# How Many People Should I Hire? An AI Medical Device Perspective. > Source: > Published: 2026-08-19 05:00:00+00:00 **Materials, methods and a caution.** AI agents did the research behind this article: they exported the FDA Device Explorer corpus, read the clearance letters, queried Explorium, captured the job postings, built the figures and drafted the text. Multiple independent panels of AI reviewers (eight Claude Opus 5 agents and eight Codex agents) then fact-checked every claim against primary sources, and the check-and-correct pass was repeated until the reports converged and no unsupported fact remained; a human spot-checked the corrections and the underlying FDA letters. Treat this as an evidence-based analysis, not a substitute for your own diligence: do independent research and verification before basing any business decision on it. That being said, it matches my intuition and lived experience for what I have seen from medical device manufacturer client team compositions. By our planning model, a first AI SaMD 510(k) is roughly **42 person-months of specialized work spread across eleven roles**. A De Novo is closer to 75. Companies cover that work by hiring the team, by borrowing it, or by taking years. The public record shows all three, and it shows what the team looks like once it exists. This article is for the founder or CFO trying to size the team a first submission needs. **The census.** Innolitics FDA Device Explorer indexes 989 AI/ML software-as-a-medical-device authorizations through August 2026 [1]. For 890 of them (decisions from 2018 onward whose letter names a classifiable correspondent) we extracted the addressee block of the FDA clearance letter, the person the agency wrote to [2], and classed it by title text with a reproducible script plus 22 checked overrides [3]. That person is either an in-house regulatory or quality title, an outside consultant or law firm, or an executive or engineer with no regulatory title. Another 37 readable letters carried no usable title and 49 De Novo and PMA records have no letter block; both are excluded, which under-represents De Novo first authorizations. We grouped applicants by how many AI SaMD authorizations they hold in total: one, two to four, or five or more. **The cohort.** Twenty-four AI SaMD companies: twelve serial submitters (Aidoc, RapidAI, [Viz.ai](http://viz.ai/), Overjet, Subtle Medical, [Qure.ai](http://qure.ai/), Bunkerhill, Brainomix, Annalise, Pearl, Lunit, Cleerly) and twelve single-authorization companies (Prenosis, Beacon Biosignals, Darmiyan, Bayesian Health, Clairity, Imidex, Fifth Eye, Rune Labs, Avenda, Optellum, Ibex, Etiometry). For each we pulled headcount and title-level staff counts from Explorium's people index [4], funding from Explorium and press, founding year, first authorization, and the correspondent on the first submission [2, 5]. **The labor market.** Twenty-two live job postings at medical device and AI SaMD companies, captured on 2026-08-17 and screenshotted because postings disappear [6], plus the RAPS 2024 compensation survey [7], and BLS wage and employer-cost data [8, 9]. Across all 890 letters, **57 percent** name an in-house regulatory or quality person, **27 percent** an outside consultant or law firm, and **16 percent** an executive or engineer. Split by experience, the picture changes. *Figure 1. Correspondent named on 890 FDA clearance letters for AI SaMD (2018 to 2026), by how many AI SaMD authorizations the applicant holds.* Among companies with a single AI SaMD authorization, only **42 percent** signed through an in-house regulatory title. **36 percent** used a consultant or law firm, and **22 percent** had a founder, CTO, COO or engineer as FDA's point of contact. Among serial submitters with five or more authorizations, **69 percent** signed in-house and the executive share fell to **11 percent**. The same gradient appears when you look only at each company's first authorization. Companies that stayed at one submission had a regulatory title on the letter 42 percent of the time; companies that went on to five or more had one 58 percent of the time. Across all 360 first authorizations, 56 percent named somebody other than an in-house regulatory person, and even among the eventual serial submitters 42 percent did. Aidoc's first 30 clearances went through Hogan Lovells before an SVP of Regulation appeared on the late-2025 letters [10]. Subtle Medical ran eight clearances through outside consultants (Decus Biomedical, then Enzyme) before an in-house Quality Engineering and Regulatory Affairs Manager signed in 2026 [11]. [Viz.ai](http://viz.ai/)'s De Novo, Cleerly, Bunkerhill and Pearl all started with outside counsel or consultants [12]. *Figure 2. Correspondent on each applicant's first AI SaMD authorization (n = 360), split by what the applicant went on to become.* The business reading: an in-house regulatory title tends to **appear on the letter after** the first clearance, not before. What varies is whether the interim solution is a law firm, a consultant, or a founder or CTO signing. Two cautions. A letter names one correspondent, not an org chart, so it cannot show when a company hired regulatory staff. And a law firm on the letter does not mean the law firm produced the evidence; somebody still ran the study, wrote the software documentation, and built the risk file. *Figure 3. Twenty-four AI SaMD companies: authorizations, years to first, indexed headcount, regulatory / quality / clinical titles, and who signed the first letter.* Serial submitters carry a median of **3 people with regulatory titles, 5 with quality titles, and 16 with clinical or medical titles**, on a median indexed headcount of 256. Single-authorization companies carry a median of **0 regulatory, 0 quality, and 5 clinical or medical**, on a median headcount of about 36 (titles are classed exclusively: regulatory first, then quality, then clinical). Six of the ten single-authorization companies with an index entry show no one with a regulatory title; Imidex and Fifth Eye are too small to be indexed at all. Their first FDA letters went to an in-house regulatory title in six cases (Beacon, Bayesian Health, Clairity, Optellum, Ibex, Etiometry), to an outside CRO or consultant in four (Prenosis via Proxima Clinical Research [13], Darmiyan via MCRA [14], Fifth Eye via Biologics Consulting, Rune Labs via a clinical CRO), and to a co-founder or chief operating officer in two (Avenda, Imidex) [5]. Clairity's regulatory signatory, an acting head of operations and regulatory, is the same person who signed Imidex's 510(k) as its COO. Two things stand out in the serial group. The quality column at Aidoc, RapidAI and Lunit is dominated by software test and QA engineers, the people who generate verification evidence, plus (at Aidoc) a layer of medical quality assurance staff who review AI output. The clinical column at Aidoc (37), Cleerly (36) and [Viz.ai](http://viz.ai/) (29) is mostly clinical data analysts, medical data interpreters and clinical research managers: the ground-truthing and validation machine. That is the production line for the evidence a submission needs. Small companies do not have that machine, so they substitute with lost time, consultants, or founder hours. Aidoc, the most prolific AI SaMD submitter in the FDA record, ran its first thirty clearances through an outside law firm, and its own regulatory officer began signing only in late 2025, by which time Explorium indexes it at roughly 500 employees [4]; the firm still appears on a 2026 letter [10]. The cohort's serial submitters reached their first AI SaMD authorization a median of **3.5 years** after founding. The single-authorization companies took a median of **7 years**. Three of the twelve took ten or more. Aidoc's CEO put the learning curve in one sentence in 2019: "It took us 18 months to get our first FDA clearance, 6 for the second one" [15]. *Figure 4. Years from founding to first AI SaMD authorization for the 24-company cohort.* Founding-to-clearance includes product development, so treat the gap as directional. FDA's receipt-to-decision clock is cleaner, though it includes time the submission sits on hold with the applicant. For Traditional 510(k)s of AI SaMD decided since 2018, first-and-only applicants waited a median of **195 days** from receipt to decision; applicants with five or more authorizations waited **146 days**. The interquartile range for first-timers runs 131 to 263 days [1]. FDA's own MDUFA reporting puts the average applicant hold after an Additional Information request at 55 to 68 days [16], so a gap of about 50 days is consistent with first-timers eating one more hold cycle than experienced applicants, which is what an incomplete package earns. The data do not identify the mechanism directly. *Figure 5. FDA receipt-to-decision time for Traditional 510(k)s of AI SaMD, by applicant experience.* Pooled across all letters, who signs moves the clock only a little (median 151 days for in-house, 155 for consultants, 166 for executives). Among first-and-only applicants the spread widens: 177 days when an in-house regulatory title signed, 198 for consultants, 201 for executives. These are unadjusted medians, not causal estimates, but they point the same way: the clock rewards experience, whether it sits inside the company or with the person who signs for it. The census says companies do not have the team at the start. The job market says what the team is. Between the FDA letters, the cohort's org charts and the postings we captured, the same roles keep appearing. *Figure 6. Eleven roles and their person-months for a first 510(k) and De Novo (Innolitics planning estimate).* Our planning estimate for a first Traditional 510(k) of an AI SaMD with a retrospective standalone performance study, no reader study, and a QMS that has to be built rather than inherited: | Role | What the submission needs from them | 510(k) person-months | De Novo person-months | |---|---|---|---| | Regulatory affairs lead | Pathway and predicate strategy, Pre-Sub, eSTAR authoring, AI responses | 4 to 9 | 8 to 18 | | Quality systems lead | ISO 13485 / QMSR build, design controls, ISO 14971 risk file, DHF | 4 to 10 | 6 to 14 | | Software lifecycle lead (IEC 62304) | SRS, SDS, architecture, SOUP, traceability, anomaly list | 4 to 9 | 6 to 12 | | Software V&V engineer | Verification protocols and reports, regression, traceability evidence | 3 to 8 | 5 to 12 | | Product cybersecurity engineer | Threat model, SBOM, architecture views, pen test, 524B package | 1 to 4 | 2 to 5 | | Clinical affairs / study manager | Protocol, data sourcing, IRB, sites, ground-truthing operations, report | 3 to 10 | 8 to 24 | | Biostatistician | Sample size, SAP, subgroup and confidence analyses | 1 to 3 | 3 to 8 | | Medical director | Intended use and clinical rationale, truthing supervision, labeling | 1 to 3 | 2 to 6 | | ML validation scientist | Locked-model testing, data documentation, model description, PCCP | 2 to 6 | 4 to 10 | | Human factors engineer | Use-related risk analysis, formative and summative usability | 0.5 to 2 | 1 to 4 | | Technical writer / doc control | Labeling, IFU, DHF assembly, eSTAR packaging | 1 to 3 | 2 to 5 | Total | 25 to 67 (typical 42) | 47 to 118 (typical 75) | FDA's own paperwork estimate for a 510(k) is 79 hours, plus 40 for eSTAR [17]. That figure covers filling in the form, not producing the evidence that goes in it, and it is usually the only number a first-time founder has seen. The 42 person-months is what the evidence costs. Three properties of this table drive everything else. The roles are **not interchangeable**: a biostatistician cannot write an IEC 62304 software description and a regulatory lead cannot run a reader study. Several are **fractional**: the biostatistician is a two-month job and the medical director a part-time one, so a full-time hire sits idle or the role goes to a contractor. And several are **sequential**: the SAP has to exist before the study, the study before the submission, the submission before the AI response. Sequencing is why headcount cannot be traded for time without limit; the mythical man month applies here too. The postings put numbers on the roles, and they describe the roles better than any job-family taxonomy. *Figure 7. Live regulatory leadership postings, captured 2026-08-17.* Overjet, ten AI SaMD clearances in, is hiring a Head of Regulatory who must "own the planning, coordination, and submission of regulatory filings, including FDA 510(k), De Novo," serve as "the primary point of contact with regulatory authorities," and "build and manage the regulatory team," with 7 or more years of experience including 2 with SaMD [6a]. Remedy Robotics posts a Head of Regulatory Affairs at **$220K to $280K** plus equity and asks for 10 or more years and "hands-on experience authoring IDE and PMA/De Novo submissions, not solely in a managerial capacity" [6b]. WHOOP posts a Staff Regulatory Affairs Associate for SaMD and AI at **$170K to $220K**, 8 or more years, PCCP literacy named as a competency [6c]. Eight Sleep bundles RA and QA into a single Director with "a successful track record of FDA clearance (510(k), De Novo)" [6d]. The RAPS 2024 survey puts average total compensation at $293K for VP and director level regulatory professionals and $168K for managers [7]. *Figure 8. Live quality, software V&V and application security postings, captured 2026-08-17.* HeartFlow, an AI SaMD company with a public listing, splits the evidence work into separate reqs: a Senior Device Quality Engineer at **$160K to $180K** who will "act as the lead for Software Verification and Validation," "manage the Design History File," and "support Regulatory Affairs in authoring technical files for global submissions, including FDA (510k/De Novo)"; a Senior Test Engineer at **$165K to $205K** who will "author clear, traceable test plans, test protocols, and test reports that meet the documentation expectations of a regulated medical device QMS," with direct SaMD testing experience required; and an Application Security Engineer at **$145K to $180K** [6e]. Baxter's Principal Engineer for Product Cybersecurity, **$120K to $165K**, will "create threat models of medical device systems" and "interface with regulatory bodies" [6f]. WHOOP's Quality Systems Specialist, **$105K to $135K**, owns "document control, post-market workflows, training management" and audit readiness under ISO 13485, 21 CFR 820/QMSR and IEC 62304 [6g]. *Figure 9. Live clinical, biostatistics and regulated-ML postings, captured 2026-08-17.* On the evidence side, Eight Sleep's Director of Clinical Affairs, **$200K to $250K**, will "own clinical study execution from protocol development through final study report" and support "pre-subs, 510(k) submissions." WHOOP's Director of Clinical Science, **$220K to $245K**, will "define clinical validation strategies, including study design, comparator selection, endpoint development, sample size justification, statistical analysis planning." Freenome's Staff Biostatistician, **$166K to $238K**, will "represent Biostatistics in interactions with regulatory authorities." WHOOP's Senior Manager for Machine Learning in SaMD, **$170K to $230K**, will "set the standard for how regulated ML gets built," including "validation rigor, and audit readiness" [6h]. The physician time is priced by the hour: BLS puts the median radiologist at $420,860 [8], physicians' stated minimum consulting rates have a $300 median with a $150 to $400 middle half [18], and one now-closed AI company posting ([Onc.AI](http://onc.ai/)) offered $400 an hour for radiologist annotation [19]. *Figure 10. Posted base salary ranges from live August 2026 postings, with RAPS 2024 total-compensation references.* Two patterns run through all 22 postings. Nearly every leadership role asks for seven to ten years and, in five of eight cases, a personal record of FDA authorizations or clearance-producing trials rather than exposure to them. And the split between narrow and generalist tracks company stage: HeartFlow and WHOOP post one req per competency, while startups collapse the same scope into a single "Head of" hire; Neko Health's first US regulatory hire, for example, must also personally hold the US Agent and Initial Importer roles and lists SaMD experience only as nice to have [6i]. Fully loaded, a **lean in-house team** that covers the eleven roles with real people is five to six full-time hires (regulatory, quality, software lifecycle, V&V, ML validation, clinical) plus two or three contractors for statistics, medical direction, cybersecurity and human factors. At $150K to $200K posted bases and the BLS employer-cost ratio (benefits are 30 percent of private-industry compensation, so a hire costs about 1.43 times base [9]), the salaried core alone is **$1.07M to $1.72M per year**, before $150K to $300K of contractor time, recruiting fees of 20 to 30 percent of first-year pay [20], ramp time, or the roughly 89 days that medtech recruiters report it takes to fill a specialised device role [21]. It is also permanent. After clearance most of it stays for post-market surveillance, complaint handling, change control and the next submission. The census classifies by the title text on one page of one letter with a reproducible script and 22 documented overrides; a hand audit of 60 rows found a residual error rate of a few percent, mostly executives who were in fact outside consultants, and no correction moved a headline share by more than three points. Some in-house people use "consultant" titles and some consultants sign under client addresses, and most De Novo letters are not in the index. Innolitics itself is the correspondent on 9 of the 240 outsourced letters. Explorium headcounts undercount small and non-US firms and count titles, not effort. Person-month estimates are Innolitics planning figures calibrated to our own project data [22, 23], not survey data, and they assume a retrospective validation study; a prospective or MRMC study moves the clinical rows substantially. Founding-to-clearance mixes product development with regulatory time, and founding years for Imidex and Bayesian Health differ across public sources (we used 2009 and 2018) [24, 25]; Annalise.ai's is taken from Harrison.ai's 2019 formation announcement [26]. The FDA-days figures come from a May 2026 snapshot of the universe. Posted salary ranges skew to states that require them. The in-house cost figure is a planning scenario built from posted salary ranges, not a survey of actual payrolls. If you are staffing a first AI SaMD submission, count roles before you count dollars. Write down the eleven, mark which ones you have on payroll today with prior submissions behind them, and price the gap: five to six full-time hires and two or three contractors, roughly $1.07M to $1.72M a year in loaded salaries before recruiting and ramp, and 12 to 18 months for a first 510(k) if every seat is filled by someone who has done the job before. Whatever mix of hires and outside help closes that gap, make sure the DHF and the QMS end up owned inside the company, because the serial submitters in the cohort all carry the function now, and seven of the twelve started with a signature that was not an in-house regulatory title. `contains_ai_ml=true` , `is_samd=true` ), retrieved 2026-08-17. `https://fda.innolitics.com/device/{record}.md` ) and hosted by FDA at `cdrh_docs/pdfYY/{record}.pdf` ); De Novo contacts from the FDA De Novo database (`accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/denovo.cfm` ).`classify_letters.py` , `letter_overrides.csv` and `letters_classified.csv` (890 letters), repository `research/ai-samd-team-headcount-2026-08/` , 2026-08-17.`first_correspondents.csv` (24 cohort first-letter addressees with the letter each was read from), 2026-08-17.