AI-Native Knowledge Engineering: Demand, Market Formation, and the Race to Capture Tacit Expertise
Executive summary
The evidence supports a qualified but fairly strong “yes”: demand is emerging for the function described as AI-native knowledge engineering, but the occupation has not yet converged on the title “knowledge engineer.” The working definition in the research brief is unusually useful because it searches by activity: observe experts doing real work, identify consequential decision points, elicit cues and judgment, combine those explanations with operational evidence, structure the result for machines, and keep updating it from later work.
That is also the opportunity that emerged in the earlier Grok discussion: not merely documenting procedures, but capturing the judgment of practitioners before it disappears and making it part of a continuing human-machine learning loop.
As of October 2026, that function is being split across at least four labor-market families rather than advertised as one profession:
Structured knowledge
Typical titles include Knowledge Engineer, Knowledge Graph Engineer, Ontology Engineer, Knowledge Graph Architect, Scientific Knowledge Engineer.
Turns documents, records and heterogeneous data into structured, provenance-aware knowledge for retrieval and reasoning.
Match to the target role: Medium.
Embedded applied AI
Typical titles include Forward Deployed AI Engineer, Forward Deployed Software Engineer, Forward Deployed Data Scientist, Applied AI Engineer, AI Solutions/Architect roles.
Works directly with users, discovers workflows, builds AI around real operating problems, iterates from field feedback.
Match to the target role: High, but usually does not explicitly elicit tacit expertise.
Human judgment/data
Typical titles include AI Trainer, Domain Expert, Expert Contributor.
Creates hard questions, evaluates model outputs, supplies demonstrations and expert feedback.
Match to the target role: High for judgment; low for embodied work.
Frontline knowledge capture
Typical titles include Product/deployment roles at connected-worker, smart-glasses and industrial-AI companies.
Captures work through cameras, microphones, sensors, remote-expert sessions and operator activity.
Match to the target role: Highest technological match, but not yet a standardized occupation.
This fragmentation matters. Current postings titled Knowledge Engineer are still dominated by knowledge graphs, ontologies, RAG, document ingestion, entity resolution and provenance. Sapience AI's current role, for example, pays 204,000–216,000 plus equity, but is primarily about turning scattered sources into a trustworthy knowledge graph. LTS explicitly includes “Knowledge Engineer” within a Senior Applied AI Engineer role, but defines it around document processing, semantic search, embeddings and enterprise knowledge management.
Cyvl's Knowledge Graph Architect combines documents with real-world LiDAR, imagery and GPS data for autonomous agent use, which moves closer to physical-world intelligence but still does not employ the “sit beside the master and ask why” model.
The activity closest to the proposed occupation is often called forward-deployed AI engineering. Palantir currently advertises Forward Deployed AI Engineers at 135,000–200,000 plus potential equity and incentives, working directly with customers, implementing end-to-end GenAI workflows, iterating with users and feeding field learning back into its AI platform. 01Health wants its equivalent embedded directly with clinicians, conducting hands-on user research and translating real-world clinical needs into deployed AI.
Govini's current defense-oriented hiring is revealing because its 53 listed openings simultaneously include a Knowledge Engineer, Lead Data Scientist–Knowledge Retrieval, Senior AI Engineers for Agentic Interactions and Agent Workflows, and Forward Deployed Data Scientists and Software Engineers. Those functions are beginning to cluster organizationally even before their titles converge.
The strongest new evidence, however, is not in job descriptions. It is in products. In 2025–2026 a small but recognizable category appeared around exactly the thesis in question. Myto says its AI glasses capture troubleshooting steps, shift handoffs and expert behavior while operators continue working, structure that knowledge, and use it to power manufacturing agents; the system then strengthens from subsequent frontline activity.
Mission Control AI launched a vest-mounted device called Bob in June 2026 that records audio, images and spatial position as experts work, sending the result into an on-premises agentic system intended to make physical know-how searchable and usable by software agents, robotics and manufacturing systems. AutoAlign's Visor spans smart glasses, mobile devices and sensors and explicitly markets the preservation of “tribal knowledge” before experts retire.
That makes the central market finding:
AI-native tacit-knowledge capture is no longer only a thesis. It is an early product category. But the human occupation that operates the category is less mature than the technology.
No incumbent I found has convincingly demonstrated at large scale the complete loop of observing work, detecting a meaningful decision, asking the expert why, combining the explanation with physical and sensor context, structuring judgment, validating it, deploying it to workers and agents, capturing outcomes, and updating the corpus.
The closest products assemble large portions of it; the “agentic elicitation” layer—an AI capable of recognizing that an expert just made a subtle judgment and asking the right question at exactly that moment—is still more visible in research than in mature industrial products. DARPA's Perceptually-enabled Task Guidance program already combines head-mounted cameras and microphones, environmental perception and mixed-initiative dialogue, while newer research is exploring LLM-driven contextual interviewing.
The demographic premise is also real, but should be targeted more carefully than “the trades are aging.” In 2025 BLS data, 54.4% of farmers/ranchers/agricultural managers were 55 or older; about 36.4% of precision instrument/equipment repairers, 31.4% of machinists, 27.7% of industrial machinery mechanics, 27.9% of aircraft mechanics, 25.1% of physicians, and 23.9% of insurance underwriters were 55+. Electricians, by contrast, had a median age of only 39.6 and about 18% aged 55+, so “electricians are all aging out” is too crude.
My assessment is that the strongest initial opportunities are therefore not necessarily the largest occupations. They are occupations or micro-industries at the intersection of:
high tacit-judgment density , combined with demographic loss , combined with expensive errors/downtime , combined with repeatable situations , combined with instrumentable work , combined with economically valuable reuse.
Industrial maintenance, precision manufacturing, specialized repair, aviation MRO, nuclear/utilities, agricultural equipment and infrastructure inspection fit especially well. Healthcare and insurance contain enormous judgment value but have materially higher regulatory, liability and data-governance barriers. BLS and DOE data reinforce the workforce-loss side: DOE reports that nuclear has 23% fewer workers under 30 than the wider energy workforce, expects significant retirements over the next decade, and reports widespread difficulty finding qualified workers.
A second strong result is that “datasets of judgment” is a useful conceptual label even though it is not yet standard industry terminology. A September 2026 research preprint, KUPAS MASTER, comes remarkably close to formalizing exactly that asset: expert experience represented as context, cues, judgment, action, boundaries and outcomes. Its authors call the broader process “experience engineering” and report converting 1,576 source files from 20 practitioners into 23,024 individual experience records and 13,113 organizational assets.
Their internal evaluation reports an agent score of 89.58 using structured experience assets versus 79.75 for raw-corpus RAG and 70.63 for the base model. That is encouraging but should be treated as early, author-reported preprint evidence, not independent validation.
Finally, the historical caution is substantial. Knowledge engineers in the 1980s already discovered that knowledge acquisition—not storage—was the bottleneck. Ethnographic work on expert systems found that turning expertise into formal representations is an interpretive act requiring extended interaction rather than simple “extraction.” Modern AI dramatically changes the cost structure of recording, transcription, searching, multimodal interpretation and structuring, but it has not repealed the underlying epistemic problem: experts often cannot fully state what they notice or why they know it.
So the market is better described as early formation rather than established occupation. The technology curve is moving faster than the job-title curve.
What the market is actually calling the work
A useful definition emerging from the evidence is:
AI-native knowledge engineering is the field-deployed practice of observing expert work in context; eliciting its latent cues, judgments, exceptions and constraints; combining those explanations with multimodal operational evidence and outcomes; structuring the result with provenance into machine-usable representations; and continuously improving those representations through subsequent human and agent use.
That definition distinguishes the emerging function from three older ones.
Traditional knowledge management asks, “What documents and organizational knowledge do we possess, and how do people find them?”
Classical knowledge engineering asks, “How do we represent domain concepts, rules, relationships and facts so software can reason over them?”
The emerging function asks a different question:
“What does the expert perceive and decide during real work that has never been adequately represented anywhere—and how do we make that judgment reusable?”
That distinction shows up clearly in live postings.
Representative current postings
Palantir
Current title: Forward Deployed AI Engineer.
The role involves: Work directly with customers; own GenAI strategy and implementation; build production LLM workflows; iterate with users; feed field learning back into AIP.
Compensation shown: 135k–200k/year, plus potential equity/bonus.
Palantir posting: https://jobs.lever.co/palantir/636fc05c-d348-4a06-be51-597cb9e07488
Sapience AI
Current title: AI/ML Data Knowledge Graph Engineer.
The role involves: Extract from messy community sources; design ontology/graph; resolve entities; preserve provenance; serve structured expertise to neuro-symbolic AI.
Compensation shown: 204k–216k + equity.
Sapience posting: https://job-boards.greenhouse.io/sapienceaicorporation/jobs/4409177009
Cyvl
Current title: Senior Knowledge Graph Architect.
The role involves: Combine LiDAR, imagery, GPS, documents, records and regulations into an agent-queryable infrastructure graph.
Compensation shown: 180k–250k base; company says 250k–350k total comp.
Cyvl posting: https://jobs.ashbyhq.com/cyvl/ac147164-3aba-43fc-bffb-2e49f86d3743
LTS
Current title: Senior Applied AI Engineer.
The role involves: Agents, RAG, evaluation plus an explicit “Knowledge Engineer” function for knowledge ingestion, semantic search and AI reasoning over enterprise knowledge.
Compensation shown: 144k–184k.
LTS posting: https://job-boards.greenhouse.io/lts/jobs/4340498009
01Health
Current title: Forward Deployed AI Engineer.
The role involves: Embed with clinicians/customers; conduct hands-on user research; improve clinical decision making; ship AI directly from real workflow problems.
Compensation shown: Not disclosed.
01Health posting: https://jobs.lever.co/32Co/7db2b314-428e-43eb-9a69-68589a1249c4
Govini
Current title: Knowledge Engineer plus FDE/agent roles.
The role involves: Current defense-tech roster simultaneously includes Knowledge Engineer, Knowledge Retrieval, Agentic Interactions, Agent Workflows, FDE Data Scientist and FDE Software Engineer.
Compensation shown: Varies.
Govini careers: https://job-boards.greenhouse.io/govini
Handshake AI
Current title: Math Expert / AI Trainer.
The role involves: Use PhD-level domain judgment to design hard questions and evaluate model reasoning; no prior AI experience required.
Compensation shown: 85–110/hour.
Handshake posting: https://jobs.ashbyhq.com/handshake/30c0f597-e3b8-4805-a9be-ab81958308ba/
Xebia
Current title: Scientific Knowledge Engineer, Ontology & Data Modeling.
The role involves: Formal knowledge representation inside AI/data consulting.
Compensation shown: Not shown in search result.
Xebia posting: https://job-boards.greenhouse.io/xebiaspain/jobs/5855104004
The compensation is notable. The technical portions of this emerging labor market already support senior-engineering compensation, not traditional documentation-specialist compensation: 135,000–200,000 at Palantir, 204,000–216,000 at Sapience, and 180,000–250,000 base at Cyvl. That does not establish a salary benchmark for the hypothetical tacit-knowledge engineer; it shows that employers already place high economic value on adjacent combinations of AI, knowledge representation and field/customer integration.
The alternative titles actually encountered in current hiring and research include Knowledge Engineer; AI/ML Data Knowledge Graph Engineer; Knowledge Graph Architect; Scientific Knowledge Engineer; Forward Deployed AI Engineer; Forward Deployed Software Engineer; Forward Deployed Data Scientist; Senior Applied AI Engineer; AI Trainer; Domain Expert; Lead Data Scientist–Knowledge Retrieval; Senior AI Engineer–Agentic Interactions; Senior AI Engineer–Agent Workflows; and Deployment Strategist.
Research is beginning to introduce additional vocabulary such as “experience engineering,” “AI Expert Twin,” and “Cognitive Operations Manager,” although those latter names are research concepts, not yet established labor-market titles.
This leads to an important hiring-market conclusion: a search for “Knowledge Engineer” materially understates demand for the function, while also over-counting jobs that are not the function of interest. The title is simultaneously too narrow and too broad.
There is not yet a credible longitudinal employment series for this exact occupation. BLS does not isolate it, and simple LinkedIn/Indeed headline counts would mix knowledge graphs, traditional KM, AI training, consulting and forward-deployed engineering while being vulnerable to duplicates and short-lived postings. I therefore would not claim a defensible year-over-year growth percentage for “AI-native knowledge engineers.” The better current evidence is the simultaneous appearance of adjacent roles, specialist products, funding, pilots and government programs.
Demand signals beyond job titles
The strongest evidence that something is changing comes from combining hiring with product formation, capital and government activity.
The 2025–2026 sequence looks like a category beginning to cohere rather than a mature category scaling linearly.
Formation of the AI-native tacit-knowledge stack
In 2025, AutoAlign launched Visor for industrial frontline AI, Augmentir expanded its industrial AI-agent capabilities, Mentra raised an $8 million seed round, and Meta acquired the always-on capture startup Limitless.
In 2026, Mentra began shipping camera-enabled open smart glasses, Mission Control launched Bob for expert-knowledge capture, and AI-first factory knowledge systems such as Myto emerged. Meta opened AI-glasses sensors to developers, while research formalized “experience engineering” and agent-ready expert corpora.
Those milestones are documented in company releases, current product pages, Reuters reporting and current research.
Capital is flowing into adjacent pieces of the stack. Mentra reports an $8 million seed round in July 2025 from investors including Y Combinator, Amazon and Toyota Ventures and began shipping its first batch of camera-equipped smart glasses in early 2026. Meta's December 2025 acquisition of Limitless is strategically interesting: Limitless had raised more than $33 million for an AI wearable that recorded, transcribed and made real-world conversations searchable; Meta said the acquisition would support its next-generation AI-wearables work.
There is also a broader capital signal around paying experts to create high-value training and evaluation data. Handshake currently offers 85–110 per hour to PhD mathematicians to create and judge difficult model problems. Reuters reported in September 2026 that Snorkel AI raised $350 million at a $3.5 billion valuation amid demand for specialized AI training data and uses expert contributors in fields including coding, law and medicine.
This is not embodied tacit-knowledge capture, but economically it validates an adjacent proposition: expert judgment itself is becoming an explicit input market for AI systems.
Acquisitions are beginning to move toward persistent organizational capture as well. In June 2026, Decidr announced an agreement to acquire Rumi.ai specifically to add an “always-on” capture layer spanning meetings, in-person conversations, Slack and CRM data and feed those signals into proprietary reasoning graphs. The announcement explicitly frames the valuable asset as how an organization's best people frame problems and weigh tradeoffs.
Because this is a buyer's own press release, its strategic language should not be treated as neutral market analysis—but the acquisition direction is directly relevant.
Enterprise deployment evidence is strongest in the older “connected worker” category. TeamViewer Assist AR already supports live expert guidance, session recording, AI-generated service-call summaries and reuse of previous session insights; TeamViewer names industrial customers including Mitsubishi Electric, Hurco and Schuler and publishes customer cases such as BOBST. Augmentir has a named Colgate-Palmolive customer and combines connected workflows, contextual training, image verification, industrial agents, remote expertise and continuous work-execution data. Its current site says its underlying platform “closes the loop between training and work execution.”
This matters because the newest entrants do not need to create enterprise acceptance of digital frontline work from zero. Remote expert assistance, AR instructions and connected-worker systems already established a commercial beachhead. The 2025–2026 change is that multimodal AI increasingly turns the captured stream into an active reasoning asset rather than merely a video call, recording or digital checklist.
Government and defense provide unusually clear evidence that the underlying need is taken seriously. DARPA's Perceptually-enabled Task Guidance program uses head-mounted cameras and microphones so AI assistants can perceive what mechanics or medics perceive, combine that with manuals and training materials, and engage in mixed-initiative dialogue during complex physical tasks. DARPA's In the Moment program attacks an even more subtle problem: characterizing how particular humans make high-stakes decisions under ambiguity so AI behavior can be aligned with those decision styles.
The U.S. Army has also experimented directly with Meta's Aria glasses for maintenance training: an Army program described capturing maintenance work, hand motions and environmental conditions through wearable devices and combining those observations with technical-manual information, later extending the effort into remote maintenance support. That is very close to a government-sponsored physical-work knowledge-capture experiment.
The federal demand signal also appears in hiring. LTS's current applied-AI role is building an AI-native platform around decades of mission-critical Veterans Affairs software and explicitly includes knowledge engineering, while Govini's defense-oriented hiring places knowledge engineering, knowledge retrieval, agent engineering and forward-deployed roles in the same product organization.
What I do not find yet is a large, separately reported “knowledge engineering services” market built around deploying teams to retiree cohorts and systematically harvesting tacit judgment. Today's evidence is mostly indirect: forward-deployed engineering, connected-worker implementation, AI-data marketplaces and early tacit-capture startups. That absence is itself informative. The function is forming inside adjacent categories before becoming a category of services in its own right.
Competitive landscape and the emerging capture stack
The competitive landscape is already recognizable, but fragmented into layers.
Myto
What is captured: Factory records plus AI-glasses recordings of troubleshooting, handoffs and expert behavior.
How AI uses it: Structures informal knowledge; builds troubleshooting/documentation/automation agents specific to the factory.
Feedback and reuse: Explicitly says system strengthens from frontline activity and physical work.
Deployment status: Commercial startup; YC and General Catalyst backing stated; no named customer found.
Evidence and thesis fit: Very high fit to full thesis.
Mission Control AI – Bob/Swarm
What is captured: Vest-mounted audio, imagery and precise spatial position while experts work.
How AI uses it: Converts physical knowledge into searchable spatial records; intended for synthetic workers, customer AI, robotics and manufacturing systems.
Feedback and reuse: Captured records become reusable agent inputs.
Deployment status: Product launched June 2026; 90-day pilots offered in energy, advanced manufacturing and logistics.
Evidence and thesis fit: Very high fit, but early-stage pilot evidence rather than proven scale.
AutoAlign Visor
What is captured: Smart glasses/mobile/sensors, industrial procedures, service activity, documentation.
How AI uses it: Contextual guidance, defect/hazard detection, workflow support.
Feedback and reuse: Company explicitly markets learning from experts and subsequent work.
Deployment status: Commercial product; company reports real-world manufacturing/service pilots.
Evidence and thesis fit: High fit; public customer evidence for Visor itself is thinner than product claims.
Augmentir
What is captured: Digital workflows, worker execution data, images/video, maintenance data, expert interactions.
How AI uses it: Industrial agents, procedure generation, contextual assistance, image verification, skills intelligence.
Feedback and reuse: Platform says training and execution form a closed loop.
Deployment status: Deployed enterprise platform, named Colgate-Palmolive reference.
Evidence and thesis fit: High fit as mature connected-worker foundation; tacit elicitation is not its sole focus.
TeamViewer Assist AR / Frontline
What is captured: Live expert sessions, audio/video, annotations, session history.
How AI uses it: AI summaries and suggestions from previous session insights; remote troubleshooting.
Feedback and reuse: Session knowledge can be reused by support teams.
Deployment status: Mature deployed product, named industrial customers and case studies.
Evidence and thesis fit: Strong enterprise precursor; more remote-expert capture than autonomous elicitation.
Strivr
What is captured: Images/videos of real procedures plus SOPs/manuals.
How AI uses it: Trains a custom visual-language model to recognize correct execution and detect deviations.
Feedback and reuse: Later worker performance is visually checked against learned patterns.
Deployment status: Commercial product.
Evidence and thesis fit: Strong “learn from expert demonstration” model; weaker on why the expert acted.
Mentra
What is captured: Point-of-view camera, microphones, narrated work and custom apps.
How AI uses it: Enables technician copilots, automatic documentation and custom AI workflows.
Feedback and reuse: Work can automatically become a record and feed guidance back to users.
Deployment status: Hardware shipping; $8M seed reported; open-source OS/SDK.
Evidence and thesis fit: Important platform layer, not itself the complete knowledge-engineering system.
DARPA PTG
What is captured: Head-mounted camera/mic plus task sources and environmental perception.
How AI uses it: Mixed-initiative AI dialogue and just-in-time task guidance.
Feedback and reuse: Designed to adjust assistance to user expertise and task state.
Deployment status: Research/government program.
Evidence and thesis fit: Demonstrates technical feasibility of major pieces; not a commercial platform.
Several other early vendors sit around this cluster—industrial smart-glasses, remote-expert and AI-assistance products from RealWear, AMA XpertEye, Carbyn, EON and smaller integration companies—but the profiles above capture the main architectural patterns visible in this research. RealWear is particularly relevant as an industrial hardware counterpoint to consumer smart glasses because it builds purpose-designed hands-free systems for harsh industrial environments.
The stack itself is becoming surprisingly complete:
Ambient capture
Purpose: Record first-person vision, speech, environmental state, location.
Current examples: Meta AI glasses, Mentra, RealWear, Mission Control Bob.
Maturity: High enough for deployment.
Biggest unresolved problem: Privacy, battery, PPE/safety requirements, consent.
Operational sensing
Purpose: Fuse machine telemetry, LiDAR, GPS, system logs, CMMS/MES/ERP.
Current examples: Cyvl, AutoAlign, Myto.
Maturity: Moderate-high.
Biggest unresolved problem: Time-aligning human decisions with machine state.
Multimodal perception
Purpose: Understand objects, actions, deviations and task context.
Current examples: Strivr custom VLMs; current multimodal model stack.
Maturity: Rapidly improving.
Biggest unresolved problem: Reliability on unusual edge cases.
Agentic elicitation
Purpose: Notice unexplained expert decisions and ask useful “why/what did you notice?” questions.
Current examples: DARPA mixed-initiative work; LLM interview research.
Maturity: Early.
Biggest unresolved problem: Knowing when to interrupt and what question exposes the hidden judgment.
Knowledge structuring
Purpose: Convert streams into cases, rules, constraints, graphs, embeddings and provenance.
Current examples: Sapience, Cyvl, KUPAS MASTER, graph/RAG tooling.
Maturity: High for explicit data; emerging for experience.
Biggest unresolved problem: Preserving context and exceptions instead of flattening them into generic rules.
Validation
Purpose: Let experts review/correct extracted knowledge and track disagreement.
Current examples: KUPAS cross-review/provenance; enterprise evaluation systems.
Maturity: Moderate.
Biggest unresolved problem: Avoiding false certainty and expert hindsight rationalization.
Delivery
Purpose: Put guidance back into the workflow through voice, AR, agents, software.
Current examples: TeamViewer, Augmentir, AutoAlign, Mentra, RealWear.
Maturity: High.
Biggest unresolved problem: Delivering enough help without deskilling or distracting workers.
Outcome feedback
Purpose: Link recommendation/action to later success, failure, correction and new conditions.
Current examples: Myto and Augmentir explicitly describe closed loops.
Maturity: Emerging.
Biggest unresolved problem: Reliable causal attribution and stale-knowledge detection.
The missing product is therefore not “smart glasses.” Nor is it “a knowledge graph.” The architecture increasingly looks like this:
The capture and learning loop
The process starts with an expert performing real work. Wearables, audio and machine telemetry capture that work. A multimodal model detects an event or decision, and an agent recognizes judgment that has not yet been explained.
The agent asks a contextual question, such as “Why this?” or “What did you notice?” The answer is combined with the situation, cues, judgment, action and boundary conditions.
Expert validation and provenance turn that material into a machine-usable experience corpus. A worker assistant, AI agent or training system then draws on the corpus.
Later outcomes, corrections, exceptions and new cases feed back into the experience corpus, updating what future workers and systems can use.
Myto, Mission Control and AutoAlign are closest to assembling the overall loop commercially; Augmentir and TeamViewer bring stronger installed-enterprise lineage; KUPAS MASTER supplies an unusually explicit model of the structured “experience corpus”; DARPA supplies evidence around wearable perception and mixed-initiative dialogue. But I found no clear market leader that publicly demonstrates the entire loop above at industrial scale. That is an inference from the surveyed products, not proof that no private deployment exists.
One 2026 development materially lowers the barrier to experimentation: Meta's current Wearables Device Access Toolkit gives developers access to camera streams, microphone/audio and other functions across its AI-glasses portfolio. However, Meta's current documentation also says publishing is still unavailable during the developer-preview phase. In other words, consumer-grade capture hardware is becoming programmable before the ecosystem is fully mature.
Mentra provides a more open alternative: its current product explicitly targets field teams and developers, supports an open-source smart-glasses OS and SDK, and markets the workflow in which technicians narrate repairs while the glasses record the job so documentation is generated from the work itself. It raised $8 million in 2025 and began shipping hardware in 2026.
That is one of the biggest differences from historical knowledge engineering: the capture interface can now disappear into the work. The expert does not necessarily need to finish a shift and spend an hour filling out KM forms.
Where the demographic opportunity is strongest
Aging is a real filter, but the data argue for being much more surgical than simply targeting “skilled trades.”
BLS's 2025 detailed-age table shows very different retirement profiles across occupations. Farmers/ranchers/agricultural managers are a genuine demographic outlier: of approximately 810,000 workers, 180,000 were 55–64 and 261,000 were 65+, yielding roughly 54.4% aged 55 or older and a median age of 56.1. Machinists were about 31.4% 55+; industrial and refractory machinery mechanics about 27.7%; precision instrument/equipment repairers about 36.4%; and aircraft mechanics about 27.9%.
Some intuitive targets are less demographically extreme. Electricians' median age was 39.6, with roughly 18% aged 55+, and HVAC mechanics/installers had a median age of 39.9. Those are large and potentially attractive knowledge-capture markets, but the argument should be “high tacit expertise and shortage/complexity,” not “nearly everyone is about to retire.” Construction and building inspectors look more exposed, with a median age of 46.1.
Professional sectors also contain substantial aging cohorts. BLS reports insurance underwriters at a median age of 44.5 with about 24% aged 55+, while its “other physicians” category had a median age of 45.6 and about 25% aged 55+. These domains contain substantial judgment, but the operational and regulatory context is very different from a machine shop.
The following score is deliberately a researcher-assigned opportunity heuristic, not a measured market-size estimate. Each dimension is scored 1–5. “Low friction” rewards easier capture/deployment, so 5 means comparatively fewer regulatory or workflow obstacles.
Industrial maintenance / MRO
Directional opportunity score: 24 out of 25.
Demographic and knowledge-loss risk scores 5 out of 5. Tacit and embodied judgment scores 5 out of 5. Cost of mistakes or downtime scores 5 out of 5.
Capture feasibility scores 5 out of 5. Low regulatory friction scores 4 out of 5, with a higher score meaning fewer obstacles.
Precision machining / tool & die / specialized manufacturing
Directional opportunity score: 23 out of 25.
Demographic and knowledge-loss risk scores 5 out of 5. Tacit and embodied judgment scores 5 out of 5. Cost of mistakes or downtime scores 4 out of 5.
Capture feasibility scores 5 out of 5. Low regulatory friction scores 4 out of 5, with a higher score meaning fewer obstacles.
Specialty field repair / equipment service
Directional opportunity score: 23 out of 25.
Demographic and knowledge-loss risk scores 4 out of 5. Tacit and embodied judgment scores 5 out of 5. Cost of mistakes or downtime scores 4 out of 5.
Capture feasibility scores 5 out of 5. Low regulatory friction scores 5 out of 5, with a higher score meaning fewer obstacles.
Agriculture / agricultural equipment
Directional opportunity score: 22 out of 25.
Demographic and knowledge-loss risk scores 5 out of 5. Tacit and embodied judgment scores 5 out of 5. Cost of mistakes or downtime scores 4 out of 5.
Capture feasibility scores 4 out of 5. Low regulatory friction scores 4 out of 5, with a higher score meaning fewer obstacles.
Aviation maintenance / MRO
Directional opportunity score: 20 out of 25.
Demographic and knowledge-loss risk scores 4 out of 5. Tacit and embodied judgment scores 5 out of 5. Cost of mistakes or downtime scores 5 out of 5.
Capture feasibility scores 4 out of 5. Low regulatory friction scores 2 out of 5, with a higher score meaning fewer obstacles.
Nuclear / utilities / critical infrastructure
Directional opportunity score: 20 out of 25.
Demographic and knowledge-loss risk scores 5 out of 5. Tacit and embodied judgment scores 5 out of 5. Cost of mistakes or downtime scores 5 out of 5.
Capture feasibility scores 4 out of 5. Low regulatory friction scores 1 out of 5, with a higher score meaning fewer obstacles.
Insurance underwriting / complex claims
Directional opportunity score: 20 out of 25.
Demographic and knowledge-loss risk scores 4 out of 5. Tacit and embodied judgment scores 4 out of 5. Cost of mistakes or downtime scores 4 out of 5.
Capture feasibility scores 5 out of 5. Low regulatory friction scores 3 out of 5, with a higher score meaning fewer obstacles.
Procedural healthcare / diagnostic specialties
Directional opportunity score: 18 out of 25.
Demographic and knowledge-loss risk scores 4 out of 5. Tacit and embodied judgment scores 5 out of 5. Cost of mistakes or downtime scores 5 out of 5.
Capture feasibility scores 3 out of 5. Low regulatory friction scores 1 out of 5, with a higher score meaning fewer obstacles.
The industrial-maintenance case is particularly strong. BLS puts industrial machinery mechanics at a median age of 45.5; 91,000 of 329,000 were 55+. First-line supervisors of mechanics/installers/repairers were even older, with a median age of 47.6 and about 92,000 of 285,000 aged 55+. Those workers also operate in environments that generate machine telemetry, work orders, parts records and visual/audio signals—meaning the expert's decision can potentially be linked to rich external evidence.
Precision repair is potentially more attractive than its small labor count suggests. BLS counted only about 55,000 precision instrument and equipment repairers, but roughly 20,000 were 55+, including 12,000 already 65+. Small occupations like this are precisely where a “capture it before it disappears” strategy can be economically interesting if the equipment being maintained is valuable enough.
Agriculture deserves more attention than it usually receives in AI-agent discussions. Its demographic exposure is extreme, but the relevant opportunity is probably not generic farming advice. It is highly contextual operating intelligence around particular crops, soils, equipment, irrigation systems, livestock, disease patterns, weather regimes and farm-specific history. BLS's 54% 55+ figure for agricultural managers means the potential loss of accumulated operating memory is unusually concentrated.
Nuclear and utilities have strong institutional incentives to preserve expertise. DOE reports an aging nuclear workforce, 23% fewer workers under 30 than the broader energy workforce, hiring difficulty across nuclear segments, and federal investment in new safety-training programs. High consequences of error, unusual equipment, long asset lifetimes and regulatory requirements all increase the value of provenance-rich knowledge—but also make deployment and validation much harder.
Aviation MRO is similarly compelling. BLS counted roughly 165,000 aircraft mechanics/service technicians in 2025, with 46,000 aged 55+. The Federal Aviation Administration has simultaneously been funding aviation-workforce development, including technician pipelines, reflecting a broader concern about maintaining skilled capacity.
The under-the-radar opportunities are likely to be narrow sub-specialties inside these categories, not necessarily occupations that appear as neat BLS rows: particular generations of industrial controls, legacy power equipment, specialty pumps and compressors, old CNC systems, marine diesels, elevator systems, laboratory instruments, agricultural machinery, industrial furnaces, municipal water systems, historic-building systems and obsolete-but-still-critical infrastructure.
The most valuable niche is therefore not simply the one with the oldest workers. It is the niche in which:
the expert knows something expensive that is not in the manual, the next worker will face substantially similar situations, the work can be observed, and the captured judgment can be reused many times.
Datasets of judgment, talent, and what AI can automate
The phrase “dataset of judgment” is not yet a standard technical category, but the underlying representation is appearing under several names: expert demonstrations, decision traces, cognitive task analysis, experience corpora, trajectories, learning from demonstration, human feedback, case-based reasoning, process supervision and expert cognition models.
KUPAS MASTER is the most direct 2026 example I found. It explicitly argues that ordinary work records omit the key ingredients agents need: which cues mattered, why a judgment was reasonable, what action followed, and when the rule should cease to apply. Its six case elements—
context, cues, judgment, action, boundaries and outcome
—are effectively a formal specification for a dataset of judgment. The platform then organizes extracted knowledge into rules, constraints, best practices, negative examples, corner cases and skills, while preserving source evidence and disagreements.
That is materially different from recording:
“Technician replaced bearing.”
A judgment-rich record might instead capture:
“At this load and temperature, the vibration signature plus intermittent metallic chirp suggested bearing damage rather than misalignment; because the bearing had been replaced only six months earlier, the technician checked lubricant contamination before authorizing replacement; the hypothesis was confirmed during teardown.”
The second record contains conditions, discriminating cues, rejected alternatives and outcome, which are what another human or agent actually needs to generalize.
The emerging research goes beyond industrial maintenance. The 2026 “AI Expert Twin” work proposes computable representations of procedural actions, semantic concepts and expert decision processes and reports a cultural-heritage workshop case study. A related manufacturing project at Carnegie Mellon's Manufacturing Futures Institute proposes extracting operator decision patterns into machine-queryable Observe-Orient-Decide-Act knowledge graphs and validating the approach on industrial troubleshooting and machine changeovers. These are still research projects, but they show the same architecture appearing independently across manufacturing and practice-based learning.
AI is also beginning to automate the interviewer. ACL 2025 research on LLM-powered in-the-moment interviews demonstrates that agents can conduct context-sensitive questioning at scale, while other recent work explicitly compares automated LLM-led knowledge elicitation with human interviewing. This does not prove that an AI can yet interrogate a master machinist appropriately while standing beside a failed press, but it weakens the assumption that every probe must always originate from a human interviewer.
The likely workflow is therefore not “replace the knowledge engineer.” It is to automate the low-value portions of the job:
Continuous video/audio capture
Likely degree of AI automation: Very high.
Human value that remains: Permission, placement, knowing what environment matters.
Transcription and indexing
Likely degree of AI automation: Very high.
Human value that remains: Validation of unusual terminology.
Object/action recognition
Likely degree of AI automation: High and improving.
Human value that remains: Edge cases, subtle sensory interpretation.
Detecting candidate decision points
Likely degree of AI automation: Moderate to high.
Human value that remains: Knowing which apparently minor event is actually important.
Asking routine follow-ups
Likely degree of AI automation: High.
Human value that remains: Social timing, trust and recognition of evasive/incomplete explanations.
Converting sessions into structured cases
Likely degree of AI automation: High.
Human value that remains: Correct abstraction level and boundary conditions.
Finding contradictions across experts
Likely degree of AI automation: High.
Human value that remains: Determining whether contradiction reflects context, error or genuine schools of practice.
Validating “expert truth”
Likely degree of AI automation: Low–moderate.
Human value that remains: Domain accountability and causal judgment.
Deciding what should become agent behavior
Likely degree of AI automation: Moderate.
Human value that remains: Safety, ethics, liability and organizational authority.
The human role consequently looks less like a data-entry specialist and more like a hybrid ethnographer, cognitive interviewer, applied AI engineer and domain apprentice.
The skills that recur across the evidence are:
domain fluency, because someone has to recognize that an unexplained decision just occurred; elicitation skill, including cognitive task analysis and critical-decision interviewing; observational discipline, because what experts do can differ from what they later say they do; AI and multimodal fluency, to configure capture and extraction; knowledge representation, to preserve contexts and exceptions; evaluation/provenance, so machine-generated structure can be audited; and trust-building, because workers will not willingly expose hard-won know-how to a system they believe is intended to eliminate them.
Historical knowledge-acquisition research and current experience-engineering work converge strongly on those requirements.
I found no mainstream university degree or widely accepted credential that currently teaches this complete bundle. Training is fragmented just like the occupation. Current programs teach knowledge graphs and ontologies, context engineering and enterprise AI information architecture, while the elicitation component still comes from cognitive task analysis, human factors, ethnography and domain apprenticeship.
For example, the Knowledge Management Institute currently offers AI-oriented information-architecture training covering taxonomies, ontologies, retrieval and agentic AI; separate 2026 programs train production knowledge-graph pipelines and context engineering. None is yet equivalent to “go onto a factory floor, instrument master practitioners, elicit their judgment and create validated agent skills.”
Interestingly, the AI-training market demonstrates a reverse pathway: start with the expert, then teach enough AI methodology to make their judgment useful to models. Handshake's math-expert role explicitly says no prior AI experience is required. That suggests future knowledge engineering may produce two career routes: AI people becoming sufficiently domain-literate to elicit expertise, and experienced domain practitioners becoming sufficiently AI-literate to codify their own domain.
The historical analogy is essential here. Expert systems experienced a knowledge-acquisition bottleneck decades ago. Researchers found that extracting expertise was difficult because mature expertise becomes procedural and contextual; experts may omit steps that have become automatic, rationalize decisions after the fact, or struggle to state the sensory cues they use. Ethnographic studies of knowledge engineering additionally showed that an elicitor is not simply transferring facts from one container to another—the elicitor is interpreting and constructing a representation of practice.
What changed in 2025–2026 is therefore not that tacit knowledge suddenly became easy. What changed is the economics and fidelity of attempting to capture it:
First-person cameras and microphones are cheap and wearable. Meta exposes camera/microphone functionality through its developer toolkit; Mentra ships programmable camera glasses; industrial systems such as RealWear already exist for harsher environments.
Multimodal models can transform raw recordings into candidate events, objects, actions and explanations instead of forcing humans to manually annotate every minute. Strivr, for example, already creates custom VLMs from captured frontline procedures and uses them to recognize deviations during later work.
LLMs can conduct and summarize interviews, propose structure and find inconsistencies. Recent LLM-interview research and systems such as KUPAS demonstrate how extraction and structuring can increasingly be automated, while leaving expert review in the loop.
Agents can put captured knowledge back into the work immediately. Augmentir, AutoAlign, Myto and TeamViewer all demonstrate different versions of contextual retrieval, work guidance, troubleshooting or agentic action tied to frontline work.
That collapses a historically expensive pipeline:
observe, take notes, transcribe, manually code, formalize, enter into a system, and train people
toward:
work, ambient capture, AI extraction, expert confirmation, machine-usable experience, and immediate reuse.
The economic change is real even though the underlying epistemic challenge remains.
Business models, risks, and what would prove the thesis
The evidence supports all three proposed business models, but for different reasons. It does not yet justify declaring one the winner.
Vertical knowledge company. The evidence for this model is strongest where context itself is proprietary. Myto's proposition is explicitly that generic Internet-trained models do not understand a particular factory's equipment, processes and standards; its AI learns from factory systems, machine history and operator behavior.
Cyvl provides another version in infrastructure, building a domain-specific knowledge representation from physical sensor data, documents and regulation. 01Health shows the embedded model in medicine, where the AI engineer sits with clinicians and continually closes the loop between real work and product behavior.
The attraction is that a vertical company can own not merely software but a progressively richer corpus of domain-specific situations and outcomes. Its weakness is economics: collecting specialized data across plants, aircraft, farms or hospitals may be expensive; data often belongs to customers rather than the vendor; one site's “wisdom” may not transfer cleanly to another; and a sufficiently capable foundation model may commoditize whatever part of the corpus was actually generic rather than truly local.
Knowledge-engineering services. Palantir's forward-deployed model demonstrates willingness to pay highly compensated technical people to live near real customer problems, while connected-worker deployments have long required implementation and workflow work. A future services company could move further upstream: arrive before a retirement wave, map critical expertise, instrument practitioners, conduct cognitive elicitation, create a validated experience corpus, and integrate it with the customer's AI stack. The appeal is that many organizations probably lack this capability internally.
The objection is scalability. The highest-value knowledge engineer may need enough domain competence and trust to recognize what matters. That looks more like elite consulting than SaaS. AI can improve margins by automating transcription, first-pass interviewing, structuring and validation workflows, but the human work may remain the bottleneck. Historically, that bottleneck is precisely what knowledge-engineering projects struggled with.
Knowledge-capture platform. This model has the clearest technology momentum. The required components increasingly exist: Meta/Mentra/RealWear at the capture layer; multimodal perception; automated transcription; enterprise integrations; graph/RAG back ends; contextual agents; and connected-worker delivery. Mission Control is explicitly attempting integrated hardware plus on-prem agent software, while Myto couples wearables and agentic automation, and AutoAlign couples glass/mobile/sensors with an industrial assistant.
Its risk is that “the platform” may not be where defensibility accumulates. Hardware can commoditize; foundation models improve; incumbent connected-worker vendors can bolt on agents; enterprise buyers may insist on owning all captured expertise; and the deepest value may reside in each customer's proprietary experience corpus rather than in the capture software. A successful platform therefore probably needs to become infrastructure for the corpus and feedback loop—not merely a camera application.
There are several reasons to remain skeptical of the overall thesis.
The first is false elicitation. A camera sees behavior, not intent. An interview records an explanation, not necessarily the actual process that generated the decision. Experts may produce plausible post-hoc explanations for pattern recognition that was largely nonverbal. Historical knowledge-acquisition research repeatedly encountered this problem.
The second is knowledge drift. Captured “wisdom” becomes dangerous if equipment, parts, regulation, disease patterns or operating conditions change while users continue treating historical rules as authoritative. KUPAS's emphasis on boundary conditions, evidence, unresolved disagreement, versioning and outcomes points toward the right architecture precisely because static rule capture is insufficient.
The third is surveillance and incentive conflict. First-person cameras and always-on microphones can capture coworkers, customers, proprietary designs, medical information or union-sensitive worker behavior. Workers may reasonably ask who owns their expertise, whether they are training their replacement, and whether the system will be used to measure or discipline them. Meta's acquisition of an always-recording wearable company underscores how strategically attractive ambient capture is, but also why consent and governance become central rather than peripheral product requirements.
The fourth is hardware mismatch. Consumer smart glasses dramatically reduce experimentation cost, but a consumer pair of glasses is not automatically appropriate in a refinery, cleanroom, aircraft-maintenance environment or hazardous electrical setting. Purpose-built industrial hardware exists for a reason. Meta's current developer environment itself remains in preview for important distribution capabilities.
The fifth is foundation-model erosion of the moat. Every piece of expertise that is actually generic, verbalizable and widely documented becomes increasingly available to generic models. The defensible remainder is most likely the knowledge that is site-specific, sensory, historical, rare, outcome-linked or otherwise absent from public corpora. That makes the data-selection problem crucial: capturing everything is not the same as capturing scarce judgment.
The sixth is lack of independent outcome evidence for the newest entrants. Myto, Mission Control and AutoAlign have compellingly on-thesis product descriptions, but many performance and deployment statements are company-reported. Mission Control is offering 90-day pilots; AutoAlign reports pilot metrics; Myto describes its architecture and investors but does not publicly identify a broad customer base on the material reviewed. In contrast, TeamViewer and Augmentir have more mature public customer evidence but are less purely focused on tacit elicitation.
That is why the best description of the competitive landscape today is:
The pieces are becoming commercial faster than the category is becoming proven.
Signals Worth Watching
The following are measurable signals that would distinguish a genuine new occupation and industry from a temporary relabeling of knowledge management, AI consulting and connected-worker software.
Dedicated tacit-knowledge job titles
Meaningful evidence would be: Hundreds of postings using terms such as AI Knowledge Engineer, Experience Engineer, Knowledge Capture Engineer or equivalent where duties explicitly include interviewing/observing experts in real work—not just RAG and ontologies.
Job-description convergence
Meaningful evidence would be: A recurring skill bundle of cognitive task analysis + field observation + multimodal capture + AI/agents + knowledge representation + evaluation.
Compensation normalization
Meaningful evidence would be: Enough dedicated postings to establish a distinct salary band rather than inferring compensation from FDE and graph-engineering jobs.
Retirement-capture programs
Meaningful evidence would be: Fortune 500, utilities, defense contractors or governments publicly budgeting programs specifically to capture expertise from retirement-eligible employees before departure.
Agentic elicitation in production
Meaningful evidence would be: Commercial systems that autonomously recognize a consequential decision and ask the practitioner a context-sensitive follow-up during work.
Structured judgment schemas
Meaningful evidence would be: Industry convergence on something analogous to context–cue–judgment–action–boundary–outcome records, rather than storing only transcripts and video.
Outcome-linked corpora
Meaningful evidence would be: Vendors demonstrating that captured expert cases are systematically connected to later equipment, safety, quality or business outcomes.
Closed-loop learning from the next cohort
Meaningful evidence would be: Public deployments showing that newer workers' exceptions, corrections and outcomes update the corpus rather than merely consuming a static expert library.
Independent ROI evidence
Meaningful evidence would be: Peer-reviewed or independently audited evidence of lower downtime, faster time-to-proficiency, higher first-time-fix rates, fewer errors or lower training costs attributable specifically to captured expertise.
Named production customers for AI-first entrants
Meaningful evidence would be: Myto-, Mission Control- or AutoAlign-like companies progressing from pilots and vendor claims to multi-site deployments with identifiable major customers.
Large financing or acquisition events
Meaningful evidence would be: A major industrial software, enterprise AI, ERP, defense or workforce-tech company acquiring a tacit-capture company specifically for its expertise-capture capability.
Wearable platform maturity
Meaningful evidence would be: Meta, Mentra or comparable ecosystems supporting stable enterprise deployment, privacy controls, device management, longer runtimes and integration with industrial-grade/PPE hardware.
Training and credential formation
Meaningful evidence would be: Universities, human-factors programs, trade schools or professional bodies creating formal curricula that combine elicitation, domain apprenticeship and AI engineering.
Worker participation models
Meaningful evidence would be: Contracts, union agreements or compensation mechanisms explicitly addressing consent and ownership when employee expertise is captured to train organizational AI.
Corpus defensibility evidence
Meaningful evidence would be: Companies demonstrating that proprietary experience data materially improves agents beyond frontier-model + generic-RAG baselines, especially across rare failures and ambiguous edge cases.
The most important of these is probably not growth in the literal title “Knowledge Engineer.” Current evidence suggests the profession could mature for several years while remaining distributed among forward-deployed AI engineers, domain experts, cognitive-task specialists, industrial AI teams and connected-worker deployments.
The decisive signal will be whether enterprises begin treating expert judgment as an asset that requires a deliberate acquisition pipeline analogous to how they already treat software, data and physical equipment.
That is the conceptual transition already visible at the edge of the market: from knowledge management to experience engineering; from collecting documents to collecting situations, cues, judgments, actions and outcomes; and from asking “What does our organization know?” to asking:
“What can our best people perceive and decide today that our people and machines will otherwise be unable to reproduce tomorrow?”