Knowledge Graph Guide · Appendix B

Knowledge Graph Cost Modeling: Build vs Buy, Layer-by-Layer Budgets, And Team Composition

Appendix B of the Knowledge Graph Practitioner's Guide. Decomposes the trilogy cost-and-benefit roll-up from Part 11c into the dollar lines a CFO can defend. Covers per-layer cost decomposition (foundation, operational, governance, agent), license-versus-infrastructure-versus-headcount splits, the build-versus-buy decision per layer, the 24-month team curve from one ontologist to a 12-person multi-discipline platform, the role-by-role compensation reality (ontologist, knowledge engineer, graph platform engineer, entity resolution engineer, governance lead, AI engineer), what ROI looks like at steady state, six failure modes in KG cost modeling, an eight-question budget diagnostic, and a tiered Do Next table for the CFO, the CDO, and the procurement officer.

Vikas Pratap Singh · · 35 min read · Data Architecture & Engineering

knowledge-graph · cost-modeling · build-vs-buy · team-composition · lakeside-bank · total-cost-ownership · vendor-procurement · graphrag-cost

Knowledge Graph Practitioner’s Guide: Overview | Part 1 | Part 2 | Part 3 | Part 4 | Part 5 | Part 6 | Part 7 | Part 8 | Part 9 | Part 10 | Part 11a | Part 11b | Part 11c | Appendix A | Appendix B | Appendix C | Part 12

Thursday Morning, Q2 2026: The CFO Has Three Questions

The Wednesday agent demo went well. The CFO had walked into the noon Müller meeting prepared, the agent had cited only gold-tier evidence, and the EU AI Act conformance assessor had read the same governance graph the BCBS 239 examiner read on Tuesday. Three days, three reading patterns, one substrate.

Thursday morning the CFO had the trilogy cost-and-benefit roll-up from Part 11c on her screen and the chief data officer in her conference room. The roll-up said the program ran at $4M to $5.6M annually at steady state and produced $16M to $24M of annual benefit. Three to four times return. Defensible at the board level if the lines underneath the totals held up.

The CFO had three questions.

The first question was where the money goes layer by layer. The roll-up showed the foundation layer at $1.8M to $2.4M, the operational layer at $0.6M to $0.9M, the governance layer at $0.4M to $0.7M, and the agent layer at $1.2M to $1.6M. The CFO wanted to see each layer split into license, infrastructure, and headcount lines so she could compare the foundation’s $1.8M to a peer bank’s published platform-team cost and the agent’s $1.2M to the AI engineering team’s headcount plan she had already approved.

The second question was the build-versus-buy decision per layer. The CDO had named ten vendors in the trilogy text. The CFO wanted to see, for each of the seven layers from Appendix A, whether Lakeside had bought a vendor product, built an in-house implementation, or done both. She wanted the rationale per pick because procurement was about to renew three of the contracts and one of the contracts (the retail master-data platform, which SAP acquired in May 2026) was on a watch list.

The third question was what could go wrong in the budget. Year one was on track. Year two was where peer banks had told her the surprises landed. The CFO wanted the failure modes named explicitly so she could put them on the agenda for the next quarterly architecture review and so the procurement runbook could be revised before the year-two RFPs went out.

This appendix is the answer to all three questions, with the layer vocabulary from Appendix A as the spine and the trilogy roll-up from Part 11c as the totals it has to reconcile to.

Recap: The Trilogy Cost-And-Benefit Roll-Up From Part 11c

Part 11c summarized the full Lakeside program as four annual cost-and-benefit lines.

LayerAnnual costAnnual benefitNet
Foundation$1.8M to $2.4Menables the layers aboveinvestment
Operational$0.6M to $0.9M$7M to $10Mnet positive year one
Governance$0.4M to $0.7M$5M to $8Mnet positive year two
Agent$1.2M to $1.6M$4M to $6Mnet positive year two
Trilogy total$4M to $5.6M$16M to $24M3-4x return at steady state

The key insight from the roll-up was that the foundation cost is amortized across the three application layers and counted once in the total, and that the agent layer is the multiplier rather than the dominant benefit; the operational layer is the largest single dollar line of recovered value (banker reconciliation time and AML throughput), and the foundation is the largest single cost line (the platform team and the graph stores). The trilogy is one program with one foundation cost and three application benefits, not three separate investments.

This appendix decomposes each layer into the dollar lines that produced those totals, anchored to public 2026 pricing and salary data where available.

Decomposing The Foundation Layer: $1.8M-$2.4M Annual

The foundation layer at Lakeside is the substrate the operational, governance, and agent layers all read from: the canonical RDF triple store (see Appendix A for the specific tools), the property-graph traversal store for long-tail counterparty groups, the IRI minter (built in-house, single source of truth for entity identity), the SHACL gate (built in-house, the validation chokepoint between extraction and assert), the entity-resolution engine at the serving edge for sub-200ms agent retrieval, and the central platform team that operates all of it.

Foundation cost lineAnnual rangeWhat it buysBuild, buy, or both
Triple store license (enterprise RDF store, ~10B triples, clustered)$250K-$400K (modeled estimate)the canonical RDF substrate with OWL plus SHACL inference and virtual-graph push-downbuy (capability gap; standards-anchored)
Property graph view (enterprise self-managed, smaller cluster)$50K-$150K (modeled estimate)sub-100ms multi-hop traversal on Müller-style 1,400-entity counterparty groupsbuy (capability gap; ecosystem maturity)
Entity resolution service (entity-resolution engine at serving edge)$100K-$300K (modeled estimate)real-time identity at sub-200ms; dropped from nightly batch in 2025 per Part 5’s locked framingbuy (commodity competence; volume-tier-licensed)
Cloud infrastructure (compute, storage, network for the three services above)$100K-$300Kwarm-and-replicated capacity at 80 percent utilization with burst headroom for quarterly stress runsbuy (cloud)
IRI minter and SHACL gate (built in-house)$0 (license); embedded in platform team compidentity discipline that survives ontology refactors and a quarantine pipeline that survives over-rejectionbuild (criticality, identity is non-negotiable internal IP)
Central platform team (1 ontologist, 1 knowledge engineer, 2 graph platform engineers, 1 ER engineer, 1 governance lead)$1.5M loaded ($250K average loaded comp, six engineers)the discipline that makes the rest of the layer hold; ramps from 3 in month 1 to 6 by month 9build (capability)
Foundation total$2.0M-$2.65M; mid-band $2.0M-$2.4Mthe substrate everything else reads frommixed

The total runs slightly above the trilogy roll-up’s $1.8M-$2.4M band at the high end, which the CDO can defend on two grounds. Entity-resolution volume-tier discounts at Lakeside scale (1.2M retail plus 22K commercial counterparties resolved continuously) drop the ER line into the $100K-$200K band rather than the $300K ceiling. The triple-store license at year three becomes negotiable as renewal terms reflect the platform’s central position in the bank’s stack rather than a procurement evaluation. Headcount is the line that does not move down: 70 percent of the foundation cost is people, and the people are what makes the layer hold.

The compensation lines deserve specificity because they reappear in the operational, governance, and agent layers. The Glassdoor 2026 ontology engineer salary range is $156K to $264K with a $202K average; the loaded cost (base plus bonus plus equity plus benefits plus overhead) at a regional US bank lands in the $250K-$320K range for senior individual contributors. ZipRecruiter’s 2026 Neo4j developer salary range is $115K to $149K base; loaded for a graph platform engineer in Chicago lands at $200K-$250K. Built In’s 2026 senior data engineer compensation reports a $164K total comp average that grows to $300K+ at FAANG; Lakeside’s mid-band is closer to $200K-$240K loaded. The ratios matter more than the absolute numbers; the foundation team’s 70-percent share of the layer cost is the structural fact, and a budget that swaps two engineers for an extra $200K of license is a budget that will not ship.

The Lakeside foundation layer's cost: 70 percent headcount, 20 percent license, 10 percent infrastructure, and what is built or bought in each The foundation layer runs $1.8M to $2.4M a year. One bar splits it: headcount 70 percent, license 20 percent, infrastructure 10 percent. A dashed line on the bar marks 60 percent: if headcount shrinks below 60 percent of the bar at your firm, the layer is at shelfware risk. Headcount is the dominant cost line in three of four layers; license is the dominant line in zero. Headcount, $1.5M a year: six engineers at an average of $250K loaded, namely 1 ontologist, 1 knowledge engineer, 2 graph platform engineers, 1 ER engineer and 1 governance lead, cross-loaded in year 1. Build: the 6 engineers operate the platform. They run the IRI minter, a one-engineer-month build and the single source of truth for entity identity; the SHACL gate, the validation chokepoint between extraction and assert; the OpenLineage-to-PROV-O bridge, a mechanical mapping from Part 7's eight-row table; the tier-policy classifier, which lives in procedural memory and is the agent dispatch point; and the response post-processing sidecar, a citation verifier with a +40 to 80 ms latency budget. Internal IP is identity, validation and dispatch: the discipline that makes the layer hold. License, $400K to $850K a year, bought from standards-anchored vendors: a W3C RDF triple store at $250K to $400K, an openCypher property graph at $50K to $150K, and a volume-tier entity-resolution engine at $100K to $300K. The reason is a capability gap: 12+ months to build at production scale. Infrastructure, $100K to $300K a year, bought as cloud: compute, storage and network; CapEx to OpEx, elastic for quarterly stress runs, a utility commodity. Year-three trajectory at Lakeside scale: the ER line falls from a $300K ceiling in year 1 to the $100K to $200K band in year 3 on volume-tier discounts; the triple-store renewal moves from price discovery in year 1 to renewal leverage in year 3, turning negotiable as the platform's central position is established; headcount does not move down, it is the structural floor, and programs that under-hire at year 1 ship the layer 6 to 12 months late. Ranges are 2026 mid-band Lakeside numbers; peer banks at $50B to $100B in assets should land within plus or minus 25 percent. Programs that over-license while under-hiring land at the most common KG cost overrun: a fancy graph store with no consumer. THE FOUNDATION LAYER: $1.8M TO $2.4M A YEAR 60% 70% 20% 10% Headcount License Infra Headcount is the dominant cost line in three of four layers; license is the dominant line in zero. Dashed line: if headcount shrinks below 60 percent of the bar at your firm, the layer is at shelfware risk. HEADCOUNT · 70% $1.5M a year Six engineers, average $250K loaded: 1 ontologist, 1 knowledge engineer, 2 graph platform engineers, 1 ER engineer, 1 governance lead (cross-loaded in year 1) BUILD the 6 engineers operate the platform IRI minter, a one-engineer-month build single source of truth for entity identity SHACL gate validation chokepoint between extraction and assert OpenLineage-to-PROV-O bridge mechanical mapping from Part 7's eight-row table Tier-policy classifier lives in procedural memory; the agent dispatch point Response post-processing sidecar citation verifier, +40 to 80 ms latency budget Internal IP: identity, validation, dispatch; the discipline that makes the layer hold. LICENSE · 20% $400K to $850K a year BUY standards-anchored W3C RDF triple store $250K to $400K openCypher property graph $50K to $150K volume-tier entity-resolution engine $100K to $300K capability gap: 12+ months to build at production scale INFRASTRUCTURE · 10% $100K to $300K a year BUY cloud cloud compute, storage, network CapEx to OpEx; elastic for quarterly stress runs utility commodity YEAR-THREE TRAJECTORY AT LAKESIDE SCALE YEAR 1 YEAR 3 ER line $300K ceiling $100K to $200K volume-tier discounts at Lakeside scale Triple store price discovery renewal leverage renewal turns negotiable as the platform's central position is established Headcount does not move down: the structural floor under-hire at year 1 and the layer ships 6 to 12 months late Ranges are 2026 mid-band Lakeside numbers; peer banks at $50B to $100B in assets should land within plus or minus 25 percent. Programs that over-license while under-hiring land at the most common KG cost overrun: a fancy graph store with no consumer.

Decomposing The Operational Layer: $0.6M-$0.9M Annual

The operational layer is the three production services from Part 11a: customer 360 (the Müller-family SPARQL query), beneficial ownership (FinCEN BOI plus 25-percent and substantial-control reasoning), and real-time transaction risk (the 100ms five-step traversal). The cost is incremental to the foundation, not standalone. The operational layer assumes the foundation is already paid for and reads from it.

Operational cost lineAnnual rangeWhat it buys
Per-business-unit ramp engineering (3 BUs at 0.5-1.0 engineer-year of incremental enhancement)$400K-$700Kservice-line integration into the operational graph; new R2RML mappings; new SHACL shapes; per-BU SLAs
Incremental compute for SPARQL endpoint scaling at peak$50K-$100Kwarm capacity for the transaction-risk service at 99.9-percent SLA; burst headroom for end-of-month batch
Per-BU onboarding ramp (one-time amortized over 24 months)$100K-$150KSHACL shape coverage; lineage validation; consumer training; runbook authoring
Operational total$0.55M-$0.95M; mid-band $0.6M-$0.9Mthree production services with explicit latency, accuracy, and audit budgets

The benefit line is the dominant one in the trilogy. Banker reconciliation time recovered from the Müller-family pattern alone is in the $3M-$4M annual range at Lakeside scale (assuming 22,000 commercial counterparties times an average of two complex reconciliations per relationship per year times $80 of banker time per reconciliation pre-KG times a 95-percent reduction post-KG; the math is sensitive to the volume assumption but the order of magnitude is robust). The AML investigation throughput improvement adds $2M-$3M; the transaction-risk service unit-cost reduction adds $1M-$2M. The $7M-$10M operational benefit band is conservative.

The operational layer is also the layer where public case studies are most directly comparable. Citi Private Banking re-engineered front-to-back data flows on a property-graph store; Lloyds reduced financial crime detection cycle time by an undisclosed but material amount; the Latin America Global 50 bank connected one trillion data relationships for real-time credit risk insights. None of those firms publish per-line cost decompositions, but the public benefit narratives are consistent with the order of magnitude Lakeside’s operational layer reports.

Decomposing The Governance Layer: $0.4M-$0.7M Annual

The governance layer is the application from Part 11b: the BCBS 239 attribute-level lineage, the ECB RDARR cross-walks, the GDPR Article 30 ROPA, and the EU AI Act Article 10 training-data provenance. The cost shape is different from the operational layer because the OpenLineage-PROV-O bridge is a consumption pattern, not new instrumentation. Lakeside’s Spark, dbt, and Airflow pipelines were already emitting OpenLineage events ($480K Run events per day across 3,200 Spark plus 1,400 dbt plus 200 Airflow); the bridge ingests those events into the named-graph governance partition rather than re-instrumenting the pipelines.

Governance cost lineAnnual rangeWhat it buys
Governance lead plus 1 governance engineer (loaded)$400K-$500Kthe human discipline that authors regulatory templates, manages quarterly FIBO releases, and operates the trust-tier-by-reporting-surface table
Quarterly FIBO release management (1 engineer-quarter per year)$60K-$80Kdual-write windows, alias-pinning, named-graph version chains across releases per Part 8
Data catalog and governance-metadata platform license (catalog UX over the governance graph)$100K-$200K (modeled estimate)active-metadata UX, MCP server, and GraphQL API per Appendix A; alternative is an open-source metadata graph at zero license but higher integration cost
OpenLineage metadata store infrastructure (incremental, since OpenLineage emission already exists)$20K-$50KOL receiver, named-graph ingestion service, retention policy
Governance total$0.58M-$0.83M; mid-band $0.4M-$0.7Mfour regulatory templates against one substrate plus the version chain that survives quarterly FIBO releases

The total runs above the roll-up’s $0.4M-$0.7M band at the upper end if all lines hit the high end of their range, which is a discipline signal: at year one Lakeside ran with one shared governance engineer (cross-loaded with the platform team) rather than a dedicated headcount, dropping the layer to the $400K-$500K band. By year two the dedicated lead joined and the layer reached the $600K-$700K band. The trade-off is whether the bank’s regulatory cycle can wait six to nine months for the governance layer to catch up to the operational layer; at Lakeside the answer was no because the EU AI Act Article 10 enforcement date (August 2, 2026) anchored the governance layer’s ramp.

The benefit lines in the governance layer are dominated by regulatory remediation cost avoidance and reduced finding cycle time. Organizations adopting knowledge-graph-backed compliance reporting report materially faster compliance cycles in published case studies, though the exact percentage varies by firm and is rarely disclosed; at Lakeside scale the dollar value translates to $3M-$5M of avoided regulatory remediation work and $2M-$3M of cross-walk authoring efficiency (writing one SPARQL template per regulation against the same graph rather than commissioning a separate per-regulation reconciliation project). The $5M-$8M governance benefit band is again conservative.

What this looks like in practice. The governance layer is the most under-budgeted layer in 2026 KG programs that did not start with a regulatory deadline. Programs without a forcing function (BCBS 239, ECB RDARR, EU AI Act Article 10) tend to defer the governance layer to year three, by which time the operational graph has accreted four years of consumer dependencies that the governance refactor has to break and rebuild. The Lakeside discipline lock from Part 11b is to ramp the governance layer in parallel with the operational layer, not after; the cost is the same, the layer arrives a year earlier, and the consumer dependencies grow with the version chain rather than against it.

Decomposing The Agent Layer: $1.2M-$1.6M Annual

The agent layer from Part 11c is the relationship-banker agent v3, including the CoALA four-layer memory mapping (semantic on the operational graph, episodic per banker, procedural per agent version), the three-policy trust-tier-aware retrieval pattern (strict-tier-floor, tier-segregated, tier-explicit-citation), the tier-policy classifier (a fine-tuned model living in the procedural-memory subgraph), and the response post-processing sidecar. The cost is the AI engineering team plus model inference plus the memory framework plus the runtime infrastructure; it is the largest non-foundation layer at Lakeside and the layer with the most cost-trajectory volatility.

Agent cost lineAnnual rangeWhat it buys
AI engineering team of ~4 (1 agent platform lead, 2 AI engineers, 1 ML engineer for tier-policy classifier and retraining loops)$1.0M-$1.2Mthe discipline that maintains the three-workflow / three-policy / three-enforcement-layer structure across model upgrades and quarterly FIBO releases
Model inference (deferred-summarization extraction for Track 2 plus a small index-time-summarization overlay for sensemaking)$50K-$150Kthe actual LLM compute; the order-of-magnitude saving comes from the GraphRAG cost cliff at comparable quality at 0.1 percent cost (1000x)
Memory framework (agent episodic-memory store, primarily open-source with managed-service exception for the bi-temporal index)$30K-$80Kthe framework that makes per-banker episodic memory operable at production latency; alternative is a managed episodic-memory tier at higher cost but less integration effort
Agent runtime infrastructure (vector index for the small embedding cache, embedding compute, observability)$50K-$100Kwarm capacity for retrieval-planner SPARQL emission and post-processing sidecar
Tier-policy classifier and workflow classifier (training data labeling plus quarterly retraining)$20K-$40Kthe procurement of labeled data for the 2,500-query training set plus the human-in-the-loop labeling cost for ongoing retraining
Response post-processing sidecar (citation verifier; small added latency budget, illustrative not instrumented)$30K-$60Kthe discipline that catches the small number of blocks per week per the illustrative Lakeside production rate from Part 11c
Agent total$1.18M-$1.63M; mid-band $1.2M-$1.6Mthe relationship-banker agent at three workflows, three policies, three enforcement layers, one substrate

The cost cliff in the model-inference line is the most volatile cost trajectory in the entire trilogy. Early index-time-summarization GraphRAG of a 5-gigabyte legal corpus cost $33,000 in early 2024; by late 2025, deferred-summarization variants had reduced the equivalent indexing cost to roughly 0.1 percent of that figure (a 1000x drop) at comparable quality. Lakeside’s pick from Appendix A (a deferred-summarization extraction pipeline for credit memos and KYC plus a small index-time-summarization overlay for narrative sensemaking) is what holds the model-inference line to $50K-$150K; a 2024-vintage index-time-summarization-only architecture at the same corpus scale would have run at $1M-$3M in pure indexing cost and would have made the agent layer indefensible.

The benefit band ($4M-$6M annual) is the smallest of the three application layers because the agent layer is the multiplier on the foundation and the operational layer rather than a primary value source. Banker productivity recovered through agent-assisted prep is the dominant line ($2M-$3M); reduced reliance on senior-banker reach-around for routine questions adds $1M-$2M; the new-revenue line from improved client-meeting throughput adds $1M. Year one of the agent layer is ramp; year three is steady state.

Team Composition Over The 24-Month Build-Out

The trilogy’s $4M-$5.6M run rate at steady state is the year-three number. The path from year one to year three runs through a team composition curve that starts with one ontologist in month one and reaches roughly 12 multi-discipline engineers by month 24. The shape of the curve is the discipline that makes the budget hold: hire ahead of the layer, not behind it.

MonthFoundation teamOperational rampGovernance rampAgent rampCumulative headcount
11 ontologist(CDO and architect-lead are pre-existing)··1
3+ 1 knowledge engineer; + 1 graph platform engineer···3
6+ 1 ER engineer; + 1 graph platform engineer+ 1 BU lead (customer 360)··6
9+ 1 governance lead (cross-loaded; transitions to governance layer at month 15)+ 1 BU lead (beneficial ownership)··8
12(foundation team at 6 IC)+ 1 BU lead (transaction risk)(governance lead transitions in 3 months)+ 1 agent platform lead10
15(steady at 6)(steady at 3)dedicated 1 governance lead + 1 governance engineer+ 1 AI engineer12
18(steady)(steady)(steady at 2)+ 1 AI engineer; + 1 ML engineer14 (peak ramp)
24(steady at 6)(steady at 3)(steady at 2)(steady at 4 with agent platform lead, 2 AI eng, 1 ML eng)12 (hold to steady state)

The compensation reality across the role types matters operationally because the loaded cost of each engineer determines whether the layer fits its band. The ranges below are 2026 US public-data anchors translated into Lakeside-band loaded compensation (base plus bonus plus equity plus benefits plus 25-percent overhead).

RolePublic 2026 base rangeLakeside loaded compFirst month on teamSteady-state count
Ontologist$156K-$264K, average $202K (Glassdoor 2026; senior IC range is wider per role title ambiguity)$200K-$280K loadedmonth 11
Knowledge engineer (Track 2 lead)$120K-$180K base (cross-mapped to senior data engineer + ontology premium)$200K-$280K loadedmonth 31
Graph platform engineer$115K-$149K base for Neo4j developer (ZipRecruiter 2026); loaded with SRE-level on-call premium$200K-$260K loadedmonth 3 (first), month 6 (second)2
Entity resolution engineer$130K-$180K base (specialist data engineer with ER-tooling experience)$220K-$280K loadedmonth 61
Governance lead$160K-$220K base (Data Governance director track at regional bank)$280K-$380K loadedmonth 15 (dedicated)1
Governance engineer$130K-$170K base$220K-$270K loadedmonth 151
Operational BU lead$140K-$190K base (senior data engineer plus product premium)$230K-$300K loadedmonths 6, 9, 123
Agent platform lead$200K-$280K base (AI engineering manager track)$360K-$480K loadedmonth 121
AI engineer$180K-$260K base (Built In senior data engineer plus AI premium)$300K-$420K loadedmonths 15, 182
ML engineer (classifier and retraining loops)$170K-$240K base$280K-$380K loadedmonth 181
Steady-state team·$4.0M-$5.4M total loaded comp at year three·12 (one BU lead drops to part-time after ramp, so 13 role seats net to 12 FTE)

The total team comp at year three reconciles to the trilogy run rate: $4.0M-$5.4M of headcount lines plus $0.6M-$1.1M of license and infrastructure across all four layers lands at the $4.6M-$6.5M total cost band, which is consistent with the roll-up’s $4M-$5.6M when the per-layer benefit lines amortize the foundation.

The Lakeside team over 24 months: cumulative headcount, and who joins when in each layer A chart of cumulative headcount over the 24-month build-out, from one ontologist to a peak of 14 and a steady state of 12. The total is 1 at month 1, 3 at month 3, 6 at month 6, 8 at month 9, 10 at month 12, 12 at month 15 and 14 at month 18, the peak ramp, when all four layers are staffed. By month 24 it is 12, the steady state, at a $4M to $5.6M run rate. Below the chart, on the same months, who joins when in each layer; a dot marks a change and its number is the layer's size after it. Foundation, 1 to 6 ICs (ontologist, knowledge engineer, graph platform engineers, ER engineer): 1 at month 1, 3 at month 3, 5 at month 6, 6 at month 9, 5 at month 15. The ontologist starts at month 1, so ontology decisions come before engineers. Operational BU leads, 0 to 3, one per BU (customer 360, beneficial ownership, transaction risk): 1 at month 6, 2 at month 9, 3 at month 12; the first BU integration starts at month 6. Governance, 0 to 2: cross-loaded under the foundation from month 9 to month 15, drawn dashed, and dedicated from month 15 with 2, in parallel with the agent ramp. Agent, 0 to 4 (agent platform lead at month 12, AI engineers at months 15 and 18, ML engineer at month 18): 1 at month 12, 2 at month 15, 4 at month 18, 3 at month 21, 2 at month 24. The agent platform lead joins at month 12 because the foundation must be operable first. Hire ahead of the layer, not behind it: programs that hire behind ship the layer 6 to 12 months late. Headcount is the dominant cost line in three of four layers; the curve's shape is the budget discipline that makes the layers ship in the right order. CUMULATIVE HEADCOUNT, MONTHS 1 TO 24 1 3 6 8 10 12 14 12 Month 18: peak ramp, all four layers staffed Month 24: steady state at 12, a $4M to $5.6M run rate M1 M3 M6 M9 M12 M15 M18 M21 M24 WHO JOINS WHEN, BY LAYER A dot marks a change; its number is the layer's size after it. Foundation · 1 to 6 ICs 1 3 5 6 5 ontologist, knowledge engineer, graph platform and ER engineers M1 ontologist starts: ontology decisions before engineers Operational BU leads · 0 to 3 1 2 3 one per BU: customer 360, beneficial ownership, transaction risk M6 first BU integration starts Governance · 0 to 2 2 dashed: cross-loaded under foundation, M9 to M15 M15 governance dedicated, in parallel with the agent ramp Agent · 0 to 4 1 2 4 3 2 agent platform lead M12, AI engineers M15 to M18, ML engineer M18 M12 agent platform lead joins: foundation must be operable first Hire ahead of the layer, not behind it. Programs that hire behind ship the layer 6 to 12 months late. Headcount is the dominant cost line in three of four layers; the curve's shape is the budget discipline that makes the layers ship in the right order.

Build Versus Buy Per Layer

The build-versus-buy decision is per layer, not per program. A simple three-factor lens I use (Capability, Complexity, Criticality) lands on the same answer at every layer Lakeside considered.

LayerBuild, buy, or bothRationale
Triple store (canonical RDF substrate)buy (the canonical RDF triple store)Capability gap (W3C OWL plus SHACL inference is not a 12-month build); Complexity high (clustered RDF at 10B-triple scale); Criticality high (substrate; all consumers depend on it). Loss of optionality if built in-house outweighs the license cost.
Property graph view (traversal)buy (the property-graph traversal store)Same logic; openCypher plus graph data science tooling is not a build candidate at Lakeside scale.
IRI minterbuild (in-house)Capability available (single source of truth for entity identity is a one-engineer-month build); Complexity low to medium (IRI scheme is documented in Part 5); Criticality very high (IRI drift is the month-7 Lakeside lesson; centralized minting is internal IP that protects identity discipline).
SHACL gatebuild (in-house)Capability available (the ontology team owns the shapes); Complexity medium (the discipline is in the shape design, not the gate); Criticality high (the gate is the chokepoint between extraction and assert; outsourcing it means outsourcing Data Quality).
Entity resolutionbuy (the entity-resolution engine at serving edge) plus retain (the retail master-data platform)Capability gap (real-time ER is a research-grade build); Complexity very high (a published entity-resolution TCO model accounts for false-positive rates that an in-house build typically misses); Criticality high (sub-200ms agent retrieval depends on it).
LLM extraction (Track 2)buy (an LLM-assisted extraction pipeline) plus build (the Lakeside-specific FIBO-locked prompt and the dedup-and-ER-and-SHACL-gate triad from Part 6)Framework is commodity by 2026; the discipline is the prompt and the post-extraction triad, which is internal IP.
OpenLineage-to-PROV-O bridgebuild (in-house)Capability available (mechanical mapping per Part 7’s eight-row table); Complexity low to medium (the bridge is a translator, not a new emission); Criticality medium (lineage is consumption pattern; failure mode is degradation, not data loss).
Quarterly FIBO release managementbuild (in-house playbook) plus buy (FIBO ontology releases, no license)The playbook is the six-stage migration pattern from Part 8; FIBO itself is open-source.
Governance metadata storebuy (the data catalog and governance-metadata platform) with an open-source metadata graph fallback on the option listCapability gap (active-metadata UX, MCP, GraphQL API); Complexity high; Criticality medium. The build alternative (an open-source metadata graph at zero license but higher integration cost) sits on the option list for renewal-leverage purposes.
Agent memory frameworkbuy (the agent episodic-memory store, primarily open-source with managed exception for the bi-temporal index)Framework competence is the 2026 commodity; the discipline is the CoALA mapping and the trust-tier policies, both internal IP.
Tier-policy classifierbuild (in-house, lives in procedural memory)Capability available; Complexity medium (the labeled-data work is the long pole); Criticality very high (the classifier is the dispatch point for the three-policy architecture from Part 11c).
Response post-processing sidecarbuild (in-house)Capability available; Complexity low; Criticality high (the citation-verifier catches the small number of blocks per week that protect against the Apex replay).

The pattern is consistent. Lakeside bought the layers where the standards are mature and the capability gap is real (graph stores, ER, governance metadata, framework competence) and built the layers where the discipline is internal IP (identity, validation gate, lineage bridge, classifiers, sidecar). The split is not 50-50 by line count; by dollar share at year three it is roughly 40-percent-build (the foundation team and the agent team comp) and 60-percent-buy (license plus infrastructure plus the bought-team-effort that licenses replace). The ratio reverses at smaller programs that lean more heavily on managed services and lighter on in-house engineering; the choice is a function of organizational capability and regulatory shape, not a universal default.

KEY INSIGHT: the most common KG cost overrun in 2026 is not over-licensing or over-hiring; it is mismatch. A program that buys a $300K triple-store license without funding the ontologist who decides what the triples mean produces a fancy graph store with no consumer. A program that hires three AI engineers without funding the foundation team produces an agent without a substrate. The layer-level build-versus-buy decision keeps the ratio honest at every layer; programs that skip the per-layer decision land at one of the two mismatches and discover the gap in year two when the application layers fail to ramp.

What ROI Looks Like At Steady State

The trilogy roll-up’s 3-4x return at steady state is one specific point on a wider distribution that the public KG case-study record substantiates within a credible band.

Forrester’s commissioned Total Economic Impact study of Neo4j reported a 417-percent ROI over three years for a composite organization, decomposed into 43-percent improved business results, 35-percent digital transformation savings, and 22-percent accelerated time-to-value. The study is vendor-commissioned and should be read directionally rather than as a literal benchmark; the order-of-magnitude is consistent with Lakeside’s 3-4x but not directly comparable (Forrester’s composite includes use cases beyond the specific operational-governance-agent trilogy Lakeside ran).

AstraZeneca’s Biological Insights Knowledge Graph is the most-cited public KG case study. AstraZeneca reported that time to generate insights decreased from over six months to less than 2.5 months, a 150-percent improvement, after adopting the platform that supports BIKG. In a domain where drug development costs $2.6B and 90 percent of candidates fail, a 150-percent acceleration on insight generation is a real benefit even if the dollar conversion is opaque. The Lakeside operational layer’s banker-reconciliation-time recovery is structurally analogous: the metric is human-hours-of-skilled-labor, the benefit is the multiplier on those hours, and the cost is the platform team that maintains the substrate.

Citi Private Banking’s case study describes front-to-back data flow re-engineering for global private banking clients; eBay’s Akutan and Beam knowledge graph ran on a 20-server deployment with 2.5 billion facts to manage 3,000-plus services architecture; Lloyds deployed a knowledge graph for financial crime detection. None of these published per-line cost decompositions, but the combination of qualitative operational benefit narratives (materially faster compliance cycles and improved product conversion reported in vendor case studies, though the exact percentages are firm-specific and rarely disclosed) and the public market-size data (graph database market projected at $3-4B in 2026 growing to $20B by 2034 at ~24-percent CAGR) puts the trilogy’s 3-4x return inside the credible distribution rather than at the optimistic tail.

The ROI lines that Lakeside watches most closely at the quarterly architecture review are the cost-trajectory lines, not the absolute ROI number. The GraphRAG cost cliff ($33,000 to $33 in 18 months at the same corpus scale, a 1000x shift driven by deferred-summarization extraction variants) is the cost-trajectory pattern that compounds across the agent layer. Lakeside’s procurement runbook treats the model-inference line as a re-evaluable cost every two quarters precisely because the trajectory is steep enough to invalidate prior procurement decisions inside the budgeting cycle.

Six Failure Modes In KG Cost Modeling

The patterns below recur across the budget-defense post-mortems this series has aggregated. Each is a diagnostic for whether the Thursday-morning CFO conversation needs to happen at your firm before year two.

  1. Under-budgeting entity resolution. The most under-budgeted layer in the 2026 KG procurement is ER, per the Appendix A lock. Programs that estimate ER as part of the triple-store cost rather than as a standalone serving-edge service land at the entity-resolution-engine cost surprise in year two when the agent layer’s sub-200ms latency requirement forces the upgrade from batch to real-time. The order-of-magnitude shift is $50K-to-$300K-per-year underbudget at Lakeside-comparable scale.

  2. Treating GraphRAG indexing as a one-time cost. The 2024-vintage architecture assumed indexing was a fixed-cost build, but the cost cliff is steep enough that a corpus indexed in early 2024 at $33,000 would have re-indexed in late 2025 at $33 with comparable quality. Programs that locked in index-time-summarization-only architectures at 2024 prices spent roughly 1000x (0.1 percent the cost in reverse) more than 2026 alternatives at no quality benefit. The procurement runbook discipline is to re-evaluate the model-inference line every two quarters.

  3. Year-2 ramp underestimation. Year one is when the foundation lands and the operational layer ships its first BU. Year two is when the foundation matures, the operational layer scales to all three BUs, the agent layer demand spikes, and the governance layer reaches dedicated headcount. Programs that budget year two at year-one’s headcount level miss the multiplier and underbudget the team by 30-40 percent. The Cutter Consortium implementation cost guidance is consistent: pilot-to-extensible-platform runs $1M-$3M, and the long-term true enterprise KG runs $10M-$20M cumulative across years one through five.

  4. Headcount-versus-license imbalance. The most expensive failure mode at Lakeside-comparable scale is over-licensing while under-hiring. A program that buys a $400K triple-store license, a $200K catalog license, and a $300K ER license but funds only three engineers cannot operate the layer; the layer becomes shelfware. The headcount-to-license ratio at the foundation layer should be in the 3-4x range (people cost roughly three to four times the license cost) at Lakeside scale; ratios below 2x are the diagnostic for shelfware risk.

  5. Vendor M&A repricing. The 2026 vendor consolidation timeline from Appendix A names six events in 18 months; any of them can change the renewal economics. The completed SAP acquisition of the retail master-data platform (closed May 2026, see Appendix A) is the active example for Lakeside’s retail master; the renewal economics are now being re-evaluated post-integration, and the procurement runbook revisits the contract at every quarterly architecture review with documented trigger conditions for the ER re-evaluation. Programs without an M&A watch process discover the repricing at renewal time and have no negotiating leverage.

  6. Tooling-without-team and team-without-tooling. The mirror images of failure mode 4. A program that issues a procurement-led RFP without an ontologist on staff (the original Lakeside RFP from Appendix A’s opener) lands at a fancy browser without a production consumer. A program that hires an ontologist and three engineers without funding the storage and ER procurement lands at twelve months of in-house RDF tooling rebuild that the open market would have sold for $300K-$500K. Both failure modes look like budget discipline up front and produce 30-50 percent budget overruns by year two.

Eight-Question KG Budget Diagnostic

The questions below separate a defensible KG budget from a budget that will be challenged by the CFO in year two. Run the list against your own budget proposal.

QuestionDefensible answerFailure mode
What is the headcount-to-license ratio at the foundation layer?3-4x at Lakeside scale; people cost is three to four times license costRatios below 2x produce shelfware (failure mode 4)
Has the operational layer’s per-BU ramp engineering been budgeted at 0.5-1.0 engineer-year per BU per year?Yes; the operational benefit depends on itOperational layer ramps slower than budgeted; benefit lags by 12 months
Is the governance layer ramping in parallel with the operational layer or deferred to year three?Parallel; the EU AI Act Article 10 enforcement date and the BCBS 239 / ECB RDARR cycle anchor the parallel rampYear-three governance refactor is 2-3x more expensive than parallel governance
Is the agent layer being asked to pay back without the foundation having shipped?No; agent layer ramp begins at month 12 only after the foundation is operablePre-foundation agent layer is the Apex Capital incident from Part 9 replayed in your firm
Does the model-inference line acknowledge the GraphRAG cost cliff?Yes; a deferred-summarization extraction budget at 0.1-percent of 2024-vintage index-time-summarization-only costA 2024-vintage index-time-summarization-only architecture in a 2026 budget is 1000x over-budget for no quality benefit
Is the build-versus-buy decision per layer rather than per program?Yes; the three-factor lens produces different answers at different layersA single platform purchase under the assumption that one vendor owns all seven layers is the most expensive procurement mistake (failure mode 6)
Is there an M&A watch process on the vendor stack?Yes; per-layer trigger conditions for re-evaluation at every quarterly architecture reviewVendor M&A repricing arrives at renewal with no negotiating leverage (failure mode 5)
Is the year-2 budget 30-40 percent above year 1?Yes; the operational layer scales, the governance layer hits dedicated headcount, the agent layer rampsFlat year-1-to-year-2 budgets are the most common cost overrun (failure mode 3)

Build-Versus-Buy Decision Tree

The decision tree below operationalizes the per-layer build-versus-buy framework. Run it for each of the seven layers in your stack.

The build-versus-buy decision tree, per layer: three questions, four answers, and Lakeside's picks A three-factor lens, applied per layer, not per program. Three questions. One, Capability: do you have in-house expertise to build a production-grade implementation in less than 12 months? Two, Complexity: is the layer's complexity dominated by domain-specific discipline, or by general-purpose engineering? Three, Criticality: is the layer internal IP that competitors cannot replicate, or is it commodity capability available off the shelf? The three answers land in one of four cells, each shown with Lakeside's picks. Buy, for no capability, general-purpose and commodity: the canonical RDF triple store, the property-graph traversal store, the governance-metadata platform and cloud infrastructure. Build, for yes capability, domain-specific and internal IP: the IRI minter, the SHACL gate, the tier-policy classifier and the response post-processing sidecar. Both, for yes capability, domain-specific and a commodity framework: the LLM-assisted extraction pipeline plus the Lakeside FIBO-locked prompt, and the OpenLineage-PROV-O bridge, from the mapping table in Part 7. Buy then build the discipline, for no capability, domain-specific and internal IP discipline: the entity-resolution engine, the agent episodic-memory store with the CoALA mapping, and the quarterly FIBO release playbook. Three guardrails when reading the tree. One: capability gap means real gap. Rebuilding standards-anchored vendors (W3C RDF, openCypher, OpenLineage) is rarely an internal-IP win; the standards age better than custom code, and the vendor competition keeps the cost trajectory honest. Two: internal IP is identity, validation, and dispatch, not implementation choice. The IRI minter, SHACL gate, and tier-policy classifier are internal IP because the discipline survives ontology refactors and model upgrades; the framework that runs them is commodity. Three: “both” is the most common answer at the framework layer. The framework is the buy and the prompt or shape or playbook is the build. Programs that try to build the framework underestimate the maintenance burden; programs that try to buy the discipline get a generic platform. At year three Lakeside's split is roughly 60 percent buy (license plus infrastructure plus bought-team-effort) and 40 percent build (foundation team plus agent team comp). Smaller programs lean more on managed services and lighter on in-house engineering; the choice is a function of organizational capability and regulatory shape, not a universal default. Programs that decide once for the whole platform land at one of the two mismatches (over-license under-hire, or under-license over-hire) and produce 30 to 50 percent overruns by year two. THREE QUESTIONS, PER LAYER not per program 1 · Capability Do you have in-house expertise to build a production-grade implementation in less than 12 months? 2 · Complexity Is the layer's complexity dominated by domain-specific discipline, or by general-purpose engineering? 3 · Criticality Is the layer internal IP that competitors cannot replicate, or is it commodity capability available off the shelf? FOUR CELLS, WITH LAKESIDE'S PICKS BUY no capability + general-purpose + commodity Canonical RDF triple store Property-graph traversal store Governance-metadata platform Cloud infrastructure BUILD yes capability + domain-specific + internal IP IRI minter SHACL gate Tier-policy classifier Response post-processing sidecar BOTH yes capability + domain-specific + commodity framework LLM-assisted extraction pipeline + Lakeside FIBO-locked prompt OpenLineage-PROV-O bridge (mapping table from Part 7) BUY then BUILD discipline no capability + domain-specific + internal IP discipline Entity-resolution engine Agent episodic-memory store (+ CoALA mapping) Quarterly FIBO release playbook THREE GUARDRAILS WHEN READING THE TREE 1 Capability gap means real gap. Rebuilding standards-anchored vendors (W3C RDF, openCypher, OpenLineage) is rarely an internal-IP win; the standards age better than custom code, and the vendor competition keeps the cost trajectory honest. 2 Internal IP is identity, validation, and dispatch. Not implementation choice. The IRI minter, SHACL gate, and tier-policy classifier are internal IP because the discipline survives ontology refactors and model upgrades; the framework that runs them is commodity. 3 “Both” is the most common answer at the framework layer. The framework is the buy and the prompt or shape or playbook is the build. Programs that try to build the framework underestimate the maintenance burden; programs that try to buy the discipline get a generic platform. AT YEAR THREE, LAKESIDE'S SPLIT IS ROUGHLY Buy 60% Build 40% Buy: license plus infrastructure plus bought-team-effort. Build: foundation team plus agent team comp. Smaller programs lean more on managed services and lighter on in-house engineering; the choice is a function of organizational capability and regulatory shape, not a universal default. Programs that decide once for the whole platform land at one of the two mismatches (over-license under-hire, or under-license over-hire) and produce 30 to 50 percent overruns by year two.

Tiered Do Next Table

PriorityActionWhy It Matters
Now (this quarter)Decompose your KG budget proposal into per-layer license-versus-infrastructure-versus-headcount lines that reconcile to a per-layer total. If you cannot reconcile the lines, the proposal cannot be defended at the CFO level.The CFO question is per layer, not per program. Roll-up totals without per-layer decomposition are the first failure mode of KG procurement
Now (this quarter)Run the eight-question budget diagnostic against your current budget. Mark each question with a defensible answer or a failure mode.The diagnostic surfaces the year-two surprises before they arrive; the diagnostic costs an afternoon and saves 30-40 percent of overrun risk
Next (next two quarters)Run the build-versus-buy decision tree for each of the seven Appendix A layers. Document the rationale per layer. Match the dollar share against the 60-percent-buy / 40-percent-build Lakeside ratio at your scale.Per-program build-versus-buy decisions land at one of the two mismatches; per-layer decisions match the Lakeside discipline that makes the layers ship
Next (next two quarters)Establish the M&A watch process on the vendor stack. Document trigger conditions per vendor. Add to the quarterly architecture review agenda.The 2026 vendor consolidation timeline names six events in 18 months; any of them can change renewal economics. Programs without a watch process discover the repricing at renewal with no leverage
Soon (next two to four quarters)Build the 24-month team curve for your program. Identify the first month each role joins. Match the curve to the layer ramp; hire ahead of the layer, not behind it.Programs that hire behind the layer ship the layer six months late; the cost is 12-18 months of foregone benefit
Soon (next two to four quarters)Reset the model-inference line to deferred-summarization-class extraction cost (see Appendix A for the specific tools). Document the 2-quarterly re-evaluation cadence.The GraphRAG cost cliff is 1000x at comparable quality; year-old architectures are a risk to the agent-layer budget
Eventually (year two and beyond)Establish the trilogy ROI tracking discipline: track the operational benefit (banker time recovered or analogue), the governance benefit (regulatory remediation cost avoided), and the agent benefit (productivity recovered) as separate lines on the program scorecard.The 3-4x return at steady state is the year-three number; without per-layer ROI tracking the program loses defensibility at year two when the agent layer is ramping but not yet paying back

What Comes Next

Appendix A named the layers, the vendors, and the picks. Appendix B decomposed each layer into the dollar lines and named the team that operates them. Appendix C closes the practical-program guidance with the politics question: handling pushback from the “just use a database” camp, building executive sponsorship that survives a leadership transition, navigating the four-or-five-store anti-pattern from Part 10 when each store has an organizational owner with budget authority, and answering the year-one objections that every KG program faces before the first SPARQL template ships. Part 12 wraps the series with what building a knowledge graph actually teaches you. Read this appendix as the cost map; read Appendix C as the political map; the Lakeside trilogy remains the worked example all three appendices refer back to. The CFO’s three Thursday-morning questions have an answer; the chief data officer’s three Friday-morning questions are next.

Sources & References

  1. Cutter Consortium: Knowledge Graph Implementation Costs and Obstacles(2024)
  2. Forrester Total Economic Impact Study of Neo4j (commissioned by Neo4j, conducted independently by Forrester)(2021)
  3. Neo4j Pricing (AuraDB and self-managed enterprise)(2026)
  4. Senzing Pricing (entity resolution SDK and managed service tiers)(2026)
  5. Senzing: Calculating Total Cost of Ownership for Entity Resolution(2024)
  6. Atlan: Collibra Pricing Explained(2026)
  7. Atlan: Best Data Catalog Tools (2026 Buyer's Guide)(2026)
  8. Microsoft Research: LazyGraphRAG sets a new standard for GraphRAG quality and cost(2024)
  9. Graph Praxis: The GraphRAG Cost Cliff: How $33,000 Became $33 in Eighteen Months(2025)
  10. AstraZeneca: Biological Insights Knowledge Graph (BIKG) on bioRxiv(2021)
  11. Databricks Customer Story: AstraZeneca (time-to-insight improvement)(2024)
  12. Glassdoor: Ontology Engineer Salary(2026)
  13. ZipRecruiter: Neo4j Developer Salary(2026)
  14. Built In: Senior Data Engineer Salary in US(2026)
  15. Glassdoor: Site Reliability Engineer Salary(2026)
  16. Ontotext: Success Stories in Financial Services Knowledge Graphs(2024)
  17. Neo4j: Customer Stories (Citi, Lloyds, eBay, AstraZeneca)(2026)
  18. SAP Completes Acquisition of Reltio(2026)

Stay in the loop

Get new articles on data governance, AI, and engineering delivered to your inbox.

No spam. Unsubscribe anytime.