Record the production boundary.
Result, baseline, version, status, owner, users, sources, actions, approvals, prohibited behavior, providers, service limits, fallback and current operating measures.
Open research protocol · no findings yet
Study the conditions that make enterprise AI operational.
Compsia is publishing the proposed method before reporting results. Fieldwork findings, prevalence estimates, benchmarks, performance claims and ROI conclusions are not yet available and must not be inferred from this protocol.
Research question01
What separates AI activity from an accepted production capability?
The study is designed to examine how organizations move a defined AI-supported business result from pilot or local use into an owned operating path. It will focus on the relationship among business consequence, production perimeter, data and action authority, acceptance, user responsibility, continuing operation and measurement.
The unit of analysis is one named AI production system or candidate system—not a company's general sentiment toward AI. A system must have a stated business result, users, sources or context, expected outputs or actions and an identifiable operational owner to enter the primary analysis.
Population and instrument02
Collect evidence at the level where responsibility exists.
The proposed population includes enterprise business owners, technology or data owners, security or risk participants, system operators and representative users involved in a specific system. Recruitment source, sector, geography, organization size, role and relationship to Compsia will be reported with the eventual sample.
Trace decisions and failure modes.
Ask who can authorize, accept, intervene, support, change and stop the system; what evidence informs those decisions; and where formal ownership differs from actual work.
Inspect available records.
Where permission allows, review diagrams, test matrices, release records, operating dashboards, incident records, change logs and user guidance rather than relying only on recollection.
Record the business measure carefully.
Capture the baseline, observation window, other changes and measurement limitations. An observed movement will not be attributed to AI without a defensible design.
Analysis and limitations03
Separate description, association and causal claims.
Qualitative material will be coded against a published codebook covering result ownership, boundary definition, technical composition, authority, acceptance, adoption, operation, evidence and stop or expansion logic. A second review will be used for a documented subset before themes are finalized.
Quantitative summaries, if sample quality permits, will report denominators, missingness and uncertainty. Small or convenience samples will be described as such. No percentage will be generalized to “enterprises” without a population and sampling design that supports that inference.
Customer, prospect or partner participation can create selection and relationship bias. Compsia's commercial interest will be disclosed. Findings that depend only on Compsia-delivered systems will be labelled separately from evidence collected from independently built systems.
The protocol does not prove that Compsia's method causes better outcomes. It creates a transparent structure for collecting and challenging evidence. External review, anonymization constraints, instrument revisions and deviations from this protocol will be recorded in the eventual release.
Release gate04
No finding is public until its evidence can be audited.
A report may be released only after the final sample, recruitment method, instrument, codebook, analysis decisions, exclusions, conflicts, limitations and evidence table are documented. Publication also requires confidentiality review and explicit permission for any identifiable company, person, quote or result.
Freeze the method record
Version the protocol, instrument, codebook, sample and deviations; retain an auditable claim-to-evidence table.
Label every evidence class
Distinguish participant report, reviewed artifact, observed system behavior, analyst inference and external source.
Publish corrections visibly
Maintain version, date, correction history and the scope within which a finding may be cited.