Where AI Delivers Real Value: Domain-by-Domain Assessment

A New Class of Institutional Asset

An AI system fine-tuned on organizational data is less a tool than a standing capability: available around the clock, retained across staff transitions, and improving on the organization’s own data as that data grows. The argument is about knowledge continuity, not loyalty or headcount. Most of what makes a senior engineer valuable is tacit, the experiential pattern recognition that lets them diagnose in minutes a failure a new hire cannot reproduce in a week. Tacit knowledge resists documentation by definition, which Polanyi captured in the observation that we know more than we can tell [1]. That is exactly why it walks out the door when its holder does, a loss the knowledge-management literature has labeled corporate amnesia, organizational Alzheimer’s, and enterprise dementia [2].

The cost surfaces in onboarding. For knowledge-based roles, Gartner’s Brian Kropp puts the floor at six months, below which an onboarding program “will lead to performance and turnover problems,” while objective ramp-to-productivity data points to twelve to fifteen months [3]. Organizations lose between a third and two-thirds of new hires inside the first year [3]. Two structural facts make this hard to manage through conventional HR practice: organizations cannot reliably predict who is about to leave, and departing staff rarely invest in transfer before they go. A system that captures knowledge continuously during employment sidesteps both.

An AI-indexed knowledge base does not reduce attrition. It changes what attrition costs. Knowledge held in past project documentation, incident reports, design reviews, and code repositories becomes queryable by successors instead of locked in the heads of people who have moved on. The institutional-asset argument and the workforce argument are the same claim seen from two sides: AI augments output, and on an owned system that augmentation accrues to the organization permanently rather than only for the duration of one person’s tenure [4].

Software Development

Software development has the most extensively studied AI-productivity evidence. In a controlled experiment by GitHub-affiliated economists, developers given GitHub Copilot finished a JavaScript HTTP-server task 55.8% faster than a control group; the confidence interval is wide, from 21% to 89%, with the understanding that the authors’ affiliation warrants reading it as a first-party result on GitHub’s own product [5]. The independent corroboration carries more weight. McKinsey’s study of developers on representative tasks found AI cut documentation time by roughly half, new-code authoring by nearly half, and refactoring by close to two-thirds [6]. A separate GitHub-Accenture randomized trial reported developer-experience gains, with 90% feeling more job fulfillment and 95% enjoying coding more [7].

These gains hold for well-scoped, generation-heavy work. On complex tasks in unfamiliar frameworks, McKinsey found time savings collapse below 10%, and the benefit there is a higher chance of finishing within the allotted time, a reliability gain rather than a speed gain [6]. That is the same boundary the controlled developer-productivity trial in the productivity-paradox analysis mapped from the other side: experienced developers on complex tasks in mature codebases ran 19% slower with early-2025 tools, and METR’s 2026 follow-up could not reliably size the current effect because wider AI adoption broke the controlled comparison, leaving point estimates that straddle zero and a qualitative read that developers are probably modestly faster than a year earlier. The McKinsey and METR results do not conflict. They mark the task-conditional line: AI accelerates generation-heavy, well-scoped work and does not reliably accelerate judgment-heavy work in complex existing systems.

The reliable wins for technical organizations are the generation-heavy ones: completing code against established internal patterns, scaffolding tests and boilerplate, modernizing legacy code, drafting API documentation, and assisting review on routine changes. Junior-to-mid developers on well-scoped tasks see the largest and most consistent effect, the same upward compression of the skill distribution the customer-support evidence shows, because the model hands less-experienced staff patterns their seniors already carry.

The on-premises advantage here is concrete rather than aspirational. Tools such as Continue.dev run production code assistance against self-hosted open-weight backends, indexing the codebase and serving context-aware completions entirely on local infrastructure, so proprietary code never crosses the network [8]. For an organization whose codebase is itself a competitive asset, that is the difference between using AI on the work and not using it at all.

Software engineering is one of the four functions McKinsey Global Institute placed at roughly 75% of the projected $2.6 to $4.4 trillion in annual generative-AI value across enterprise use cases [9]. The macro estimate and the developer-level experiments point the same direction.

Software Testing and QA

Test generation is the highest-return, lowest-risk AI application in the software lifecycle. Generating syntactically correct unit tests against existing code does not require the model to understand business intent, and the validation loop is immediate: the tests run and pass, or they do not.

At ICSE 2023, CodaMosa showed that large-language-model-augmented test generation escapes the coverage plateaus search-based tools cannot pass. Where a search fails on methods whose state spaces need specific input combinations it cannot discover, a model primed on the code’s semantics proposes those combinations directly [10]. Mehmood et al. found Copilot effective for unit-test generation and boilerplate scaffolding on well-documented methods with clear specifications [11]. Yuan et al. mapped the far boundary, where ChatGPT produces reliable tests on clean function signatures and degrades on focal methods with intricate logic [12].

A claim that circulates, that AI lifts coverage from 40% to 70%-plus on existing codebases, is not traceable to a single peer-reviewed study. What the literature supports is narrower and still useful: model-augmented generation produces meaningful coverage gains beyond automated testing alone, concentrated on the methods where traditional tools plateau. Expect gains; expect their size to depend on how much of a given codebase is well-scoped enough for reliable generation.

The adjacent applications share the profile: suggesting edge and boundary cases, synthesizing test data for sensitive domains without exposing real records, expanding regression suites, and generating adversarial inputs for fuzz and security testing. Each is generation-heavy with fast validation; none asks the model to judge business intent. The management read is direct: expanding coverage without proportional headcount is a quality and risk lever. On owned infrastructure, the proprietary code, test data, and security findings stay internal.

Technical Documentation

Documentation is chronically under-resourced because it is high-effort, low-visibility, and detached from delivery metrics. AI removes most of the friction that produces that pattern.

The cost is documented at scale. The Consortium for Information & Software Quality (CISQ) put accumulated US software technical debt at roughly $1.52 trillion as of 2022, with documentation debt a structural component of it [13]. McKinsey separately estimates technical debt at about 40% of IT balance sheets, with around 30% of CIOs reporting that more than a fifth of nominally new-product budget is diverted to servicing it [14]. Documentation debt is not a line item in either source. It surfaces through the costs they do measure: longer onboarding, errors propagating from undocumented systems, audit overhead, and the difficulty of changing legacy code without breaking it.

The deployable applications are practical and validatable, with humans in the editorial loop rather than the generative one: API documentation drawn from code and comments, draft architecture-decision records from ticket threads, runbooks from incident logs, system documentation derived from infrastructure-as-code, and inline comments.

The higher-leverage use of the same capability is institutional-knowledge capture. An AI system can interview subject-matter experts through structured prompting, synthesize the responses into reference documentation, and index it for retrieval. The output is not documentation in the ordinary sense; it is queryable institutional context that grows with each interview and each project. The case for funding documentation as a discipline has always been sound and routinely deprioritized. AI shifts the economics enough that deprioritizing it is no longer rational.

Project Management and Executive Decision Support

AI does not replace judgment. It compresses the time needed to assemble the inputs judgment depends on.

McKinsey Global Institute finds current generative AI could automate work activities absorbing 60% to 70% of employee time, against a pre-generative estimate near 50% [9]. Knowledge workers spend a disproportionate share of that time synthesizing information. McKinsey’s product-management study quantifies the effect on a controlled population of forty product managers: gen-AI tools raised product-management productivity by 40% and pulled time-to-market in by 5%, with nearly twice the impact on content-heavy tasks, gathering and synthesizing information, drafting, and brainstorming, as on content-light ones [15]. Product management is the closest available analog to executive decision support, turning distributed information into decision-ready summaries.

The deployable applications are familiar: meeting summarization and action-item extraction, status reports assembled from ticket systems and standups, risk registers populated from project documentation, dependency analysis from unstructured notes, and stakeholder communication drafts. What they address is not decision quality but the time and effort to reach the point of deciding. Halving the path from “the data exists” to “a decision-ready summary” changes which decisions get made deliberately rather than by default.

Agentic infrastructure extends this from on-demand generation to continuous monitoring. An agent wired to project tools, document repositories, and communication channels surfaces exceptions and emerging risks, producing status without being asked; the compute and architecture this needs are covered in the agentic-scaling analysis. The step from “AI generates a report when prompted” to “AI surfaces what warrants attention unprompted” is a workflow step, not a capability step.

Institutional memory closes the loop. When a project begins, a system with access to past project records can surface relevant prior decisions, risks, and mitigations, the functional equivalent of consulting a program manager who has been at the organization for a decade. Morgan Stanley’s GPT-4 deployment, which gives advisers query access to a corpus of roughly 100,000 internal research documents, is the named public instance of this architecture; it is in production rather than theoretical [16], [17].

McKinsey’s 2025 Superagency survey notes that 92% of companies plan to raise AI investment while only 1% call their deployment mature [18]. In this category, the gap between aspiration and realized value is almost entirely workflow design.

Troubleshooting: Software and Hardware Systems

Troubleshooting is costly because it is unpredictable, expert-dependent, and largely undocumentable in advance. Each session resolves a problem and generates training signal that organizations almost never capture. AI changes both halves of that economics.

Gartner’s solution criteria for AI-for-IT-operations (AIOps) platforms name five capabilities: cross-domain event ingestion, topology generation, event correlation, incident identification, and remediation augmentation [19]. Gartner’s strategic framing is unequivocal, that “there is no future of IT Operations that does not include AIOps” [20]. In March 2025, Gartner renamed the category from AIOps platforms to Event Intelligence Solutions, shifting emphasis from the AI label to the use case [21]. Outcomes reported by adopters cluster around a 40% reduction in mean time to resolution, though the figure varies widely by environment and instrumentation and should be read as directional rather than a guaranteed result. The workload runs on local infrastructure for the reasons this paper presses throughout: logs and telemetry are sensitive, and the integration surface is internal.

For software, the deployable applications are log analysis, stack-trace interpretation, error-message explanation, root-cause hypothesis generation, and fix suggestion, integrated into existing observability stacks such as Grafana, Prometheus, and OpenSearch through agentic pipelines. An agent that watches logs and surfaces anomalies with plain-language explanations is operational today, not a research concept. For hardware, the value is symptom-to-cause mapping for known failure modes, cross-referencing vendor documentation against historical incident logs, and proposing investigation paths. Environments with tangled interdependencies across network, storage, compute, power, and thermal produce diagnostic problems where the bottleneck is correlation across more data sources than a human can read in time. The machine reads them all.

The institutional-knowledge angle is the most operationally direct in this section. Each logged and indexed troubleshooting session becomes training signal; the system learns the organization’s specific failure patterns, and later diagnoses are faster because the corpus is richer. A static runbook is frozen at the moment it was written and decays from the next day. This does not.

Engineering Design

For organizations doing mechanical, electrical, systems, or software-architecture design, AI helps at several stages, though the published evidence is younger and thinner than in software development.

Concept generation is the most mature application in the peer-reviewed record. At ASME’s design-engineering conference, Chong et al. showed that computer-aided-design (CAD)-prompted generative models produce designs that are more feasible and more novel than text-only generation, addressing the central weakness of earlier approaches, whose output needed heavy manual validation before anything usable could be extracted [22]. Llopis-Albert et al., in The International Journal of Advanced Manufacturing Technology, present a framework for model-assisted CAD that uses GPT-4o to generate and manipulate 3D models from text, blueprints, images, and voice [23]. The data-scarcity problem that blocks on-premises adoption is the subject of separate 2025 work: a generative-design method using Low-Rank Adaptation (LoRA) fine-tuning cut the training data required from over 16,600 samples to roughly 200 while holding generation correctness near 90%, a volume small enough to be feasible on a proprietary design corpus [24].

Specification drafting and standards-compliance checking are deployable now but thinner in the productivity literature. Given design intent in natural language, AI drafts structured specifications, flags missing requirements, and surfaces internal inconsistencies; with the relevant standards available through retrieval-augmented generation, it checks a draft against applicable codes and flags likely noncompliance, cutting expert review time on the routine majority of compliance work so reviewers concentrate on the consequential remainder. The capability is operational. The before-and-after productivity studies are not yet published.

Design-review preparation fits the same profile: synthesize a design package into a structured brief, surface open questions, and generate review criteria, with the human in the editorial role.

The frontier is moving quickly. Multimodal models now interpret diagrams and images in production forms that did not exist eighteen months ago with generative design maturing from novelty toward usable output. The honest characterization for engineering design is high ceiling, real workflow-integration cost, and an evidence base that still lags the demonstrated capability.

Scientific and Technical Research Assistance

In specific scientific domains the line between “AI accelerates research” and “AI conducts research” is genuinely blurring as of 2025 to 2026. For the use cases this paper concerns, the accurate framing is the former: AI sharply accelerates the research process for human researchers. The latter holds only in narrow domains, autonomous chemistry being the canonical case, and should not be generalized.

Literature-review acceleration has the most concrete evidence. Van Dinter et al., in Springer’s Computing, document that AI-assisted screening in systematic reviews can surface the large majority of relevant primary studies while reviewers examine a small fraction of the candidate pool [25]. The screening step that consumes weeks of expert time compresses to a fraction of that without losing the studies that matter.

Hypothesis generation, experimental-design assistance, and data-analysis narration are operational with less quantitative validation. Each addresses a real constraint. Hypothesis generation gives researchers a structured counterpart that enumerates plausible explanations and finds gaps. Design assistance flags confounders and statistical-power considerations, which is most valuable to teams without a dedicated statistician. Narration describes a dataset’s structure, proposes analyses, generates code, and renders results for non-specialist stakeholders.

Technical writing, drafting sections of reports, proposals, and patent applications from structured notes, compresses the path from “experiment complete” to “results documented” substantially. The published productivity numbers are not available at the granularity software development has, but the task profile is exactly the one where AI works.

The frontier is genuinely the frontier. A March 2026 Nature editorial documented the spread of AI research assistants from Google, OpenAI, and Anthropic, and noted that one system, Sakana AI’s The AI Scientist, produced a paper that cleared first-round peer review at a machine-learning workshop, with human curators filtering the most promising outputs before submission [26], [27]. The same editorial calls current outputs “limited and rarely innovative so far,” which is the calibrated read [26]. In autonomous experimentation, Berkeley’s A-Lab ran for seventeen days and synthesized 36 compounds from 57 targeted, computationally predicted inorganic materials with minimal human intervention, a documented instance of autonomous discovery, though such systems’ automated novelty and characterization claims have drawn scrutiny that underscores the continuing need for human verification [28]. For organizations deploying AI in technical research over the next two years, the value is in accelerating human-driven research; autonomous research will arrive in specific domains before it generalizes.

Lessons Learned and Institutional Knowledge Capture

This is the section’s highest-strategic-value application for executives, where the on-premises advantage aligns most directly with the organization’s interest.

Most engineering organizations hold extensive institutional knowledge only in the heads of senior staff. When those people leave, through retirement, resignation, reassignment, or restructuring, the knowledge goes with them. Ramp-to-productivity for complex systems runs from a six-month floor to twelve to fifteen months on the evidence already cited, and a third to two-thirds of new hires are lost inside the first year [3]. The knowledge base is the asset; storage is an implementation detail handled in the storage-architecture discussion earlier in this paper.

The deployable architecture is concrete. A maintained internal knowledge base, fed by past project documentation, incident reports, design reviews, meeting notes, code repositories, and structured expert interviews, becomes queryable institutional memory through retrieval-augmented generation methods: new members ask questions and get answers grounded in the organization’s actual history. Morgan Stanley’s GPT-4 system, giving tens of thousands of advisers access to roughly 100,000 internal research documents, is the named public case [16], [17]. McKinsey’s enterprise framing of generative AI includes exactly this corporate-brain function, answering questions across organizational knowledge with answer quality bounded by both the question and the system’s access to the relevant data [9].

Lessons-learned documents illustrate the difference between a passive archive and an active asset. Most are filed and never reopened. A system that indexes them and surfaces the relevant one when a similar project begins converts the archive into an advisory resource; the lesson appears when it is needed, without anyone having to remember it exists, find it, and read it.

The compounding is structural. Unlike tacit knowledge that departs with its holder, the indexed base grows cumulatively, becoming more valuable as the organization changes rather than less. MIT Sloan and BCG’s 2024 research finds that organizations pairing AI learning with organizational learning are better placed to manage uncertainty than those treating AI as a one-off tool deployment [29]. The framing for executives follows. This is strategic capital, closer to how technology firms treat proprietary databases and codebases as balance-sheet assets than to an IT line item. The infrastructure depreciates on schedule; the knowledge base it indexes appreciates.

Realistic Expectations

Not every use case here delivers equally; implementation quality dominates. Software development, testing, and AIOps have the densest peer-reviewed and observational productivity data. Documentation, project management, and institutional-knowledge capture have well-documented enterprise deployments and clear architectural patterns, with outcomes that depend more on organizational implementation than on raw capability. Engineering design and scientific-research assistance carry the highest ceilings and the heaviest workflow-integration cost.

The productivity-paradox analysis establishes the governing pattern: the difference between organizations that see measurable AI gains and those that do not is almost always workflow integration, training, and expectation-setting, not model capability. Every domain here is subject to it. The Faros telemetry on review burden swallowing coding speedup is the cautionary case for leaving workflows unchanged around the AI step; the Brynjolfsson, Li, and Raymond customer-support deployment, with its 15% average and roughly 30% gain for less-experienced agents, is the affirmative case for redesigning them [30].

The on-premises advantage compounds these effects rather than substituting for them. A team that owns its infrastructure can iterate on workflows, fine-tune on domain data, and build integrations against internal systems, none of which is possible against a locked vendor product. The ceiling on a fixed SaaS product is the one the vendor chose to ship; the ceiling on an organization-owned system is the one the organization is willing to invest to reach. For the use cases in this section, those are not the same ceiling. The distance between them is what the financial case has to price.