Exec intake pipeline

Executive Intake → AI Adoption Roadmap

A 19-field executive form becomes a founder-ready roadmap through four stages. Every heuristic, knowledge base, rubric, and guidance document is shown at the point it is injected.

Intake fields

Processing stages

Knowledge docs

~1,900

Target output words

Pipeline flow

Stage 1

Intake Processing

19 fields → 4 org_context rows. Confidence scoring and gap analysis.

smb_intake_executive_service.py

Stage 2

Prompt Engine

A ~220-line system prompt — 11 tenets, 5 principles, 14 rules.

adoption_roadmap_agent_executive.txt

Stage 3

LLM Generation

Claude Sonnet 4.5, validation loop with up to 2 retries.

adoption_roadmap_agent_executive.py

Stage 4

Post-Processing

Confidence merge, phase extraction, source manifest, email notification.

adoption_roadmap_executive_service.py

Stage 1 — intake processing

19 fields across four sections

About your business (5): business_name, industry, employee_count, what_they_do, what_makes_different.
What’s working, what’s not (5): biggest_bottleneck, time_consuming_tasks, success_definition, top_priorities, business_impact.
How work gets done (4): founder_role, team_description, delegation_gap, decision_style.
Tools & guardrails (5): essential_tools, document_storage, compliance_requirements, investment_comfort, non_negotiables.

Four org_context rows + confidence

company_profile — business snapshot. strategic_goals — goals & ambition. smb_operations — how work gets done. smb_constraints — constraints & guardrails.

Confidence engine: four sections × 25 points = 100 max (capped at 95). 19 field-level completeness checks. Field fill ratio → section score → total. Level: high ≥75, medium ≥50, low <50.

Gap analysis

Fires for every missing field — 19 field-specific impact statements plus 19 cohort exercises.

Plain-language impact

Each missing field gets a specific statement of what it affects. Example: “Without the founder’s role, the roadmap cannot identify delegation opportunities.” All 19 are hand-authored, not templated.

Cohort exercise

Each missing field maps to a concrete, week-numbered exercise. Example for delegation gaps: “Week 2: for one week, every time you do something someone else could handle, write it down.”

Confidence summary

Three-tier plain-language summary. High: “the roadmap is well-grounded.” Medium: “direction is sound but recommendations may shift.” Low: “a starting point, not a finished plan.”

Stage 2 — the prompt engine

Core IP

A ~220-line system prompt containing embedded knowledge frameworks, design principles, an output schema, and generation rules. Not a generic “write a roadmap” prompt — it encodes a complete methodology for honest, evidence-grounded AI adoption guidance.

Embedded knowledge frameworks

AI maturity model (5 × 5)

Five levels — Aware → Ready → Capable → Mature → Native. Five dimensions — Information Readiness, Process Clarity, Decision Governance, Team Capability, Tool Foundation. Bottleneck principle: overall maturity equals the lowest dimension score — prevents inflating the assessment by cherry-picking the strongest dimension. Backed by ai-maturity-model.md.

The eleven tenets (AESOP methodology)

1 State Change · 2 Problems Before Solutions · 3 Evidence Over Eloquence · 4 Know What You Need · 5 People Are the Center · 6 Humans Decide · 7 Multiple Perspectives · 8 Context and Brevity · 9 Guardrails Not Gates · 10 Trace the Connections · 11 The Questions Stay the Same.

Five non-negotiable design principles

CLEAR

Skimmable at 6am. Short words, short sentences — under 20 words per sentence. Business language, not academic.

CONCISE

Five pages maximum. Every paragraph earns its place. Substantial and detailed, never padded.

ACTIONABLE

Every section points to something they can do. No passive observation without a practical implication.

NO JARGON

Banned: LLM, API, vector database, embedding, token, agent, platform, ecosystem, solution. Write for a business founder.

NO ASSUMPTIONS — the hardest rule

Only say what the intake data supports. If they didn’t tell you about their customers, don’t describe their customers. Thin data → shorter, more tentative output. Before every claim: “did they tell us this?”

Writing discipline — banned words

significant · particularly · exceptionally · demonstrates · comprehensive · sophisticated · leverage · synergy · paradigm · automation · solution · platform · ecosystem · integration · implementation

Output validation schema — the rubric

maturity_snapshot

current_level + current_level_number. bottleneck_dimension + bottleneck_why (2–3 sentences) + bottleneck_what_this_means (2–3 sentences). Four stage_gaps (1→2, 2→3, 3→4, 4→5), each with whats_missing + what_it_unlocks. Three confidence_gaps, each with gap name, why_it_matters, how_to_close.

movements[] + insights

Three movements — name, rough_timeframe, focus, why_first — each with 4–5 practical guidance bullets, 2–3 what_to_avoid anti-patterns, and one observable risk_signal. Plus the_honest_truth (2–3 uncomfortable sentences), the_one_move (single highest-leverage 30-day action), and what_changes (3–4 tangible outcomes).

Fourteen generation rules — self-enforced

Quality constraints

1. Target ~1,900 words (1,600–2,200).
2. Maturity snapshot gets real word count — each gap 2–3 sentences, not bullets.
3. Exactly 3 movements with detailed guidance.
4. Exactly 3 confidence gaps ranked by roadmap impact.
5. No assumptions: 80%+ of claims traceable to intake fields.

Voice constraints

6. Mirror without announcing — never say “you told us” or “from your intake.”
7. The honest truth: direct, not cruel, grounded in their specifics.
8. What changes: [Person] can [do thing] without [current bottleneck].
9. What to avoid: concrete anti-patterns, never generic advice.
10. Risk signals: observable warnings, not abstract risks.

Guardrails

11. Banned: specific products, AI model names, internal scaffolding language.
12. Thin intake (<12 of 19 fields) → shorter, more tentative output.
13. Self-check before output: count claims, scan for jargon, check sentence length, verify sections present.
14. why_first never opens with “why this phase comes” — substantive content first.

Stage 3 — LLM generation + validation

Model configuration

Model: Claude Sonnet 4.5 (Anthropic API, streaming). Temperature: 0.2 — factual, consistent. Max tokens: 32,000. Provider: configurable — Anthropic or DeepSeek. Agent: AdoptionRoadmapAgentExecutive.

User message construction

Builds from the four org_context sections (content_summary, falling back to full content), the confidence assessment (score, level, known gaps). If total characters exceed 30,000, warns about the model input budget. Appends: “Produce an executive AI adoption roadmap as JSON. See the system prompt for the exact output format and generation rules.”

Structured validation loop — up to 2 retries

Attempt 1

LLM generates the full roadmap JSON.

Parse

Extract JSON — handles markdown wrapping, regex fallback.

Validate

20+ schema rules: maturity fields, 4 gaps, 3 confidence gaps, 3 movements, honest truth, one move, what changes (3+), word-count range.

Failed?

A specific violation list returns to the LLM with a retry prompt.

Retry ×2 max

Full multi-turn. Each retry ~$0.02–0.04. Warnings attached if still failing after two retries.

Pass

Validated JSON proceeds to post-processing.

Stage 4 — post-processing & delivery

Phase extraction

Movements → structured phase objects with phase_number, name, timeframe, focus, why_first, guidance (preserved as a bullet list), what_to_avoid, risk_signal. Timeline computed from first/last movement timeframes.

Confidence merging

Heuristic score (field completeness) adjusted downward by LLM-identified gaps — each gap deducts 5 points (min 5, max 40). Produces base_score + adjusted_score + adjusted_level + a merged summary.

Executive insights extraction

the_one_move, what_changes[], the_honest_truth, stage_gaps, confidence_gaps → an executive_insights JSONB column. Separates v3.0 fields from legacy columns for forward compatibility.

Source manifest — full provenance

org_contextdocument_count + document_types + provenance (“org_context table, is_active=true, all types”).

confidencescore, level, gap_count, provenance (“compute_confidence_executive() — field completeness from 19-field executive intake”).

agentname, temperature, prompt_file, max_tokens, input_tokens_used, output_tokens_used, cost.

validationwarnings list + warning_count.

What the founder receives

Maturity snapshot

Current level + bottleneck + four stage-gap maps.

Three movements

Practical phases with guidance, anti-patterns, risk signals.

Honest truth

2–3 uncomfortable sentences grounded in their data — plus the one move for the next 30 days.

Knowledge base

Eleven reference documents in backend/engine/knowledge/ provide the methodological foundation. They exist independently as versioned, auditable assets — updatable without retraining or rebuilding the pipeline.

aesop-methodology.md

The eleven tenets, autonomy classification (4 levels), build summary artifacts (5 types).

ai-maturity-model.md

5 levels × 5 dimensions, bottleneck principle, full scoring protocol with evidence requirements.

prd-decision-framework.md

Assistant vs agent classification, 8-dimension scoring, decision tree, hard gates.

feasibility-rules.md

Platform connector registry, constraint rules, trigger compatibility matrix, file-format support.

hybrid-architecture-patterns.md

9 patterns for combining deterministic + non-deterministic logic, and anti-patterns.

discovery/triage-questions.md

7 triage questions, red flags, complexity signals, routing rules.

discovery/capability-thresholds.md

7 capability domains with quantitative breaking points, migration paths, cost implications.

discovery/project-failures.md

7 common failure patterns with early warning signs and red-flag signals.

discovery/trade-off-frameworks.md

Decision frameworks for scope, quality, and speed trade-offs in AI adoption.

discovery/question-banks.md

Structured question banks for discovery interviews across domains.

discovery/platform-capabilities.md

Platform-specific capability matrices and constraints for build targeting.

Why the output is quality

Multi-layered confidence

Heuristic field-completeness score, adjusted by LLM-identified gaps, merged with provenance. No single source of truth about quality.

Bottleneck principle

Overall maturity equals the weakest dimension. Prevents inflating the assessment by cherry-picking. Backed by an explicit scoring protocol in ai-maturity-model.md.

Structured validation with retries

20+ schema rules checked programmatically. Up to two automatic retries with specific violation feedback; warnings attached if still failing.

Evidence traceability

80%+ of claims must trace to specific intake fields, enforced in the prompt. “Did they tell us this?” before every claim.

No-jargon enforcement

15+ banned words keep out generic consulting language. Written for someone who runs a business, not a technology department.

Mirror without announcing

References their specifics naturally, never “you told us” or “from your intake.” They feel seen, not analyzed.

Compass, not GPS

Direction with enough texture for Monday morning, but no false precision. Three movements, not twelve steps. Risk signals, not risk matrices.

Gap → impact → exercise chain

Every missing field produces a plain-language impact and a concrete cohort exercise with a week number. Actionable, not diagnostic.

Full source provenance

The source manifest records every input: org_context types and counts, confidence provenance, agent config, token usage, cost, validation warnings.

Eleven Tenets foundation

Philosophical rigor from aesop-methodology.md — People Are the Center, Evidence Over Eloquence, Guardrails Not Gates, Trace the Connections — so every output embodies a coherent methodology, not ad-hoc heuristics.

Architecture note. This is the executive path — a lightweight variant of the full AESOP Build Pipeline. The full pipeline adds readiness assessment, portfolio with strategic prioritization, department context, multi-agent quality analysis, interviews, custom rubric + scenario generation, adversarial red-team, functional testing, and synthesis reporting. The executive path exists for founders who need a focused, honest, actionable roadmap — not an enterprise-scale analysis.