunotherapy have specialized structure and vocabulary. |
| Who confirms and who signs? | The draft should never be final without a person; the clinician should sign every note. |
| Can I see the source of each statement? | Review is faster and safer when you can check a line against what was said. |
| Does it know the skin-test results when it drafts the note? | Re-typing test results into the note is where transcription errors happen. |
| What is the regulatory status of any image feature? | Image analysis of skin tests is an evolving area; ask directly. |
| Does it apply my practice's criteria? | Positivity criteria differ between aeroallergen, drug and venom testing. |
| Will you sign a BAA, and where does the data run? | PHI requires a contract and known infrastructure. |
| How is it priced, and can I try it without a long contract? | A short trial in your own clinic beats any demo. |
How Medora approaches AI in allergy
Medora: frontier AI built only for allergy
Medora is the AI copilot built only for allergy and immunology practices, built together with Allergy Affiliates, a working allergy practice in Bradenton, FL. All modules share one patient context.
- Medora Copilot listens to the allergy visit and drafts an allergy-structured SOAP note (HPI, ROS, assessment and plan, workup), with templates for new patients, follow-ups, skin testing, immunotherapy, asthma, and drug and food challenges. It suggests ICD-10 codes for the clinician to review. The clinician reviews, edits and signs every note.
- Evidence Mapping links lines of the note, including review of systems and past, family and social history, back to where they were said in the visit, so the clinician can check the source before signing.
- Skin testing with Medora Vision (investigational, not FDA-cleared): the nurse photographs the panel, Medora drafts wheal sizes using the printed ruler card for scale and flags sites it isn't sure of, a person confirms every value, and the clinician signs. Nothing is recorded as "no reaction" until a person confirms it.
- Patch testing is live as a workflow: ordering, 48-hour and Day 3–4 readings on the ICDRG scale, photos and clinician sign-off. Photo-assisted reading of patch tests is in development.
- Patient Intelligence shows a snapshot of the prior visit and a longitudinal allergen profile, with cross-reactivity alerts for the clinician to review. Everything it shows is advisory.
In a 60-day pilot at Allergy Affiliates, provider sign-off per skin-test visit went from 18 minutes to 4. Medora signs a BAA (through AdvanceAI LLC) and runs on HIPAA-eligible AWS. Read the full description of Medora.
What comes next
My view, as a builder rather than a clinician: the next two years in allergy AI will be decided less by model size and more by specialty data and workflow design. The AAAAI work group was right that specialists need to be involved in design and validation.[2] The useful tools will be the ones built inside allergy clinics, measured against what allergists actually do, and honest about what they don't do yet.
Frequently asked questions
What is AI in allergy and immunology?
It covers language models that draft and summarize clinical notes, computer vision that analyses skin test and patch test images, and predictive models that are mostly still in research. The AAAAI has published a framework calling for specialists to be involved in the design, validation and implementation of these tools.
Can AI diagnose allergies?
No tool should be treated that way. A positive skin test shows sensitization, not allergy, and practice guidelines say test results must be read together with the clinical history. Medora does not diagnose; the clinician decides and signs.
Is there evidence that AI scribes help clinicians?
Two randomized trials published in NEJM AI in 2025 found reduced documentation time or burden with ambient AI, alongside occasional inaccuracies that require clinician review. Neither trial was specific to allergy.
Does AI read skin prick tests?
Research devices with AI-assisted readouts have been published, mostly outside the US. In Medora, Vision is investigational and not FDA-cleared: it drafts wheal sizes for the nurse to confirm, and the clinician signs.
What should an allergist ask before adopting an AI tool?
Whether it was built for allergy, who confirms and signs, whether each statement can be traced to its source, the regulatory status of any image feature, whether it applies your practice's criteria, and whether the vendor signs a BAA.
Related guides
Sources
Last reviewed October 8, 2026. Summaries paraphrase the cited documents; read the originals before making clinical decisions.
- American Medical Association. Augmented intelligence in medicine. www.ama-assn.org
- Khoury P, Srinivasan R, Kakumanu S, et al. A framework for augmented intelligence in allergy and immunology practice and research: a work group report of the AAAAI Health Informatics, Technology, and Education Committee. J Allergy Clin Immunol Pract. 2022;10(5):1178–1188. doi.org
- Goktas P, Damadoglu E. Future of allergy and immunology: is artificial intelligence the key in the digital era? Ann Allergy Asthma Immunol. 2025;134(4):396–407.e2. doi.org
- González-Díaz SN, Morais-Almeida M, Ansotegui IJ, et al. Artificial intelligence in allergy practice: digital transformation and the future of clinical care. World Allergy Organ J. 2025;18(8):101078. doi.org
- Goktas P, Karakaya G, Kalyoncu AF, Damadoglu E. Artificial intelligence chatbots in allergy and immunology practice: where have we been and where are we going? J Allergy Clin Immunol Pract. 2023;11(9):2697–2700. doi.org
- Sinsky C, Colligan L, Li L, et al. Allocation of physician time in ambulatory practice: a time and motion study in 4 specialties. Ann Intern Med. 2016;165(11):753–760. doi.org
- Lukac PJ, Turner W, Vangala S, et al. Ambient AI scribes in clinical practice: a randomized trial. NEJM AI. 2025;2(12). ClinicalTrials.gov NCT06792890. doi.org
- Afshar M, Baumann MR, Resnik F, et al. A pragmatic randomized controlled trial of ambient artificial intelligence to improve health practitioner well-being. NEJM AI. 2025;2(12). ClinicalTrials.gov NCT06517082. doi.org
- Bernstein IL, Li JT, Bernstein DI, et al. Allergy diagnostic testing: an updated practice parameter. Ann Allergy Asthma Immunol. 2008;100(3 Suppl 3):S1–S148. www.aaaai.org
- Seys SF, Hox V, Chaker AM, et al. Artificial intelligence (AI)-assisted readout method for the evaluation of skin prick automated test results. Nat Commun. 2025;16:8637. (Several authors are employees or shareholders of the device maker.) doi.org
- Morales-Palacios MP, Núñez-Córdoba JM, Tejero E, et al. Evaluation of a novel automated allergy intradermal skin test reader: a diagnostic accuracy study. Clin Exp Allergy. 2024;54(12):1006–1009. (Company-sponsored.) doi.org
- Fonacier L, Bernstein DI, Pacheco K, et al. Contact dermatitis: a practice parameter—update 2015. J Allergy Clin Immunol Pract. 2015;3(3 Suppl):S1–S39. www.aaaai.org
- Ivens U, Serup J, O’goshi K. Allergy patch test reading from photographic images: disagreement on ICDRG grading but agreement on simplified tripartite reading. Skin Res Technol. 2007;13(1):110–113. pubmed.ncbi.nlm.nih.gov
- Tang HS, Ebriani J, Yan MJ, Wongvibulsin S, Farshchian M. Artificial intelligence in patch testing: comprehensive review of current applications and future prospects in dermatology. JMIR Dermatol. 2025;8:e67154. doi.org
- Khan DA, Banerji A, Blumenthal KG, et al. Drug allergy: a 2022 practice parameter update. J Allergy Clin Immunol. 2022;150(6):1333–1393. www.aaaai.org
- NIAID-Sponsored Expert Panel (Boyce JA, et al.). Guidelines for the diagnosis and management of food allergy in the United States. J Allergy Clin Immunol. 2010;126(6 Suppl):S1–S58. pmc.ncbi.nlm.nih.gov