MBT

Governance

AI Adoption & Research Governance

The 8 Governance Disciplines of MBT — how peer-reviewed science, evidence law and operational safeguards govern research and AI adoption.

Discipline 1: Input Integrity & Institutional Scoping

Scientific lineage: Paul DiMaggio & Walter Powell (1983) — Institutional Theory (The Iron Cage Revisited)

Organisations within the same field face identical structural and regulatory pressures. Shared environmental pressures must be formally separated from company-specific conditions before any diagnostic begins.

  • Specialised multi-model firewalls. Models that search public internet evidence never see internal documents. Internal reasoning evaluates business context at temperature zero and never browses the open web.
  • Zero model training & enterprise privacy. Customer inputs, saved research sessions and uploaded documents are never used to train models and are never sold. Sensitive runs can use zero-retention mode.
  • Mandatory relevance verification gate. Every typed input or uploaded internal file is screened against the locked industry, geography and scope. Incompatible files are rejected; partially matching documents are downgraded to inference-level evidence so they cannot carry a rating alone.
  • Sector & geography-first research anchoring. External searches anchor on industry and jurisdiction first. The company's own name is used only to evaluate organisational exposure, so corporate PR cannot bias the research.

Discipline 2: Grounded Evidence & Inductive Discovery

Scientific lineage: Barney Glaser & Anselm Strauss (1967) — Grounded Theory (The Discovery of Grounded Theory)

Explanations must be built up inductively from primary observable evidence, not chosen first and rationalised after the fact.

  • Outside-in evidence collection. External industry, statutory and economic evidence is gathered and logged before internal organisational problems are framed.
  • Standardised epistemic tiers. Every finding is labelled on screen as Fact, Evidence, Inference or Hypothesis.
  • Origin-based source independence. Corroboration counts distinct publisher origins, not syndicated copies. Commercial interests (vendor-authored, sponsored content, analyst-paid) are flagged so marketing claims cannot pass as independent proof.
  • Mandatory contested evidence disclosure. When internal documents conflict with public evidence, the clash is disclosed, never silently resolved: "Workspace document states X; public sources suggested Y; internal data used per policy."

Discipline 3: Causal Depth & Directed Structural Reasoning

Scientific lineage: Judea Pearl (2000) — Causal Directed Acyclic Graphs (Causality: Models, Reasoning, and Inference)

Cause and effect must run strictly one way. A reasoning chain that loops back on itself explains nothing.

  • The 7-level causal hierarchy. Root Cause → Problem → Challenge → Complexity → Scenario → Symptom → Business Impact.
  • No symptom-level AI band-aids. Prevents buying software to treat surface pain (turnover, slow ticket resolution) when the true failure sits upstream in role design or policy.
  • Zero-temperature determinism. Reasoning runs at temperature zero, so the same inputs produce the same findings.
  • Provenance tracking. Every hierarchy cell carries its origin — research, radar, internal documents or researcher entry — with a valid-as-of date and re-check horizon.

Discipline 4: Frontline Empirical Validation & Falsification

Scientific lineage: Karl Popper (1959) — Empirical Falsificationism (The Logic of Scientific Discovery)

A claim is never established by collecting confirming examples; it earns standing only by surviving deliberate attempts to disprove it.

  • Downward probing (Discover & Relate). The survey asks about concrete operational friction across five families: Occurrence, Frequency, Ownership, Friction and Impact.
  • Active falsifier probes. Deliberate test questions confirm whether an issue is genuinely absent or whether leadership is overestimating operational health.
  • Contradiction detection. Conflicting answers raise contradiction notices that must be reconciled.
  • Evidence-derived governing mandates. Survey patterns set the directive for each objective:
    • Confirmed gap — "Establish…": the basic capability must be built before software is introduced.
    • Inconsistent practice — "Standardise and embed…": processes must be unified before automation can run reliably.
    • Insufficient evidence — "Clarify / retain as hypothesis…": internal checking comes first, so AI is not deployed into blind spots.

Discipline 5: Governed Capability & Bottleneck Framing

Scientific lineage: Eliyahu M. Goldratt (1984) — Theory of Constraints & Current Reality Trees (The Goal)

A system is limited by a small number of real constraints; treating every symptom equally wastes capital and effort on the wrong problems.

  • Theory of Change objective formula. Every objective is written as an operational capability: [Verb] [Capability] within [Scope], enabling [Execution] so that [Failure reduced] and [Impact improved], subject to [Dependencies].
  • Root-cause anchored capabilities. No AI capability, digital tool or transformation act is recommended unless it traces back through the causal hierarchy to an evidenced root cause.
  • Structured objective metadata. Every objective carries four governance fields:
    • Enables — the operational capability unlocked.
    • Protects — the organisational, financial or compliance failure prevented.
    • Subject to — prerequisites, dependencies and governance constraints.
    • Context — the workflow, workforce and geographical boundary.
  • Built-in SMART check. Each objective can be checked for being Specific, Measurable, Achievable, Relevant and Time-bound.

Discipline 6: Pre-Adoption Gatekeeping & Socio-Technical Balance

Scientific lineage: Eric Trist & Fred Emery (1951) — Socio-Technical Systems Theory (Tavistock Institute studies)

Work is a joint system of people and technology. Optimising technology without restructuring the social and organisational system reliably produces failure.

  • Parallel social and technical workstreams. Every target outcome is split into changes to organisation design, policy and governance, alongside what digital systems can carry.
  • Honest technology contribution rating. Technology's true share of the solution is rated None, Low, Medium or High. A low or nil rating is a valid and expected finding.
  • The "HR-alone" limit rule. Where a root cause sits in governance, funding, structure or leadership mandate, MBT states that HR-led change alone will fail and names the wider transformation as a prerequisite.
  • Prerequisite wave sequencing. Data definitions, field-level ownership and process harmonisation come first; AI is held to later waves so organisations never automate chaos.
  • Vendor neutrality. Recommendations name capabilities and categories only — no software brands.

Discipline 7: Multi-Dimensional Prioritisation & Decision Utility

Scientific lineage: Ralph Keeney & Howard Raiffa (1976) — Multi-Attribute Utility Theory (Decisions with Multiple Objectives)

Sound prioritisation evaluates multidimensional business value against the real cost and friction of change on a declared scale.

  • 7-factor balanced scoring (1–10). Business Impact, Urgency, Likelihood of Success, Evidence Strength, Technology Contribution, Implementation Effort and Future Relevance.
  • Impact × Effort law. Initiatives are plotted as impact against effort (cost, complexity, dependency, change load) — never against technology potential.
  • Quarterly horizons. Roadmaps are sequenced into delivery phases:
    • Q1 (0–90 days) — mobilisation, foundational data governance, policy containment.
    • Q2 (3–6 months) — controlled pilots and core execution.
    • H2 and beyond (6–12+ months) — scaled institutionalisation and advanced automation.
  • Rejection gating. An objective the researcher marks Rejected is excluded from the active portfolio.

Discipline 8: Operational AI Oversight & Executive Traceability

Scientific lineage: Igor Ansoff & Francis Aguilar (1975 / 1967) — Weak-Signal Detection & Environmental Scanning

Systemic disruptions announce themselves quietly at the periphery long before they become obvious at the centre.

  • Current Affairs Radar & status checks. User-run scans of regulatory, labour and statutory developments. Each item is re-checked for its status as of the run date, so superseded measures are not presented as current.
  • Agentic AI oversight checklist. Every automated scenario must define:
    • Human-in-the-loop checkpoint — where a named person reviews, approves or overrides.
    • Guardrails — compliance, privacy, fairness and ethical boundaries.
    • Prerequisites — data accuracy, connections and process stability before go-live.
    • Success signal — the measure used to detect drift and stop the agent.
  • Workforce sustainability safeguards. Flags job redesign, re-skilling needs and psychological-safety risks, with precautionary safeguards (bias audits, notice, appeal paths) for AI used in hiring and evaluation.
  • Human adoption gate & database-derived reports. AI never makes the final call: the researcher records a verdict (Validated, Investigate further, Rejected or Not sure). Dossiers, decks and specifications are generated from the saved study record, without softening root causes or technology limits.

Theorist names, theory titles and seminal works are cited for scholarly attribution and methodological lineage only, as permitted for nominative academic reference. The named authors, their estates and their affiliated institutions do not sponsor, certify, endorse or have any affiliation with MBT. All descriptions, wording and software are original works of MBT and are not authorised derivatives of the cited publications. To query any attribution, contact contact@hrmbt.com. Attribution terms