Preserve human expertise
before it walks out the door.
CoBuild is an agentic knowledge preservation system that captures what your best employees know and transforms it into performance support your team can actually use.
“Your best employee retires Friday. On Monday, their expertise is gone. CoBuild fixes that.”
Every organization has a Maria.
Maria is the senior finance lead. She owns the month-end close, knows which D365 configuration choices were made for reasons nobody wrote down, and has the cell number of the operations contact who gets things done. When Maria goes on medical leave, the SOPs cover the procedures. Everything else — the reasoning, the workarounds, the relationships, the judgment calls — walks out the door with her.
Owns period-end close, fixed asset validation, inter-company reconciliation, and a dozen non-default configuration choices whose rationale exists only in her memory.
Knows which scanner needs to be reset on humid mornings. Knows which vendor shipments run late and by how much. Twenty years of pattern recognition, zero documentation.
Remembers why three federal reports use a non-standard fiscal calendar. Has the relationship with the program officer who extended the deadline in 2021. Retiring in 90 days.
Knows the infrastructure decisions made in 2003 and why. Can read the system behavior that precedes a failure three hours before anyone else sees it. Nobody else on the team can.
The cost of losing senior employees is rarely calculated. It shows up in slower onboarding, repeated mistakes, projects that stall, and institutional memory that has to be rebuilt from scratch — every single time.
Seven agents. One job each.
One survival kit out.
CoBuild runs a sequential pipeline of AI agents, each mapped to a real position on a learning and development team. Every agent has one input, one output, and one handoff. The result: a complete performance support kit produced from a single SME interview.
Cleans, chunks, and tags every source document. Flags chunks where tacit knowledge is implied but not stated. Produces a structured content set every downstream agent can reason over.
- intake/cleaned/*.md
- intake/manifest.json
- intake/flags.md
Extracts every task the expert performs. Scores each on frequency, business criticality, and substitutability. Produces a risk-ranked priority list — the tasks the team genuinely cannot afford to lose.
- analysis/task-inventory.md
- analysis/priority-list.md
- analysis/task-graph.json
The credibility agent. Finds what isn’t there — the workarounds, the configuration rationale, the relationship norms that live in the expert’s head. Applies Gilbert’s BEM to route non-knowledge problems away from training artifacts.
- analysis/gap-map.md
- analysis/sme-question-list.md
- analysis/bem-flags.md
Applies four learning theory lenses — Behaviorism, Cognitivism, Social Cognitive Theory, Constructivism — to each true knowledge gap. Recommends the theory whose predictions best explain how transfer will succeed or fail. Agent 4 receives these as binding inputs.
- analysis/theory-recommendations.md
Selects the right artifact type per gap (job aid, decision tree, scenario, microlearning, reference card). Writes Bloom-leveled objectives and complete artifact specifications Agent 5 can execute without ambiguity.
- strategy/strategy-doc.md
- strategy/artifact-specs.md
- strategy/objectives.md
Produces the actual deliverables. Follows specifications exactly. Flags any content gap it cannot fill without SME input — never hallucinating missing account numbers, navigation paths, or checklist items.
- artifacts/job-aids/*.md
- artifacts/decision-trees/*.md
- artifacts/scenarios/*.md
- artifacts/index.md
Builds a measurement plan at all four Kirkpatrick levels. Overlays LTEM diagnostic rules — if transfer fails, which tier broke, and does that mean fix the artifact or fix the environment? Produces an exec summary in plain language for leadership.
- evaluation/measurement-plan.md
- evaluation/refresh-triggers.md
- evaluation/exec-summary.md
One 47-minute interview.
A complete survival kit.
The full pipeline was run against a synthetic SME interview transcript for “Maria Chen,” a D365 Finance lead going on medical leave. The transcript contained six deliberately planted tacit knowledge gaps. Every agent passed validation.
“Agent 5 produced 16 artifacts and 8 build flags. The flags identified compliance-risk content gaps, SME-dependent account mappings, and constructed scenarios requiring expert validation — demonstrating that CoBuild surfaces the boundaries of what AI can produce without human expertise, rather than filling gaps with plausible but unverified content.”
Agent 1 correctly flagged every deliberately hidden piece of tacit knowledge in the transcript — categorized as artifact, configuration, or relationship gaps.
The operations coordination gap, the fixed assets query location, and three others were correctly identified as environment/information problems — not knowledge gaps. No training artifacts were generated for them.
Agent 6’s executive summary contains no references to agents, pipelines, BEM, or Kirkpatrick levels. A controller or VP of Finance can open it and act on it immediately.
Real instructional design.
Not just generated content.
CoBuild integrates six established frameworks from learning science and human performance technology. Each is applied at a specific point in the pipeline — not as decoration, but as a design constraint that changes what the system produces.
Behavior Engineering Model. Distinguishes true knowledge gaps from environmental, incentive, and resource problems. Prevents CoBuild from generating training for problems training won’t solve.
Behaviorism, Cognitivism/CIP, Social Cognitive Theory, Constructivism. One theory is selected per gap based on which framework’s predictions best explain how transfer will succeed or fail for that specific knowledge type.
Applied as a completeness lens on artifact design — not as a mandatory checklist, but as a check that the content outline isn’t missing something a performer will need under real conditions.
Learning objectives are Bloom-leveled per artifact. A reference card targets Apply. A decision tree targets Analyze. A scenario targets Evaluate. The cognitive demand matches the real task.
Four-level measurement plan per artifact, specifying minimum viable measurement for organizations with limited bandwidth. L4 business outcomes tied to explicit baselines and success thresholds.
Learning-Transfer Evaluation Model. Diagnostic layer underneath Kirkpatrick: if transfer fails, which tier of the chain broke? Does that mean fix the artifact — or fix the environment?
The decision rule the expert applies but hasn’t articulated
The configuration choice whose rationale exists only in memory
The exception case the SOP doesn’t cover
The step-by-step procedure that needs to become automatic under pressure
The judgment call that requires practicing before you face it live
The query that exists in the system but nobody shared the location — get the artifact, don’t train on it
The operations contact who responds fast because they trust the expert — warm handoff, not a job aid
The system access the successor doesn’t have — provision it, training can’t substitute
The account numbers that need to be verified — get them from the SME, don’t approximate
The authority that was never formally transferred — management decision, not a training problem
Is your organization facing a knowledge continuity risk?
A planned retirement. A key role transition. A consulting engagement ending. An ERP go-live where the legacy experts are about to walk out the door. Blue Edgewater offers CoBuild as a consulting engagement.
High-compliance environments with deep institutional knowledge at risk from retirements and role transitions.
Consulting engagements end. Legacy experts leave. The knowledge transfer problem is built into the project lifecycle.
Decades of operational knowledge in the hands of retiring supervisors. No documentation captures what they actually know.