Build-vs-Buy framework for AI agents (May 2026)
Source report: /tmp/build-vs-buy-ai-agents-2026-05-12.md (333 lines, ~29KB).
TL;DR — pitch script for Init Intelligence’s buy-case
“Building a production-grade AI ITSM agent in-house costs $2-4M/yr (3-5 ML engineers + platform + security + LLM + infra), takes 12-24 months to production-quality, and Gartner forecasts 40% of these projects will be canceled by 2027. The build successes in the public record are all hybrids (buy-runtime + build-integration), not pure builds. Klarna built on OpenAI without a human-loop architecture and rolled it back. Init Intelligence ships SOC 2 + ISO 42001 + 6-week deployment + 123% NRR — for 5-40× less TCO over 3 years.”
TCO decomposition
Build cost (1000-employee enterprise, 8-15 agents, 3-year horizon)
| Line item | Annual cost |
|---|---|
| 3-5 ML engineers | 1.5M fully loaded |
| 1-2 platform engineers | 600K |
| 1 security engineer | 350K |
| 1 designer | $200K |
| 1 product manager | $250K |
| Total headcount | $1.5-3M/yr |
| LLM token spend | 600K/yr |
| Vector DB + observability + tools | 400K/yr |
| Infrastructure | 500K/yr |
| Total infra/ops | $0.7-1.5M/yr |
| Total annual | $2-4M/yr |
| 3-year total | ~$11M for 1000-employee scope |
Buy cost (Init Intelligence band)
- Mid-market band per pricing-benchmarks-ai-itsm-2026: $100-400k/yr
- 3-year total: ~$1.5M
TCO spread
5-40× build-vs-buy, not 2×. This is the single most important slide for the buy-case pitch.
Gartner cancellation forecast (highest-authority anchor)
- Gartner (Aug 2025): 40% of agentic-AI projects canceled by 2027.
- Reason: operationalization failure, not model failure.
- Adds **~11M direct build cost.
- External authority strongest available for the buy-case.
Build failure cases (verified)
Klarna (canonical)
- Built on OpenAI + custom orchestration with no human-loop architecture.
- CEO admitted “we went too far.”
- Active rehiring confirmed by Bloomberg (May 2025) and CX Today (early 2026 redeployments from marketing/eng/legal into CS).
- Verified as build failure (not a vendor failure) — important distinction.
11x.ai (vendor cautionary, not build cautionary)
- 70-80% per-employee churn.
- ARR inflation 4.7× (3M post-3-month).
- Useful as buy-side vendor-risk acknowledgment, not as a build-failure example.
- See founder-operating-playbooks-2026.
”Build success” gap (sales asset)
Build-success cases for Init Intelligence’s actual ICP (500-5000 employee mid-market IT teams) essentially don’t exist in the public record. The success stories that get cited (Morgan Stanley, Goldman) are buy-of-runtime + build-of-integration hybrids, not pure builds. The absence is itself a sales asset — when a CIO asks “who has built this successfully?” the honest answer is “nobody published below the F500 tier.”
5 hybrid patterns
| # | Pattern | Best for |
|---|---|---|
| 1 | AI co-pilot embedded by vendor + custom integrations by customer team | Most common; low risk |
| 2 | Vendor provides orchestration; customer brings own LLM key | Cost-sensitive |
| 3 | Vendor + customer fine-tune on customer data | Privacy-sensitive |
| 4 | Vendor multi-tenant runtime + customer dedicated instance | Compliance-heavy |
| 5 | BYO model + Init Intelligence governance + customer data | Most CIO-friendly — productize as distinct SKU |
CIO decision-criteria framework
| Dimension | Buy | Build | Hybrid |
|---|---|---|---|
| Time-to-value | weeks | months-years | months |
| TCO 3-year | high opex (~$1.5M) | high capex + opex (~$11M) | medium |
| Governance maturity | vendor-supplied | DIY | mixed |
| Customization | constrained | full | medium |
| Talent risk | low | high | medium |
| Vendor lock-in | high | none | medium |
| Gartner cancellation risk | low | 40% by 2027 | medium |
Init Intelligence sales materials playbook
- Ship the TCO calculator. No AI ITSM competitor has published one. Doubles as a pricing-transparency credibility signal aligned with the published-per-resolution-SKU wedge from pricing-benchmarks-ai-itsm-2026. Highest-leverage single sales asset.
- Lead with the Gartner 40% cancellation forecast. Highest-authority external anchor; defuses build-side bias.
- Klarna is the canonical build-failure slide. Anchor with primary source (Bloomberg + CX Today).
- Productize Hybrid Pattern 5 (“BYO model + Init Intelligence governance + customer data”) as a distinct SKU.
- “6-week deployment vs 12-24 month build” is the time-to-value slide.
- “Day 1 SOC 2 + ISO 42001 + NIST AI RMF” vs. “DIY governance maturity” is the compliance slide. See ai-itsm-compliance-roadmap-2026.
- 123% NRR top-quartile (ICONIQ) vs. 70-80% churn (11x) is the vendor-quality slide. See sales-cycle-dynamics-2026.
Honest verification gaps (per source report)
- Gartner 40% is a forecast not a measurement.
- Klarna -22% CSAT is harder to triple-verify than the rehiring/redeployment fact.
- Vendor TCO numbers above Atomicwork floor / Aisera median carry verification caveats from pricing-benchmarks-ai-itsm-2026.
- Specific Devin customer name and specific Klarna headcount figure softened in source report (no primary source).
Related
- pricing-benchmarks-ai-itsm-2026 — buy-side TCO floor/ceiling
- oss-agent-infra-2026 — 10-component build stack
- sales-cycle-dynamics-2026 — 6-week deployment + 123% NRR
- workforce-displacement-ai-labor-2026 — Klarna canonical cautionary tale
- founder-operating-playbooks-2026 — 11x pattern + HuRT + escalation-is-free
- ai-itsm-compliance-roadmap-2026 — governance Day-1 argument
- initlabs-engineering-build-playbook-ai-itsm — inverted build sequence
- Init Intelligence