EGP CDAIO — Strategic vs Routine vs Gray Dashboard

Top-Line KPIs

Strategic

118

59.0% of files · $21.24 Claude spend

Routine

3.0% of files · $1.08 Claude spend

Gray Zone

38.0% of files · $13.68 Claude spend

Total Claude spend (2026-04-13): $36.00 · cost model $0.18/file blended Opus 4.6
Local LLM persona: Vesta (backed by llama3.2:3b primary, qwen2.5:3b fallback · Phase 1 not yet live)

Gray Zone — 10 Sub-Categories

Marc's directive: break Gray down granularly, flag which are SOP-able (for routing to Vesta) vs need-human-triage (stay on Claude). If a tool is being executed, that's an SOP opportunity.

Sub-Category	Count	Spend	Routing	SOP Note
1. Python multi-step glue scripts	2	$0.36	SOP-able	Pipeline orchestrators, subprocess chains — SOP them for Vesta execution
2. HTML/PDF renderers	1	$0.18	SOP-able	Deterministic template filling — Vesta can run these after SOP
3. HTML content generators	27	$4.86	SOP-able	Brand-compliant HTML from markdown — template-bound, ideal for Vesta
4. No-signal (low-confidence, needs human triage)	40	$7.20	Stay Claude	Triage first; reclassify into a concrete bucket before routing
5. Config + credential scripts	0	$0.00	Stay Claude	Strategic — credential handling, stay on Claude
6. Data analysis scripts	0	$0.00	Stay Claude	Judgment required — stay on Claude unless pattern-locked
7. Tool execution wrappers	0	$0.00	SOP-able	API wrappers with stable schemas — SOP them, route to Vesta
8. Status aggregation	0	$0.00	SOP-able	Read + count + summarize — straightforward Vesta work
9. Template filling (MD -> HTML -> PDF)	0	$0.00	SOP-able	Pattern-locked rendering — move to Vesta immediately
10. Other / unclassified	6	$1.08	Stay Claude	Needs human review before any routing decision

Sub-Category

Count

Spend

Routing

SOP Note

1. Python multi-step glue scripts

$0.36

SOP-able

Pipeline orchestrators, subprocess chains — SOP them for Vesta execution

2. HTML/PDF renderers

$0.18

SOP-able

Deterministic template filling — Vesta can run these after SOP

3. HTML content generators

$4.86

SOP-able

Brand-compliant HTML from markdown — template-bound, ideal for Vesta

4. No-signal (low-confidence, needs human triage)

$7.20

Stay Claude

Triage first; reclassify into a concrete bucket before routing

5. Config + credential scripts

$0.00

Stay Claude

Strategic — credential handling, stay on Claude

6. Data analysis scripts

$0.00

Stay Claude

Judgment required — stay on Claude unless pattern-locked

7. Tool execution wrappers

$0.00

SOP-able

API wrappers with stable schemas — SOP them, route to Vesta

8. Status aggregation

$0.00

SOP-able

Read + count + summarize — straightforward Vesta work

9. Template filling (MD -> HTML -> PDF)

$0.00

SOP-able

Pattern-locked rendering — move to Vesta immediately

10. Other / unclassified

$1.08

Stay Claude

Needs human review before any routing decision

Shift-to-Vesta Savings Projection

Horizon	What Shifts	Spend → Vesta	% of Claude Spend
Day 1	6 already-routine files	$1.08	3.0%
Day 30	Routine + SOP-able Gray (multi-step-glue, renderers, HTML gen, tool-exec, status-agg, template-fill)	$6.48	18.0%

Horizon

What Shifts

Spend → Vesta

% of Claude Spend

Day 1

6 already-routine files

$1.08

3.0%

Day 30

Routine + SOP-able Gray (multi-step-glue, renderers, HTML gen, tool-exec, status-agg, template-fill)

$6.48

18.0%

Real upside is in Day-30 ramp. Day-1 captures only 6 already-classified routine files ($1.08). The meaningful savings come from SOP-ing the Gray sub-categories marked SOP-able (30 files, $5.40) and routing them to Vesta after SOPs are written + Vesta Phase 1 is live. Expected trajectory: $36.00/day → $29.52/day = 18.0% cost reduction.

Honest Notes

Classification is heuristic; gray = needs human review, NOT defaulted to either bucket.

Cost model: $0.18/file blended Opus 4.6, conservative.

Local LLM = Vesta (EGP internal persona) — underlying models: llama3.2:3b primary, qwen2.5:3b fallback. Phase 1 not yet live.

Day-1 savings only captures the 6 already-routine files. Real upside is in SOP-ing the gray-subcategories marked sop_candidate=true and routing them to Vesta over a 30-day ramp.

Gray sub-categories that are SOP-able: multi-step-glue Python, HTML/PDF renderers, template fillers, tool-exec wrappers, status aggregation. These burn $X/day today and should drop to $0 once Vesta is live + SOPs are written.

Next Actions

Write SOPs for the 4 SOP-able Gray sub-categories that are non-zero today: HTML content generators (27 files, $4.86), Python multi-step glue (2 files, $0.36), HTML/PDF renderers (1 file, $0.18).

Stand up Vesta Phase 1: Ollama + llama3.2:3b on EGP HQ hardware, expose via localhost MCP endpoint parallel to the Willemena MCP daemon.

Build routing harness: Betty's task classifier checks the file → routes to Claude (Strategic) or Vesta (SOP-able Routine/Gray) at task dispatch time.

Shadow-test 7 days: run both Claude and Vesta on SOP-able Gray tasks; compare output quality before fully cutting over.

Cut over after shadow-test metrics hit 90%+ quality parity.