{ "schemaVersion": "1.0", "language": "en", "canonicalUrl": "https://catalysphere.co.jp/en/articles/ai-adoption-operating-model/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/ai-adoption-operating-model/index.json", "publisher": { "name": "Catalysphere", "url": "https://catalysphere.co.jp/", "description": "A research and consulting studio connecting AI and big-data research with practice." }, "article": { "slug": "ai-adoption-operating-model", "title": "What Organizations Using AI Are Standardizing", "summary": "Distributing tools alone does not build organizational capability. Drawing on field research, this article outlines how to redesign work, learning, expertise, and control.", "takeaways": [ "AI adoption is not about increasing the number of users, but about leaving behind organizational mechanisms that allow results to be reproduced.", "Effects vary by proficiency and work structure; reducing individual time alone does not change the volume or composition of work.", "Standardize usage scenarios, inputs, quality criteria, reviewers, and stop conditions, and update them regularly based on on-the-ground learning." ], "publishedAt": "2026-07-12", "updatedAt": "2026-07-24", "readingMinutes": 8, "topics": [ "AI Adoption", "Organizational Design", "Change Management" ], "domains": [ "generative-ai", "organization-adoption" ], "evidenceKind": "research-review", "status": "published", "image": { "prompt": "Abstract scientific editorial illustration of organizational AI adoption: several knowledge clusters connected through a shared intelligent network, individual nodes becoming a coherent learning system; deep navy, muted blue and restrained gold; precise, calm, premium research-journal aesthetic; no text, no logo, no watermark.", "negativePrompt": "robots, humanoids, office stock photography, readable text, logos, neon cyberpunk, purple gradients, clutter", "alt": "Abstract diagram of multiple knowledge clusters connecting to an organizational network that learns", "aspectRatio": "16:9", "assetPath": "article-art/ai-adoption-operating-model.png", "status": "approved" }, "lead": "There is a gap between issuing AI accounts and changing how an organization works. Field research shows that effects vary depending on individual proficiency and work structure, and that AI influences knowledge transfer and the boundaries of expertise. Adoption should be designed not by usage rates, but as a learnable work system.", "visual": { "eyebrow": "EVIDENCE MAP / ADOPTION", "title": "From Distribution to Work Redesign.", "caption": "Implementation effects are not determined solely by the number of users. Observe work flow, learning, and connections to expertise.", "metrics": [ { "value": "66 companies", "label": "Randomized Experiment in Daily Work", "source": "Dillon et al., 2025", "referenceId": 3 }, { "value": "+36%", "label": "Resolved Cases for Low-Skill Groups", "source": "Brynjolfsson et al., 2025", "referenceId": 1 }, { "value": "791 participants", "label": "Individual/Team x AI Experiment", "source": "Dell’Acqua et al., 2026", "referenceId": 2 } ] }, "sections": [ { "heading": "Distribution Alone Does Not Change Work Structure", "paragraphs": [ "In a randomized experiment on generative AI involving 66 companies and 7,137 people, actual users reduced their email time by approximately two hours per week. However, providing AI to individuals alone did not produce statistically detectable changes in the volume or composition of work. [3] This is not a result indicating AI lacks value, but rather shows that time savings do not automatically translate into work redesign.", "Login rates and prompt counts may serve as leading indicators of adoption, but they are not outcomes. Which meetings, approvals, records, and customer interactions change, and where the saved time is reallocated, must be decided. Only by defining the entry and exit points of work does individual convenience become organizational capability.", "Implementation effects should be tracked in the order of usage, process changes, confirmed quality, and business outcomes. If usage increases and time decreases, but processing volume and quality remain unchanged, this may be due not only to insufficient AI capabilities but also to constraints in downstream processes or a lack of design for reallocation. [3] Link the entry and exit points of target work to the same records, and do not rush company-wide rollout until the owners of processes where changes have stalled hold improvement plans." ], "bullets": [ "Target Work: Define where AI is used and where it is not.", "Input: Prepare accessible data and permissions.", "Output: Determine quality criteria, reviewers, and record formats.", "Reallocation: Decide which outcomes the saved time will be shifted to." ] }, { "heading": "Viewing AI as a Device for Transferring Tacit Knowledge", "paragraphs": [ "In a phased implementation study of customer support, while the average effect of AI assistance was 15%, the lowest-skilled group before implementation saw a 36% increase in resolved cases per hour. For the top-tier group, there was no significant productivity gain, and evidence has been reported that some learning effects remained even after AI was stopped. [1] AI may function less as uniform automation and more as a device that conveys excellent response patterns to those with less experience.", "Therefore, education should not end with general prompt training. Embed what information experts review, where they make judgments, and what they treat as exceptions into AI proposals, reviews, and feedback. Share the work conditions under which results were reproduced as shared assets, rather than just the prompts that produced results.", "Conditions adopted for knowledge transfer must ensure that not only do the results of less experienced people improve, but that judgment procedures remain even when AI is stopped or for other cases. Although some learning effects are shown in research, there is no guarantee they will replicate across all job types. [1] From review records, separate verbatim copying of proposals, understanding of reasoning, and exception judgment; if the latter two do not improve, revise the education design." ] }, { "heading": "Redefining the Boundaries of Expertise", "paragraphs": [ "In a randomized experiment involving 791 people at Procter & Gamble, combining individual and two-person teams with the presence or absence of AI showed that individuals using AI achieved results equivalent to teams not using AI. Additionally, the bias toward R&D or commercial specialties in AI users' proposals decreased. [2] AI may change not only work speed but also who can handle which specialized knowledge.", "However, this does not conclude that experts are unnecessary. These results were observed in short-term product development tasks; long-term collaboration, implementation responsibility, and organization-specific tacit knowledge require separate verification. We test whether expert roles can shift from creators to designers of questions, evaluators, and exception judges.", "Experiments that change the boundaries of expertise measure not only outcomes but also who bears final responsibility, who detects exceptions, and who hands off to implementation. Even if AI-using individuals match team performance in short-term tasks, this does not mean they have replaced long-term collaboration or implementation responsibility. [2] If only the waiting time for expert reviews increases, treat it as work transfer rather than role redesign." ] }, { "heading": "Measuring Exception Liability as Standard Maturity", "paragraphs": [ "Studies showing significant improvements in less experienced groups indicate that the value of the same support varies depending on user conditions. [1] However, these studies do not measure an organization's volume of exception handling. Here, we define 'exception liability' as the state where cases deviating from standard procedures remain unresolved and accumulate, using it as an analytical concept to evaluate AI adoption from the perspective of operational feasibility.", "Exceptions are not just failures. They include cases held due to insufficient input, returned to experts, nearly used for prohibited purposes, or failed to reach on-the-ground results despite following standards. If usage rates rise but exception dwell time and concentration on the same personnel increase, this operational hypothesis suggests that standards are not absorbing on-the-ground differences. Track not just the number of cases, but the reasons for closure and recurrence.", "Recommended operations manage exceptions not by individual ingenuity, but by status: received, provisional measures, owner, resolved, and reflected in standards. If the same exception recurs, leave competing hypotheses such as insufficient training, input conditions, role distribution, and application boundaries. For businesses where liability grows beyond resolution capacity, prioritize standard repair over adding users." ] }, { "heading": "Managing Operational Standards as Time-Bound Hypotheses", "paragraphs": [ "Distributing standards as fixed answers leaves procedures in place even when premises change. Therefore, describe each standard as a time-bound hypothesis: 'Given these inputs, roles, and confirmation conditions, work closes with this quality.' Standards versions should include not only application conditions but also falsifying failures, next review dates, and post-abandonment transitions. This is not a finding from existing research, but a recommendation to prevent halting organizational learning.", "Consider a hypothetical example. When transferring a summarization procedure replicated in one department to another, do not simply copy common parts; make transfer conditions include whether reference permissions, document update frequencies, and final reviewers are the same. Branch parts with different conditions as regional versions, and decide after a certain period whether to elevate them to common standards or leave them as local procedures. Focus not on the number of branches, but on the ability to explain the differences.", "Recommended standard reviews confirm the percentage of cases closed as hypothesized, exception recurrence, knowledge returned from local to common versions, and the abolition of unused procedures. Do not extend the life of standards based solely on success cases; treat non-use reasons as falsifications. If owners are absent, verification deadlines have passed, or alternative procedures are superior, remove them from valid standards even if they are referenceable." ], "bullets": [ "Hypothesis: State in one sentence the input, role, confirmation, and exit conditions expected to hold.", "Falsification: Define signals to review standards, such as quality failures, exception recurrence, and non-use.", "Version and Deadline: Record owners, verification period, next review date, and local differences.", "Exit: Record judgments and transitions for commonization, local maintenance, replacement, or abolition." ] }, { "heading": "Standardizing Not Answers, But Boundary Conditions", "paragraphs": [ "Experimental results showing increased outcomes within capability ranges and decreased accuracy outside them indicate that AI application boundaries are uneven within work. [4] Organizational standards should center not on universal prompts, but on available data, quality criteria, prohibitions, escalation paths, and stop conditions.", "On the other hand, leave the formulation of questions and room for trial to the field. Record successes and failures in the same format, and have management, operations, data, IT, and legal/risk periodically update boundaries. Adoption is not about increasing the number of AI users, but creating a state where the organization can learn when to use, doubt, and stop AI.", "Operational meetings should be structured so that business owners own quality and application boundaries, data owners own available usage scopes, risk owners own stop conditions, and technical owners own versions and recovery. For businesses where failures outside capability ranges increase, revert to approved procedures even if usage rates are high. [4] Resumption requires both improvement based on failure samples and the ability of the field to hold or stop execution. Judgment histories are carried over to the next boundary update." ], "bullets": [ "Commonize: Data, permissions, confirmation criteria, records, stop conditions.", "Leave to Field: Questions, trials, expression, improvement proposals.", "Continuous Observation: Quality, correction volume, learning speed, exceptions, non-use reasons.", "Regular Updates: Not just success cases, but failures and capability limits." ] } ], "note": "Research results include single-company/single-job-type studies and short-term experiments. Do not confuse usage rates with causal business outcomes; you must verify job types, proficiency, usage voluntariness, evaluation systems, and long-term learning within your own organization.", "sources": [ { "id": 1, "authors": "Erik Brynjolfsson, Danielle Li, Lindsey R. Raymond", "year": 2025, "title": "Generative AI at Work", "venue": "The Quarterly Journal of Economics, 140(2), 889–942", "url": "https://doi.org/10.1093/qje/qjae044", "finding": "Average 15% productivity improvement in support work for 5,172 people. Effects were larger for lower-skilled groups, with evidence of learning reported.", "limitation": "Phased implementation in a single company and job type, targeting stable products and inquiry environments." }, { "id": 2, "authors": "Fabrizio Dell’Acqua et al.", "year": 2026, "title": "The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork", "venue": "Organization Science, Articles in Advance, published June 12, 2026", "url": "https://doi.org/10.1287/orsc.2025.20702", "finding": "Randomized 791 people at P&G; individuals using AI achieved results equivalent to teams without AI, and bias between specialties also decreased.", "limitation": "Short-term product development tasks at one company; long-term team relationships and implementation outcomes were not measured. P&G employees were among the authors, and Harvard Business School provided partial funding." }, { "id": 3, "authors": "Eleanor W. Dillon, Sonia Jaffe, Nicole Immorlica, Christopher T. Stanton", "year": 2025, "title": "Shifting Work Patterns with Generative AI", "venue": "NBER Working Paper 33795, revised November 2025", "url": "https://doi.org/10.3386/w33795", "finding": "In an experiment with 66 companies and 7,137 people, actual users' email time decreased by about 2 hours per week, but changes in work volume/composition were not detected.", "limitation": "Working paper; quality and sales were not directly measured." }, { "id": 4, "authors": "Fabrizio Dell’Acqua et al.", "year": 2026, "title": "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality", "venue": "Organization Science, 37(2), 403–423", "url": "https://doi.org/10.1287/orsc.2025.21838", "finding": "Outcomes improved for tasks within AI capabilities, while accuracy dropped by an average of 19 percentage points for tasks outside capabilities.", "limitation": "Targeted high-skilled consultants; only one type of task outside capabilities was tested." } ], "url": "https://catalysphere.co.jp/en/articles/ai-adoption-operating-model/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/ai-adoption-operating-model/index.json", "topicFacets": [ { "id": "organization-adoption", "name": "Organization & Adoption", "url": "https://catalysphere.co.jp/en/topics/organization-adoption/" } ] } }