Cybos: playbooks and casesOpen playbooks, cases and infrastructure guides for CEOs building AI-native mid-market and enterprise companies.https://cybos.ai/2026-10-28T00:00:00ZCyboshello@cybos.aiWestpac: 12,500 banker hours saved in the first month of operation, July 2026https://cybos.ai/cases/westpac-l3-01/2026-08-05T00:00:00Z2026-10-28T00:00:00ZWestpac: 12,500 banker hours saved in the first month of operation, July 2026, annualised at 150,000 hours (AFR, 5 August 2026).Foundation Risk Partners: $10M EBITDA uplift from two agent systems, Partners Group sayshttps://cybos.ai/cases/foundation-risk-partners-p2-02/2026-08-03T00:00:00Z2026-10-28T00:00:00ZFoundation Risk Partners: New-client policy processing cycle cut 94%; close rates doubled; 120 bps EBITDA margin uplift worth USD 10M; first use case went from concept to production in 14 weeks.C.H. Robinson: 2–3 minutes per ocean or air freight quote, down from one to three dayshttps://cybos.ai/cases/c-h-robinson-p2-01/2026-10-28T00:00:00Z2026-10-28T00:00:00ZC.H. Robinson: NAST adjusted operating margin 40.3% and Global Forwarding adjusted operating margin 36.6% in Q3 2026; annual token cost under $2M on a run-rate basis. Ocean and air price quotes that took a day (air) or two to three days (ocean) now take two to three minutes; missed-pickup agents autL3Harris: 10 minutes to scope a supply shortage, down from 3 months and 30 peoplehttps://cybos.ai/cases/l3harris-l2-02/2026-09-10T00:00:00Z2026-10-28T00:00:00ZL3Harris: Scoping which programs a glass shortage for circuit boards affects now takes ~10 minutes, versus ~3 months and ~30 people before; a fine-tuned open model on the same data beat a frontier model on country-of-origin monitoring at 95% lower model cost.Shopify: 96% lower estimated serving cost for one agenthttps://cybos.ai/cases/shopify-l4-01/2026-08-05T00:00:00Z2026-10-28T00:00:00ZShopify: Estimated serving cost fell from ~$27M/year on a frontier model to ~$1M/year (-96%); quality surpassed the frontier baseline on Shopify's judge; end-to-end latency -38% and time-to-first-token -19% at 350 req/min; ~14% fewer GPUs for the same traffic.Bridgewater Associates: 13.8x lower cost per task, with higher accuracy than the best frontier modelhttps://cybos.ai/cases/bridgewater-associates-l4-04/2026-06-30T00:00:00Z2026-10-28T00:00:00ZBridgewater Associates: Accuracy 84.7% vs 78.2% for the best frontier model (29.8% fewer mistakes) at 13.8x lower inference cost per task; frontier models averaged ~50% with simple prompts and stayed under 80% with expert prompts.Zurich Commercial Insurance: 80% less manual submission triage timehttps://cybos.ai/cases/zurich-commercial-insurance-l3-10/2026-09-22T00:00:00Z2026-10-28T00:00:00ZZurich Commercial Insurance: 80% reduction in manual submission triage time; digitisation accuracy from below 80% to 98%; live across more than 20 markets.AT&T: Up to 90% lower AI cost, AT&T sayshttps://cybos.ai/cases/at-t-l1-01/2026-07-23T00:00:00Z2026-10-28T00:00:00ZAT&T: AI costs reduced 'as much as 90%', 'already saving millions'; 45 billion tokens/day on average.Pinterest: <8% of closed-model cost per transactionhttps://cybos.ai/cases/pinterest-l1-02/2026-08-05T00:00:00Z2026-10-28T00:00:00ZPinterest: Cost per transaction with open models at less than 8% of comparable closed proprietary models, with 'superior performance for our use cases' (e.g., Pinterest Assistant).Uber: 52% lower cost per agent session since the June peakhttps://cybos.ai/cases/uber-l1-04/2026-08-27T00:00:00Z2026-10-28T00:00:00ZUber: Cost per session down 52% from its June 2026 peak, with the model held constant. Uber's uReview benchmark showed switching models improved F1 while sharply cutting cost per review.Databricks: 30%+ lower average task cost from smart routing, internal coding workhttps://cybos.ai/cases/databricks-l1-06/2026-10-22T00:00:00Z2026-10-28T00:00:00ZDatabricks: Unit costs down 'as much as 90%' in some cases; model switching alone 50%+; spend reports ~10%; Smart Router cut average task cost by more than 30% at roughly top-model quality (65% of Opus 5 cost per task on internal coding workloads); output close to doubled with the same team.Meta: 85% of cases resolved by versioned rules, LLM only for the resthttps://cybos.ai/cases/meta-l1-07/2026-06-25T00:00:00Z2026-10-28T00:00:00ZMeta: ~85% of traffic resolved deterministically in single-digit milliseconds (~40 ms with context assembly); LLM path handles ~15% at roughly 400x the compute cost and seconds of latency.DoorDash: 53.6% recall for a two-model reviewer vs 30.7% single-passhttps://cybos.ai/cases/doordash-l1-08/2026-07-06T00:00:00Z2026-10-28T00:00:00ZDoorDash: Production reviewer (Sonnet 4.6 scout + Opus 4.8 reviewer): 53.6% weighted recall vs 30.7% for a single-pass GPT-5.5 baseline, at $3.91/PR. Swapping the scout to open-weight Kimi K2.6: $2.35/PR at 45.8% recall.Replit: 65% lower cost at the same quality, Replit's testinghttps://cybos.ai/cases/replit-l1-09/2026-08-26T00:00:00Z2026-10-28T00:00:00ZReplit: Same output quality at 65% lower cost than the previous version of Max Mode, in Replit's testing.Ramp: 30% lower LLM cost on its own productshttps://cybos.ai/cases/ramp-l1-11/2026-08-19T00:00:00Z2026-10-28T00:00:00ZRamp: Cut Ramp's own LLM costs 30% while improving speed and intelligence, CTO Rahul Sengottuvelu said in July 2026.Grab: 500+ services behind one LLM gatewayhttps://cybos.ai/cases/grab-l1-12/2026-07-24T00:00:00Z2026-10-28T00:00:00ZGrab: 500+ services and 50+ MCP servers on the framework; billions of tokens per month through one gateway (no cost figure published).Targa Telematics: 3.4B+ tokens processed on-prem in seven weekshttps://cybos.ai/cases/targa-telematics-l1-13/2026-10-14T00:00:00Z2026-10-28T00:00:00ZTarga Telematics: First seven weeks: ~30,000 developer requests, 3.4B+ tokens processed, 65M tokens generated, ~150 engineers using it.Atos: 1% → 99% valid structured output after fine-tuning a small modelhttps://cybos.ai/cases/atos-for-a-german-public-sector-customer-l1-14/2026-10-27T00:00:00Z2026-10-28T00:00:00ZAtos: Fine-tuned Phi-4 Mini JSON validity rose from 1% to ~99% (GPT-4.1: 96%); on a 34-point rubric the best fine-tuned model (Qwen 2.5 7B, 30.8) edged GPT-4.1 with a detailed prompt (30.5); training needed at most 3 GPUs.Descript: 1–2 hours to evaluate a new model, down from a week or morehttps://cybos.ai/cases/descript-l1-16/2026-09-15T00:00:00Z2026-10-28T00:00:00ZDescript: 1 to 13 models in production (Aug 2026); time to evaluate a new model cut from a week or more to 1-2 hours.Ito: 55% lower median cost per pipeline run after moving to an open modelhttps://cybos.ai/cases/ito-l1-17/2026-07-13T00:00:00Z2026-10-28T00:00:00ZIto: Median cost per pipeline run -55%; weekly spend -28% while volume +8%. Trade-offs: median step 23 to 29 min; failed terminal states 2.4% to 4.5%; open model used 2.18x output tokens.Uniper: 27 days faster time-to-sign for candidateshttps://cybos.ai/cases/uniper-l2-01/2026-09-09T00:00:00Z2026-10-28T00:00:00ZUniper: Average time-to-sign for candidates cut by 27 days; screening phase shortened by 13 days. Before production, the same prompts run on raw HR-system data returned generic answers, while the process-grounded agent named Uniper-specific bottlenecks.CVS Health: 2 months for a 1% forecast-error gain that last took a yearhttps://cybos.ai/cases/cvs-health-l2-04/2026-08-26T00:00:00Z2026-10-28T00:00:00ZCVS Health: Weighted MAPE (forecast error) improved 1% in the 2 months after deployment, versus about a year for the previous 1% gain; the speaker values that at roughly $5–10M a year in reduced overtime and overstaffing. Data scientists had been spending ~20 hours a week acting as analysts.Kirkland & Ellis: Days → minutes for fund-formation analysis and drafting, the firm sayshttps://cybos.ai/cases/kirkland-ellis-l2-06/2026-08-03T00:00:00Z2026-10-28T00:00:00ZKirkland & Ellis: Analysis, discussion and drafting work that took a lawyer days now takes minutes (firm statement quoted on Palantir's earnings call).NVIDIA: 86.7% accuracy from a model trained on planners' decision recordshttps://cybos.ai/cases/nvidia-l2-07/2026-09-10T00:00:00Z2026-10-28T00:00:00ZNVIDIA: On a development set, the model trained on planner decision records reached 86.7% accuracy on allocation decisions, versus 55.5% for a large general model and 17.5% for the untrained base model.The Elmet Group: 10 apps on a legacy ERP in two monthshttps://cybos.ai/cases/the-elmet-group-l2-08/2026-09-10T00:00:00Z2026-10-28T00:00:00ZThe Elmet Group: CRM data integration that normally takes months was completed by Palantir's team in about two days (the executive thought perhaps two hours), the company says; ten operational apps delivered in two months.Uber: Thousands of internal tools exposed to agents with zero service changeshttps://cybos.ai/cases/uber-l2-10/2026-10-01T00:00:00Z2026-10-28T00:00:00ZUber: Hundreds of MCP servers and thousands of tools catalogued, exposed with zero changes to downstream services.Experian: 25% higher win rates in one division, vendor-reportedhttps://cybos.ai/cases/experian-l2-13/2026-09-30T00:00:00Z2026-10-28T00:00:00ZExperian: 25% lift in win rates.Cisco: 26% higher win rates in early-adopting teams, Gong reportshttps://cybos.ai/cases/cisco-l2-14/2026-09-30T00:00:00Z2026-10-28T00:00:00ZCisco: Early-adopting teams show roughly 32% larger deal sizes and 26% higher win rates, Gong reports.Unilever: Days → minutes for finance analysis, the company sayshttps://cybos.ai/cases/unilever-l2-15/2026-07-17T00:00:00Z2026-10-28T00:00:00ZUnilever: Finance analysis that took days now takes minutes; 1,200+ users targeted; multi-million-euro annual cost avoidance expected from customer analysis and credit-risk monitoring.Ondas: ~3 months for post-acquisition integration, the company sayshttps://cybos.ai/cases/ondas-l2-16/2026-09-10T00:00:00Z2026-10-28T00:00:00ZOndas: Post-acquisition integration cycles compressed from a typical 12–24 months to about 3 months.Atlassian customers: 20% more Jira work items completed by adoptershttps://cybos.ai/cases/atlassian-customers-l2-17/2026-08-06T00:00:00Z2026-10-28T00:00:00ZAtlassian customers: Rovo adopters complete 20% more Jira work items and create/edit 25% more Confluence pages than non-adopters.Stripe: 26% shorter median compliance review timehttps://cybos.ai/cases/stripe-l3-02/2026-06-26T00:00:00Z2026-10-28T00:00:00ZStripe: 26% reduction in median review handling time, with over 96% helpfulness rating from reviewers.Coinbase: 75% of pull requests created by agentshttps://cybos.ai/cases/coinbase-l3-04/2026-08-31T00:00:00Z2026-10-28T00:00:00ZCoinbase: 75% of all pull requests created by agents (Cursor customer story); some teams cut idea-to-production from 20 days to under 2; security reviews from three days to five minutes.DoorDash: 2,000 sessions auto-graded a day before launchhttps://cybos.ai/cases/doordash-l3-06/2026-06-30T00:00:00Z2026-10-28T00:00:00ZDoorDash: Quality signal grew from ~1 employee feedback report to 2,000 auto-graded sessions a day; agent quality score up 8 points and error rate nearly halved before nationwide launch; a manual regression test that took more than six hours now runs in ~20 minutes.Uber: 70%+ of pull requests attributed to agentshttps://cybos.ai/cases/uber-l3-07/2026-08-27T00:00:00Z2026-10-28T00:00:00ZUber: More than 70% of pull requests attributed to agents; 30K+ agent skill executions a day; cost per session down 52% from its June 2026 peak.Spotify: Months → days for codebase-wide migrations, run by one engineerhttps://cybos.ai/cases/spotify-l3-08/2026-06-03T00:00:00Z2026-10-28T00:00:00ZSpotify: Migrations that took hundreds of teams weeks or months now take one engineer a few days; the most recent Java migration across backend services took three days.Siemens: 90% of inbound calls handled without a personhttps://cybos.ai/cases/siemens-l3-11/2026-07-07T00:00:00Z2026-10-28T00:00:00ZSiemens: 90% of inbound calls handled autonomously (resolved or correctly routed) without human intervention.LG Uplus: 70% → 96% agent accuracy through evaluation-driven developmenthttps://cybos.ai/cases/lg-uplus-l3-12/2026-08-17T00:00:00Z2026-10-28T00:00:00ZLG Uplus: Agent accuracy raised from 70% to 96% (+26 points).Getmobil: 60% lower contact-centre costhttps://cybos.ai/cases/getmobil-l3-13/2026-06-09T00:00:00Z2026-10-28T00:00:00ZGetmobil: 100% of inbound contacts handled by AI agents first; 75% resolved end to end without escalation; contact-centre costs down 60%; response times 3x faster.TwinMed: 3,200 purchase orders a month supported by agentshttps://cybos.ai/cases/twinmed-l3-14/2026-09-30T00:00:00Z2026-10-28T00:00:00ZTwinMed: Agents support about 3,200 purchase orders a month across 260 suppliers.Commerzbank: 20 h → 1 h manual document-review workhttps://cybos.ai/cases/commerzbank-l3-15/2026-10-08T00:00:00Z2026-10-28T00:00:00ZCommerzbank: Manual document-review work cut from 20 hours to 1 hour.Shopify: ~7x faster responses from a post-trained open modelhttps://cybos.ai/cases/shopify-l4-02/2026-10-21T00:00:00Z2026-10-28T00:00:00ZShopify: Judge score rose from 37.7 (base) to 74.5 after six loops and ~61,000 repaired examples, vs 73 for the frontier model; median full response 0.87 s vs 6.50 s (~7x faster); ~90% lower serving cost.Travelers: Own LLM beat commercial models in internal tests, no figures publishedhttps://cybos.ai/cases/travelers-l4-03/2026-06-30T00:00:00Z2026-10-28T00:00:00ZTravelers: In testing on tens of thousands of insurance questions it outperformed commercial models on quality, cost and speed (no figures disclosed).Harvey: 0.853 → 0.913 rubric pass rate after RL post-traininghttps://cybos.ai/cases/harvey-l4-05/2026-06-22T00:00:00Z2026-10-28T00:00:00ZHarvey: Rubric pass rate 0.853 to 0.913, above Opus 4.8 Max and GPT-5.5 xhigh; all-pass rate 0.059 to 0.126; grading cost cut 40x by a smaller model with batched criteria.Harvey: ~2x held-out LAB legal tasks completed vs base Kimi K3https://cybos.ai/cases/harvey-l4-06/2026-08-20T00:00:00Z2026-10-28T00:00:00ZHarvey: Harvey Tenet (Kimi K3 post-trained with Fireworks) completes almost twice as many held-out LAB tasks as base Kimi K3 (+9 pp all-pass rate). Separately post-trained Harvey models: a GLM-5.2 Review Table model (with Applied Compute) improves answer quality by 3.6 points and citation quality byFin: 16.5% fewer hallucinations than its previous own modelhttps://cybos.ai/cases/intercom-fin-l4-07/2026-10-07T00:00:00Z2026-10-28T00:00:00ZFin: Versus Apex 1.0: hard resolution +8.6%, resolution +1.7%, hallucinations -16.5%; live for every Fin customer.Kasisto, deployed at 40+ financial institutions: 20-50x lower cost per query than GPT-4.1https://cybos.ai/cases/backbase-l4-08/2026-07-30T00:00:00Z2026-10-28T00:00:00ZKasisto, deployed at 40+ financial institutions: Answer quality 6.21 vs 5.72 for GPT-4.1 on banking QA; ~$0.001 per query, 20-50x cheaper and 3-5x faster than GPT-4.1; training cost ~$1,800; deployed at 40+ financial institutions.Heidi Health: 25 s → 7 s scribe latency on a post-trained open modelhttps://cybos.ai/cases/heidi-health-l4-09/2026-10-22T00:00:00Z2026-10-28T00:00:00ZHeidi Health: Scribe latency fell from 25 s to 7 s (3.5x), with POC to production in 4 weeks; post-trained Nemotron 3 Ultra wins ~54% of live clinician comparisons against the frontier model in production (base: ~42%).Factory: 0.698 recall vs 0.588 for the best frontier model at a 5% false-alarm budgethttps://cybos.ai/cases/factory-l4-10/2026-07-01T00:00:00Z2026-10-28T00:00:00ZFactory: At a strict 5% false-alarm budget, recall 0.698 vs 0.588 for GPT-5.5 and 0.574 for Opus 4.8 (non-overlapping 95% CIs), at lower cost and latency.Clay: Up to 47% lower inference cost at retained qualityhttps://cybos.ai/cases/clay-l4-11/2026-10-27T00:00:00Z2026-10-28T00:00:00ZClay: Inference cost down up to 47% at retained quality; 48.2% savings at the 92.5% quality point vs always using the frontier model; can route hundreds of millions of queries a month in a fraction of a second per decision.