Solution Archetypes: Grouping the Opportunities

Purpose and Method

The companion papers in this series — The Architecture of Work (the structural model), the Opportunity Assessment (each Layer 3 activity assessed in isolation), and the Sector Atlas (the same assessments organized by industry) — deliberately treated every activity separately. This paper does the opposite. It reads across all 136 activity assessments and asks: when the individual assessments are compared, do the opportunities cluster? Do many activities, scattered across unrelated patterns and industries, call for the same kind of solution?

They do. The 136 AI-reasoning assessments collapse into seven reasoning groups. The robotics assessments collapse into six robotics groups. The traditional-automation assessments collapse into five automation groups. This paper defines each group, states the attributes that bind its members, names representative members from across the economy, and specifies the solution types each group calls for.

The grouping logic differs from the companion papers’ logic, and the difference matters. Layer 2 patterns group activities by cognitive structure — what operation, on what substrate. The groups in this paper cluster activities by solution demand — what has to be built, bought, governed, and staffed to capture the opportunity. The two cut across each other. Two activities in the same Layer 2 pattern can demand different solutions (rubric grading and holistic admissions are both Adjudication, but only one is a straight-through candidate), and two activities in different patterns can demand the same solution (AML alert adjudication and insurance claims handling sit in different patterns but call for an identical solution shape). Solution demand, not structural kinship, is what an implementing organization actually procures against — which is why this fourth cut through the material earns a standalone document.

The attributes used to cluster are those that the isolated assessments implicitly turned on:

For AI reasoning: whether governing criteria are explicit or tacit; whether the substrate is fully representational; whether the output is a new artifact or a judgment about an existing one; whether volume permits an exception-routing economics; whether the ground truth is verifiable, delayed, or contested; whether an adversary adapts against the method; and whether institutional authority (law, license, liability) caps machine autonomy regardless of capability.

For robotics: the structure of the environment; whether the task is sensing, movement, or manipulation; the dexterity demanded; the consequence of contact error; the proximity of untrained humans; and the regulatory posture of the operating domain.

For traditional automation: whether the system’s job is keeping state, moving data, executing rules, sensing and controlling in real time, or performing fixed kinetic work — and whether determinism is a limitation to be transcended or a guarantee to be preserved.

Each group below is presented the same way: defining attributes, why the members belong together, representative members drawn from across the Sector Atlas, and the specific solution types the group calls for.

Part I — The Seven AI Reasoning Groups

R1. Codified Adjudication Engines

Defining attributes. Explicit, published, versioned criteria; evidence that arrives as records; a required reasoning trail from evidence through criterion to conclusion; high case volume with a difficulty distribution — a large tractable body and a hard tail; and a consequential, attributable output that some institution must stand behind.

Members. Bank examination, financial statement audit, tax examination, insurance claims adjudication, immigration and benefits determination, patent examination, regulatory submission review, quality acceptance, appraisal, rubric-based grading, AML alert adjudication, sanctions-hit resolution, contract playbook review, eligibility screening in social services, prior-authorization review, building plan review.

Why they cluster. Every member is the same machine wearing different uniforms: ingest a case file, retrieve the applicable criteria, apply them, document the chain, and route by confidence. The assessments for these activities were nearly interchangeable — “very high; straight-through resolution of the clear majority with human review of the contested residual” — because the solution is nearly interchangeable. Note the membership spans five Layer 2 patterns; what binds it is the solution shape, not the pattern.

Solutions called for. Architecture: retrieval-grounded reasoning over an authoritative, versioned criteria corpus; extraction and normalization of the case file; chain-of-citation output in which every finding links its evidence and its criterion; calibrated confidence scoring that routes cases into auto-resolve, machine-draft-for-review, and human-first lanes. Quality regime: a gold-case evaluation harness built from historical adjudications; drift monitoring when criteria versions change; sampling-based human audit of the auto-resolved lane, mirroring the sampling logic these fields already apply to themselves. Governance: model risk management of regulatory grade — documented validation, challenge processes, appeal rights for affected parties, and clear human accountability for the residual and for the system itself. In these domains the pacing constraint is institutional acceptance, so the governance artifacts are not overhead; they are the product. Human-role design: the practitioner becomes a reviewing official — calibrating the machine, owning the hard tail, and adjudicating the appeals. Staffing models, training (see R7’s note on formation), and career paths must be rebuilt around review rather than first-pass production.

R2. Constraint-Bound Generators

Defining attributes. The output is a new artifact — code, a drawing, a molecule, a clause, a lesson, a campaign — that must satisfy an explicit constraint set; candidate artifacts are cheap to produce and cheap to test in representational form; and a verification mechanism (compiler, simulator, checker, test suite, code-compliance engine, assay) can score candidates faster and more objectively than a human reader.

Members. Software development, structural and civil design, chip design, contract and statute drafting, drug and molecule design, process and plant design, curriculum and assessment generation, industrial design, financial structuring, experiment and protocol design, game and narrative asset production, brand creative.

Why they cluster. Each is a generate-and-verify loop, and the economic transformation is the same in all of them: generation becomes nearly free, so the binding scarcity moves to specification (stating requirements and constraints well) and verification (judging candidates efficiently and catching the constraint the spec forgot). The assessments repeatedly located the durable human contribution in exactly those two places.

Solutions called for. Architecture: a generation engine coupled tightly to the domain’s verifier — test suites for code, analysis models for structures, design-rule checks for silicon, simulators for plants, assays for molecules, playbook checkers for contracts — with the loop run at machine speed and only verified candidates surfaced. Process design: invest in the specification layer: machine-assisted requirements elicitation, constraint libraries, and executable acceptance criteria. The organizations that win in this group are those whose constraints are written down and testable. Quality regime: verification coverage is the metric that matters; an unverified generated artifact is inventory, not output. Provenance tracking of what generated what, under which constraint version, for liability and revision. Human-role design: practitioners shift from making candidates to setting requirements, curating constraint sets, exercising taste among verified candidates, and owning the professional stamp. Seat counts in first-draft production fall; verification engineering and specification craft become the hiring profile.

R3. Telemetry Diagnosticians

Defining attributes. An opaque system misbehaving; a fault model — accumulated knowledge mapping symptoms to causes; probes that expose hidden state at a cost; and hypothesis competition resolved by evidence, ending in an intervention whose success verifies the diagnosis.

Members. Software debugging and incident response, equipment and vehicle troubleshooting, network and grid fault isolation, medical differential support, root-cause analysis after failures, organizational diagnostics, environmental and epidemiological source attribution, remediation validation.

Why they cluster. All are abduction with a reference model, and their assessments converged on the same machine role: hold the full fault model (which no individual can), generate the complete differential, reason about which probe discriminates most per unit cost, and update as results arrive. The members differ mainly in how physical the probes are — which is a robotics question (Group B3/B4), not a reasoning one.

Solutions called for. Architecture: agentic systems with tool access to the probes — query the logs, order the test, run the trace, pull the maintenance history — plus an explicit hypothesis ledger with likelihoods, so the reasoning is inspectable mid-investigation, not only at conclusion. Knowledge assets: the fault model is the moat. Solutions must capture every resolved case back into the model — symptom, probes run, cause found, fix verified — converting the organization’s incident history from narrative archive into diagnostic capital. Autonomy design: probe-and-remediate authority granted in tiers by blast radius: free rein over read-only probes, policy envelopes for reversible interventions, human gates on irreversible ones. The software-incident members will reach closed-loop remediation first and serve as the template. Human-role design: humans own premature-closure defense (the classic abductive failure now transposed to machine anchoring), intervention approval, and the novel case that the fault model has never seen.

R4. Live Planners

Defining attributes. A continuously updated representation of many resources and demands; combinatorial allocation under constraint; a plan that decays on contact with reality and must be re-solved continuously; and fast, unambiguous feedback (the schedule held or it didn’t).

Members. Production planning and scheduling, freight and network routing, OR and clinic scheduling, project replanning, grid dispatch and DER orchestration, air-traffic flow support, warehouse wave and labor planning, exam-engagement scoping, incident-resource allocation, crew and fleet recovery.

Why they cluster. Every member already runs on a model of its world — the APS, the network model, the schedule, the surveillance picture — so the reasoning lever slots into an existing representational socket. The assessments were uniform: the optimization core is machine-superior now; the frontier is fusing learned prediction (durations, demand, failure risk) with the solver, and handling the human negotiation around the plan.

Solutions called for. Architecture: hybrid stacks — learned predictors feeding constraint solvers — operating on a live state representation (a digital twin in the physical members, the tracked plan elsewhere), replanning on every state change rather than on planning cycles. Data infrastructure: the constraint is usually state freshness, not solver quality; instrument the gap between reality and its representation (Group T2’s territory) before buying better optimization. Autonomy design: machine-issued plans within policy envelopes; human-on-the-loop for envelope exceptions and for the persuasion layer — the stakeholders who must accept the plan are not variables in it. Human-role design: planners become envelope-setters and exception negotiators; the status-gathering half of coordination work (meetings held to learn state) is eliminated by inferring state from the work systems directly.

R5. Bounded Interaction Agents

Defining attributes. The substrate is a cooperative mind engaged through conversation; the interaction has a service objective (resolve, teach, guide, comfort, sell); competence is partly tacit but a large tier of the volume follows recognizable scripts; and failure modes are relational and safety-laden, not merely factual.

Members. Customer support and success, tutoring and adaptive instruction, sales preparation and qualification, benefits navigation, patient and member engagement, structured therapeutic support, hospitality concierge service, coaching and onboarding assistants, companionship supplements in care settings.

Why they cluster. Every member splits the same way in its assessment: a transactional tier the machine handles end-to-end, and a relational tier — the distressed, the complex, the high-stakes, the genuinely personal — where the human relationship is the product. The solution problem is identical across them: build the machine tier well, and engineer the boundary honestly.

Solutions called for. Architecture: conversational agents grounded in the organization’s knowledge and the individual’s context and history; persistent memory with consent; multilingual by default; instrumented for outcome, not just deflection. Boundary engineering: the escalation design is the product’s most important feature — detection of distress, confusion, risk, and value; warm handoff carrying full context to the human tier; and no dark patterns that trap people in the machine tier. In care-adjacent members, honest presentation of the agent as an agent is an ethical floor, and safety behavior (crisis detection and referral) must be engineered and tested to a higher standard than the happy path. Quality regime: conversation review sampling, outcome tracking against the human baseline, and adversarial testing for manipulation, unsafe advice, and boundary failures. Human-role design: the human tier shrinks in headcount and rises in skill and stakes — staffed, trained, and paid as the relational specialists they now are, with the machine preparing their context rather than competing with them.

R6. Adaptive Sentinels

Defining attributes. An adversary who studies the defense and adapts; a monitored stream (transactions, traffic, content, access events) in which hostile events hide among benign volume; asymmetric error costs; and a method that decays by design, because the substrate routes around whatever is deployed.

Members. Cybersecurity detection and response, fraud scoring and interdiction, AML scenario monitoring, content moderation and platform integrity, insider-threat detection, border and customs targeting, counterintelligence analytics, and the anti-gaming layer of any Group R1 system whose criteria invite manipulation.

Why they cluster. These assessments alone carried the phrase “competitive necessity”: the adversary is adopting the same machine capabilities, so standing still is losing ground. Statically trained models — the previous generation’s answer — inherit the decay; the group’s real demand is an institutionalized adaptation loop, which is a different thing to build than a detector.

Solutions called for. Architecture: layered detection — deterministic rules as the floor (Group T3), learned models above, and reasoning-grade investigation of the flagged residual (an embedded R1/R3 engine that adjudicates and documents each alert). Adaptation loop: confirmed-outcome feedback wired directly into retraining; red teams — increasingly machine red teams — probing the defense continuously; detection engineering that converts fresh intelligence into deployed logic in days, not quarters. Governance: precision matters ethically as well as economically — false positives here freeze accounts, remove speech, and accuse colleagues. Contestability, appeal, and human review of consequential actions are structural requirements, and in several members (moderation, insider threat, targeting) civil-liberties constraints properly cap autonomy below technical capability. Human-role design: analysts move up the stack to novel-typology hunting, adversary emulation, and judgment on the cases where the model is least certain and the stakes are highest.

R7. Open-Inquiry Engines

Defining attributes. The question has no answer key and no reference model — no one knows; method is a discipline, not a determinant; validation is slow, social, and statistical; and the failure mode is confident falsehood propagating downstream.

Members. Basic and applied research, pharmaceutical discovery reasoning, economic and policy research, market and user research synthesis, intelligence assessment, investigative journalism analysis, prospectivity analysis in exploration, mathematical inquiry.

Why they cluster. These assessments shared a distinctive double edge: the reasoning lever’s ceiling is enormous (superhuman literature synthesis, hypothesis generation, simulation), and its characteristic risk is unique — fluent, plausible, wrong. Every other group has fast or eventual ground truth; this group’s ground truth is expensive, so epistemic discipline must be engineered into the system rather than awaited from the world.

Solutions called for. Architecture: research agents with full provenance — every claim traceable to source or computation; explicit uncertainty representation and competing-hypothesis tracking (the intelligence community’s analytic tradecraft, generalized); simulation and formal verification wherever the domain permits (the mathematical member shows the ceiling: machine-verified proof). Validation coupling: couple the reasoning engine to the fastest available truth source — closed-loop laboratories (Group B1), pre-registered analyses, replication pipelines — so hypotheses meet reality at machine cadence rather than publication cadence. Quality regime: institutional defenses against machine-scale error: adversarial review of machine findings, source-integrity checking as generated content contaminates the literature itself, and explicit confidence language in every product. Human-role design: problem selection, taste, interpretation, and accountability for the claim — the activities the assessments consistently reserved — plus a new one: epistemic supervision of systems that read more than any human ever will.

A note spanning all seven groups. Formation (Layer 2 Pattern 8) appears inside every group’s solution rather than as a group of its own: each reasoning system changes what its human counterparts must know, and each generates, as a byproduct, the case libraries and simulation environments with which those humans can be trained. Solution programs that omit the retraining pipeline strand their own human tier.

Part II — The Six Robotics Groups

B1. Contained Kinetics

Defining attributes. An engineered, bounded environment that can be shaped around the machine; repeatable objects and tasks; the option of physical segregation from people; and throughput economics — the robot is bought with an ROI spreadsheet, not a research program.

Members. Warehouse goods-to-person and sortation, manufacturing machine tending and rigid assembly, welding and painting cells, laboratory automation and pathology pre-analytics, pharmacy dispensing robotics, food-preparation cells in standardized formats, automated parking and storage/retrieval, print finishing and bindery, container-terminal cranes and yard automation.

Why they cluster. These are the environments where robotics already works, and the assessments uniformly rated them “high to very high” with the same reasoning: when the world can be structured, the robot’s perception and dexterity burden collapses. The 3–5 year gains come less from new robot capability than from cheaper integration — faster teaching, vision-guided tolerance for variation, and standard cells — which pushes viability down into higher-mix, lower-volume operations that could not previously justify the engineering.

Solutions called for. Procurement posture: buy, don’t build — this is a mature vendor market of proven cells and fleets; competitive advantage lies in deployment speed and integration quality, not robot invention. Facility design: the highest-leverage act is structuring the environment — standardized totes, fixtures, presentation, and flow — because every unit of environmental structure is a unit of robot capability you don’t have to purchase. Integration layer: fleet orchestration tied into the WMS/MES (Group T1) and the live planner (Group R4); the robots are executors of the plan, and the plan’s quality now binds throughput. Human-role design: the residual human work is exception handling, changeover, and maintenance of the fleet — a technician profile, hired and trained deliberately rather than assumed.

B2. Bounded Movers

Defining attributes. The task is navigation, not manipulation; a definable operational design domain — fixed corridors, geofenced districts, private land, segregated rights-of-way, reserved airspace; safety certification and regulatory permission as the true gate; and fleet economics with remote supervision.

Members. Hub-to-hub autonomous trucking, geofenced robotaxi operations, agricultural autonomy on open fields, mining and quarry haulage, automated transit on segregated guideways, drone delivery on fixed corridors (medical logistics foremost), yard trucks and airport ground movement, autonomous marine and aerial survey transits.

Why they cluster. Every assessment of vehicle autonomy carried the same qualifier — in bounded domains — because the technical problem is the domain, not the vehicle. Members succeed in the order of their domains’ simplicity: private land first, fixed corridors next, dense mixed traffic last. The solution work is therefore domain engineering and regulatory engineering as much as autonomy engineering.

Solutions called for. Domain definition: rigorous ODD specification — routes, weather, hours, hand-off points — and route/infrastructure investment (mapping, transfer hubs, corridor sensing) that buys down the autonomy problem the way facility structure does in B1. Operations model: remote supervision centers with defined supervisor-to-vehicle ratios that improve over time; degraded-mode and minimal-risk-condition design; incident response as a built discipline. Regulatory strategy: certification evidence — safety cases, disengagement data, simulation miles — treated as a first-class product; jurisdiction sequencing as the actual rollout plan. Human-role design: driving work converts to supervision, first/last-leg handling, and hub operations; the transition plan for the existing workforce is part of the deployment, not an afterthought.

B3. Instrumented Eyes

Defining attributes. The robot’s job is sensing, not touching — access to places that are hazardous, elevated, submerged, confined, remote, or simply expensive to visit; near-zero contact risk; and output that is data for the reasoning layer, not a completed physical task.

Members. Drone and crawler inspection of bridges, pipelines, dams, track, turbines, and boilers; construction site capture; roof and catastrophe capture for claims; conservation census platforms and camera/acoustic arrays; environmental sampling drones and buoys; security patrol robots as mobile sensors; wreckage and hazard-zone mapping; downhole and subsea survey.

Why they cluster. Across wildly different sectors the assessments described the identical value chain: platform captures → data pipeline normalizes → AI interprets → human acts. Because nothing is manipulated, the safety case is easy and adoption is fast; these members were consistently rated the most mature robotics opportunity in their sectors. The robot is best understood as the mobile end of a perception system whose intelligence lives in Groups R1/R3/R6.

Solutions called for. System design: buy the capture platform, build the data product — flight/route planning, capture standards, georeferenced time-series storage, and the interpretation models that turn imagery into condition scores, deficiencies, counts, and alerts. The differentiation is entirely downstream of the airframe. Baseline discipline: value compounds only if capture is repeatable — same angles, same resolution, same intervals — so that change detection, the real product, is possible. Capture standards are the group’s equivalent of criteria versioning. Integration: condition outputs must land in the systems that program work (CMMS, claims platforms, enforcement queues — Group T1), or the data pipeline ends in a folder. Autonomy trajectory: progress runs from piloted, to autonomous capture with human review, to continuous unattended sensing (fixed arrays and resident robots) — plan the program as that progression.

B4. Guided Hands

Defining attributes. High-consequence physical intervention where human judgment must stay in command; the robot contributes precision, steadiness, reach, or survivability that human hands lack; autonomy enters as supervised subtasks inside a human-directed procedure; and credentialed operators within regulated practice.

Members. Surgical robotics and interventional platforms, bomb disposal and hazardous entry, nuclear and contaminated-environment work, subsea and space manipulation, energized line maintenance, remote heavy-equipment operation in dangerous zones, laser conservation cleaning, teleoperated crisis intervention devices.

Why they cluster. The assessments drew the same line in each: teleoperation established, autonomous subtasks realistic within the horizon (suturing runs, milling-to-plan, standardized manipulations in well-imaged fields), full autonomy out of scope. The binding constraints are consequence and accountability, so the solution shape is human-commanded machines that earn autonomy one verified subtask at a time.

Solutions called for. Capability design: haptic and sensory feedback fidelity; imaging and registration that give the machine (and the operator) a truthful model of the worksite; subtask autonomy packaged as discrete, certifiable functions with explicit engage/disengage. Verification: subtask-level validation against defined performance envelopes; simulation-based rehearsal on the actual case (the patient’s imaging, the site’s scan) before the intervention. Credentialing: operator training and proficiency on the platform as a formal program (Group R5/Formation adjacency: simulators generate the training). Human-role design: the operator’s craft shifts toward planning, supervision, and exception takeover — the aviation model of automation, imported into procedures.

B5. Cohabitant Assistants

Defining attributes. Operation among untrained people in shared, semi-structured spaces; narrow tasks (carry, fetch, lift, guide, patrol, deliver); safety and social acceptance as the design center; and value measured in relieved human effort rather than replaced human roles.

Members. Hospital and facility transport robots, delivery robots in public and campus spaces, patient-lift and transfer assistance, person-directed assistive manipulators and smart mobility devices, hotel and restaurant runners, autonomous cleaning fleets in occupied buildings, telepresence platforms, monitoring companions in care settings.

Why they cluster. The assessments in care, hospitality, and facilities converged on the same shape: robotics succeeds here exactly insofar as it takes the logistics out of human-facing work — the carrying, the fetching, the lifting — and returns that time to the relational work that Groups R5 and the Care pattern reserve for people. The person-directed members (disability assistance) sharpen the principle: the robot as capability extension under the served person’s command, never as a substitute presence.

Solutions called for. Safety and social design: inherently safe hardware (speed, force, mass limits), legible behavior — people must be able to predict the machine — and interaction design tested with the actual population, including those with impairments. Task discipline: resist scope creep toward dexterity and intimacy; the group’s failures are attempts to make cohabitant platforms do B6 work. Narrow, reliable, boring tasks are the winning profile. Operations: fleet management, recovery-from-stuck as a service discipline, and facility accommodations (door integration, elevator access, charging) planned as infrastructure. Ethical posture: in care settings, deployment must be measured by whether human presence increases where it matters — the explicit test the assessments imposed on every care-adjacent robotics claim.

B6. The Dexterity Frontier (a residual group, held honestly)

Defining attributes. Unstructured, improvisational manipulation — varied objects, confined and cluttered spaces, force-feel and compliance, flexible materials, occupied environments; and, per the horizon discipline of this series, no realistic expectation of economic autonomy within 3–5 years.

Members. Household and building trades repair, varied garment assembly, general kitchen craft, hands-on personal care (bathing, dressing), delicate harvesting at scale, general construction assembly in the open site, equipment repair in the field.

Why they cluster. These are the assessments that said “remains human through the horizon,” and grouping them is as important as grouping the opportunities: it marks where robotics capital should not chase substitution. General-purpose humanoid platforms are advancing and will begin entering the adjacent structured niches (B1’s edges) within the horizon, but the members of this group are the last, not the next.

Solutions called for. Augmentation, not substitution: AR-guided procedures and remote expert support (importing Group R3’s diagnostician to the human’s eyes), instrumented and powered tools, exoskeletons for load and repetition, and better capture (B3) so the human trip is single and prepared. Upstream redesign: move the work into structured settings where the other groups apply — prefabrication and modular construction, garment designs engineered for automated sewing, crop varieties and trellising engineered for mechanical harvest, appliances designed for modular swap-out rather than in-place repair. The durable answer to unstructured work is often to stop generating it. Workforce implication: these trades gain leverage and scarcity value through the horizon; formation pipelines for them (Pattern 8) are an investment, not a legacy cost.

Part III — The Five Traditional Automation Groups

T1. Systems of Record and Workflow Backbones

Defining attributes. Authoritative state — the case, the asset, the order, the patient, the contract — plus routing, status, and audit trail; long-lived, deeply integrated, compliance-laden platforms; and the property, visible throughout this series, that they are the representational capture on which every reasoning and robotics lever depends.

Members. Core banking and policy administration, EHRs, ERP/MES, CMMS and asset registries, case management in courts, agencies, and social services, CLM, WMS/TMS, LIMS, project and exam workflow platforms, land and corporate registries.

Why they cluster. In assessment after assessment, the “traditional automation” line named a system of this kind and the “AI reasoning” line presupposed it. The group’s solution demand is therefore strategic: completeness and cleanliness of representational capture is the single variable that raises the ceiling of every other group in this paper.

Solutions called for. Finish digitization — every case element born digital, no paper shadow processes; API-first and event-emitting architectures so reasoning systems can read and write as first-class actors; data-model hygiene and master-data discipline (identity resolution above all); embedded audit-trail capacity sized for machine actors, whose action volume dwarfs human clerks’; and explicit “agent access” design — permissions, rate controls, and attribution for non-human users — which the coming period will demand of every platform in the group.

T2. Sensing, Telemetry, and Deterministic Control

Defining attributes. Real-time capture of physical state and closed-loop deterministic control against setpoints; hard-real-time reliability and safety certification; and the role, throughout the physical economy’s assessments, of being the substrate the reasoning layer reads and the safety floor beneath anything it commands.

Members. SCADA/DCS and historians, protective relaying and safety-instrumented systems, building automation, physiological monitoring and smart infusion, cold-chain and condition telemetry, structural and environmental sensor networks, avionics and vehicle control layers, track-and-trace scanning infrastructure.

Why they cluster. The assessments never proposed replacing this layer; they proposed thickening it (more sensing) and supervising it (learned optimization writing setpoints within its guarantees). Its determinism is not a limitation to transcend but the property that makes machine reasoning deployable in physical domains at all.

Solutions called for. Instrumentation build-out prioritized by the reasoning use cases downstream (predictive maintenance, condition-based programs, live planning); historian and time-series infrastructure treated as an analytic product, not an archive; alarm rationalization so human and machine attention lands on signal; a strict architectural contract in which learned systems optimize within envelopes that deterministic safety systems enforce — the pattern (advisory → supervisory → closed-loop within limits) that the energy, process, and medical assessments all converged on; and cybersecurity for the control layer as a precondition, since Group R6’s adversaries target exactly this substrate.

T3. Rules Engines and Straight-Through Execution

Defining attributes. Fully codified logic executed deterministically at volume — the same input always yields the same output; auditability and fairness by construction; and a boundary with Group R1 that is the most consequential design decision in the entire automation stack: what stays rules, what becomes judgment.

Members. Payment processing and settlement finality, screening thresholds and eligibility calculations, laboratory autoverification, claims auto-pay rules, protective logic and interlocks, CI/CD gates and policy-as-code, tax and payroll computation, matching and reconciliation engines.

Why they cluster. The assessments called this group “the floor”: the layer that already resolves the clear majority everywhere it exists, leaving the exception tail that Group R1 now addresses. Its determinism is a feature the reasoning layer cannot replicate — provable consistency — so the solution demand is not replacement but principled division of labor.

Solutions called for. Rules-as-code discipline: versioned, tested, human-readable rule sets with regulatory traceability; a designed rules/judgment boundary — deterministic logic wherever criteria are truly mechanical (with machine reasoning used offline to author and verify rules rather than to execute them), reasoning-layer routing for the genuinely judgmental residual; idempotent, replayable pipelines; and continuous reconciliation of the two layers’ outputs, so drift between what the rules decide and what the judgment layer decides is itself surfaced as a finding.

T4. Connective Plumbing

Defining attributes. Movement of structured data across organizational and system boundaries — formats, protocols, translations, acknowledgments; no judgment, no state ownership; and a large installed base of brittle, semi-manual bridges whose cost is invisible until counted.

Members. EDI and B2B document exchange, healthcare interface engines and interoperability standards, customs and trade single-window interfaces, open-banking and payment APIs, ETL/ELT and event streams, RPA deployed as screen-level bridging, e-filing and e-submission gateways.

Why they cluster. Whole professions cataloged in the Sector Atlas — freight forwarding’s documentary chain, revenue cycle intermediation, back-office repair queues — exist substantially because this plumbing is incomplete, and their AI-reasoning assessments amounted to “a model now reads what the interface couldn’t parse.” That is the group’s strategic tension: reasoning models can paper over missing integration by reading unstructured artifacts, which is invaluable at the boundary you don’t control and technical debt at the boundary you do.

Solutions called for. API- and event-first modernization inside the estate, with RPA and model-based document reading explicitly designated as bridges to be retired on a schedule wherever both endpoints are controllable; canonical data models and identity resolution at the boundaries; machine-readable standards adoption in industry consortia (the shipping-container move, replayed in data); and, at uncontrolled boundaries, hardened extraction — the reasoning layer as a permanent, monitored adapter with confidence scoring and human fallback.

T5. Fixed Kinetic Automation

Defining attributes. Single-purpose machines executing one physical transformation at maximal rate and consistency — no perception, no flexibility, no reprogramming worth the name; peak unit economics on stable, high-volume tasks; and a design boundary with Group B1 that is the physical world’s version of T3’s rules/judgment line: what should be hard automation, what should be flexible robotics.

Members. Conveyance and sortation hardware, CNC and dedicated machining lines, filling, forming, and packaging lines, printing presses, milking parlors and grain handling, automated analyzers, irrigation and dosing systems, dedicated test rigs.

Why they cluster. The assessments’ “very high and mature” traditional-automation ratings in manufacturing, agriculture, laboratories, and logistics largely name this group. It is not legacy: for any transformation whose specification is stable at volume, fixed automation beats flexible robotics on cost, speed, and reliability, and the correct system design hardens stabilized robot tasks into fixed automation over time.

Solutions called for. A deliberate hard/flexible boundary review as product variety and volumes shift — fixed automation for the stable core, B1 robotics for the variable edge, with migration paths both directions; instrumentation retrofits (Group T2 sensors on Group T5 iron) so the installed base joins the representational layer and becomes visible to predictive maintenance and live planning; and changeover engineering — the group’s classic weakness — as the highest-return investment where mix is rising.

Closing Synthesis: How the Groups Compose

The groups are levers on the same machine, and the assessments’ recurring architecture assembles them in a consistent stack. T1 holds the state and T4 moves it; T2 senses the physical world and enforces its safety floor; T3 executes everything truly mechanical; T5 and B1 perform the stable physical work while B2 moves things and B3 watches things; upon that substrate, R4 plans, R1 judges, R3 diagnoses, R2 generates, R6 defends, R7 inquires, and R5 speaks — with B4 and B5 extending human hands and relieving human legs where the physical world resists, and B6 marking where it still refuses.

Read as a portfolio, the groupings yield three practical rules. First, sequence by substrate: reasoning groups pay off in proportion to the completeness of T1/T2 capture, so representational debt is the true backlog. Second, buy maturity, build judgment: the robotics and automation groups are predominantly vendor markets where integration wins, while the reasoning groups’ quality regimes, criteria corpora, fault models, and escalation designs embody the organization’s own judgment and must be owned. Third, design the boundaries as products: rules versus judgment (T3/R1), autonomy envelope versus human command (T2/R4, B4), machine tier versus relational tier (R5), hard versus flexible automation (T5/B1) — in every group, the boundary line, not the technology on either side of it, is where the consequential engineering lives.