Chapter 10
Appendix: benchmarks by lever, the figures not to cite, the modeling assumptions, and the glossary
The numbers behind the cards, in one place, each with its grade and its source chapter. The sources themselves are listed at the end of each chapter, with the URL and the date it was opened.
Chapter 10 of 10Appendix9 min
Benchmarks by lever
Grades: S strong (independent study, regulator, statute, primary filing or press release with method); M medium (vendor case with a named customer, or an independent figure via a secondary outlet); W weak (vendor claim without a named customer, or argued rather than measured).
| Lever | Benchmark | Figure | Grade | Chapter |
|---|---|---|---|---|
| P1 | Refactoring and duplication in AI-era code (GitClear, 623M changes) | Refactoring −70%, duplication +81% | S | Before close |
| P1 | AI-generated code failing security tests (Veracode, 100+ models) | 45%; Java 72% | S | Before close |
| P1 | Lower-middle-market tech diligence cost (ACG) | $25K to $75K; PE-grade to $150K | S | Before close |
| P2 | Contract review time (Harvey, Bridgewater) | 2 days to 2 hours | M | Before close |
| P2 | Legal research tool error rates (Stanford) | 17% to 34%, on the wrong task | S | Before close |
| P3 | Commercial diligence cost (DiligenceSquared) | $50K vs $500K to $1M | M | Before close |
| P3 | Letter of intent to close (SEG) | 5 to 7 weeks; customer calls ≈1 week | M | Before close |
| P4 | Public software multiples (Meritech, Apr 2026) | Median 3.2x ARR; AI-tailwind names fell more | S | Before close |
| P5 | AI in diligence and valuation (KPMG, n≈700) | 56% | S | Before close |
| P5 | PE benefits within 12 months / efficient implementation (FTI, n=555) | 66% / 31% | S | Hub |
| P5 | Required EBITDA growth for 2.5x (Bain) | 10% to 12% a year; holds ≈7 years | S | Before close |
| P7 | HIPAA readiness and zero retention on one organization (Anthropic) | Cannot coexist; 30-day retention | S | Before close |
| P8, R1 | Enterprise list-price increase (Salesforce, Aug 2025) | +6% average | S | Before close |
| P8, R1 | SaaS spend growth on flat app count (Zylo) | ≈ +8% in 2025 | S | Before close |
| R1 | Regulatory-workflow pricing power (Workiva Q2 FY26) | GRR 97%, NRR 111%, premium tiers >20% | S | Revenue |
| R2 | AI adopters' ARR growth (Atlassian Q4 FY26) | >2x non-adopters, no matched cohort | S | Revenue |
| R2 | Seat expansion alongside AI credits (Figma Q2 26) | Two-thirds of renewals added seats | S | Revenue |
| R2 | Ambient documentation time saved (JAMA, 1,809 clinicians) | 16.0 min per 8 hours | S | Revenue |
| R3 | Hybrid pricing adoption (Growth Unhinged, 230+) | 37%; AI credits 29% | M | Revenue |
| R3, R11 | Deflection and net adds in the same quarter (HubSpot Q2 26) | 72% resolved; net adds cut to 5 to 6K | S | Revenue |
| R4 | ARR per success manager (Vista, n=17) | $5.4M to $6.7M | M | Revenue |
| R5 | Retention targeting rule (Ascarza, JMR) | Highest-risk are not the best targets | S | Revenue |
| R5 | Bootstrapped SaaS retention medians (SaaS Capital) | GRR 91%, NRR 103% | S | Revenue |
| R6 | Questionnaire turnaround (Conveyor, Intellistack) | −70% | M | Revenue |
| R7 | AI Overviews and position-1 clicks (Ahrefs, 300K keywords) | −58% | S | Revenue |
| R7 | Zero-click searches (SparkToro/Similarweb, Jan to Apr 2026) | 68.01% | S | Revenue |
| R7 | AI-written vs human cold email (head-to-head tests) | 1.4% vs 2 to 4% positive | M | Revenue |
| R8 | Net payments take (Toast Q3 25) | 49 bps; fintech 61 bps | S | Revenue |
| R8 | Fintech vs SaaS gross margin (Toast FY25) | ≈36% vs 79% | S | Revenue |
| R8 | Usage revenue and take (ServiceTitan FY25) | 22.5% of revenue; ≈25 bps of GTV | S | Revenue |
| R8 | Payments share and penetration (EverCommerce 10-K) | ≈20% of revenue; ≈13% attach | S | Revenue |
| R8 | Facilitator build (three vendors) | 12 to 24 months; ≈$2B volume break-even | W to M | Revenue |
| R9 | Non-payments fintech (Toast) | $58M GP, 11 bps in a quarter | S | Revenue |
| R9 | Embedded insurance economics (Authentic via Tidemark) | 5 to 25% of premium; >3% attach | M | Revenue |
| R9 | Shopify Capital credit drift | Current 93.7% to 91.9% | M | Revenue |
| R10 | Regulated-vertical margin (Veeva FY26) | 44.9% non-GAAP operating margin | S | Revenue |
| G1 | Agent resolution (Intercom, 12,000+ customers) | 76%, incl. soft resolutions | M | Gross margin |
| G1 | Containment vs resolution gap (Lorikeet) | ≈20 points | W | Gross margin |
| G1 | Assistant productivity (Brynjolfsson et al., 5,179 agents) | +14%; +34% novices | S | Gross margin |
| G2 | Implementation hours (Workday Deployment Agent) | "Estimated 30%", target 50% | M | Gross margin |
| G3 | Services vs subscription gross margin (Guidewire Q4 FY26) | 12.5% vs 74.5% | S | Gross margin |
| G4 | Cloud waste (Flexera, 750+) | 29% | S | Gross margin |
| G5 | Inference price decline for fixed capability (Epoch, to Dec 2024) | 9x to 900x per year | S | Gross margin |
| G5 | Tokens per answer with retrieval (Atlassian) | −48% | S | Gross margin |
| G5 | Gross margin while shipping AI (Figma Q2 26) | 85%, +2.5 pts sequential | S | Gross margin |
| G5 | Target AI product gross margin (Growth Unhinged) | ≈50% | M | Gross margin |
| O1 | Experienced developers with AI (METR RCT) | 19% slower; believed 20% faster | S | Opex |
| O1 | Field experiment, 4,867 developers (Cui et al.) | +26.08% tasks | S | Opex |
| O1 | Downstream effects (Faros, 22,000 developers) | +54% bugs, 5x review time | M | Opex |
| O1 | R&D share (Vista, n=54) | 21.9% to 19.2% | M | Opex |
| O2 | AI-authored migration changes (Google) | 80% of landed changes; reviewer bottleneck | S | Opex |
| O4 | S&M share (Vista, n=55) | 30.4% to 25.6% | M | Opex |
| O6 | Monthly close (APQC) | Median 6 days; top ≤5 | S | Opex |
| O6 | Close reduction (Choi and Xie, 79 companies) | 7.5 days | M | Opex |
| O7 | No-touch invoices (Vic.ai named cases) | 72% to 78% | M | Opex |
| O8 | Legal AI productivity (Schwarcz et al., RCT) | +50% to +130% on 5 of 6 tasks | S | Opex |
| O8 | Outside-counsel spend direction (CLOC, n=135) | Expected increases 58% to 37%; no decline | S | Opex |
| O9 | Compliance time per year (Vanta State of Trust, n=3,500) | 12 weeks, up from 11 | M | Opex |
| O10 | Time to hire, high volume (Paradox, 7-Eleven) | >10 days to <5 | M | Opex |
| O11 | Employee assistant adoption (Moveworks, Databricks) | 10% to 73% | M | Opex |
| O11 | Ticket resolution with generative AI (SolarWinds, 60,000+ records) | 17.8% faster; top 54.3% | M | Opex |
| O12 | Unused licenses (Zylo, 40M licenses) | 36%; $9,455 per employee | S | Opex |
| O12 | Savings on managed spend (Forrester/Zip) | 3.3% | M | Opex |
| O14 | SG&A median vs first quartile (Hackett, 1,000 companies) | 16.2% vs 7.8% | S | Opex |
| C1 | DSO (APQC) | Median 38 days; top ≤30 | S | Cash |
| C1 | DSO reduction (Nucleus/HighRadius) | 80 to 58 days | M | Cash |
| C1 | Retry logic vs account updater (Zapier, 2021) | 1.24% vs 2.76% uplift | M, dated | Cash |
| C2 | Appeal overturn rates (KFF) | MA 67%, Medicaid 47%, Marketplace 43% | S | Cash |
| C2 | Post-acute denial rate (npj Digital Medicine) | 10.0% to 22.7% | S | Cash |
| C2 | Revenue-cycle AI adoption gap (HFMA/AKASA, n=519) | 64% large vs ≈20% small | S | Cash |
| X2 | EU AI Act dates (omnibus in force Jul 27 2026) | Art. 50 Aug 2 2026; high-risk Dec 2 2027 | S | Risk |
| X2 | Vendor liability for AI screening (Mobley v. Workday) | Nationwide collective certified | S | Risk |
| X2 | Accuracy-claim enforcement (FTC/Workado) | 98% claimed, 53% actual | S | Risk |
| X4 | Add-on share of software deal value (PitchBook, 2026) | ≈45%; platforms 41% | S | Risk |
| X6 | Multiples by growth band (SEG 4Q25) | ≤10%: 2.4x; 10 to 20%: 5.8x; 20 to 30%: 12.7x | S | Risk |
| X6 | Multiples by Rule of 40 band (SEG 4Q25) | 20 to 30: 6.0x; >40: 14.0x | S | Risk |
| X6 | AI product premium (McKinsey, n=471) | ≈130%; no growth control | S, caveated | Hub |
Figures that circulate and are not cited
These were checked and either do not exist at the claimed source, are marketing content without method, or contradict a primary source: the "Gartner 20 to 30% average, 40 to 60% best-in-class" service-desk deflection benchmark; the "65% deflection" and "75% deflection" vendor figures absent from the pages cited for them; the "integration costs run 1 to 3% of deal value" line for software buy-and-build; the payments-versus-subscription multiple ranges from a banker's marketing page; the "Canva 90% AI task cost reduction" from an unsourced secondary; Software Equity Group's above-30%-growth multiple band, which inverts on sample noise; the Faros "+210% tasks" figure that appears in neither Faros report; incident.io's "18% MTTR" figure, which belongs to SolarWinds; Vanta's "85% of evidence automated," which carries no citation; and any AI governance program cost, because no credible one has been published. Amazon's 4,500 developer-years and $260M measure different things and are cited separately.
Modeling assumptions
Every figure in the five archetype profiles and bridges, every "at close" value in the metrics blocks, the 90% flow-through of price, the 25% to 35% of labor value pricing rule for modules, the $38 cost per ticket and 55-day DSO in the case study, the 42 to 50 basis point take and 40% to 44% payments gross margin in archetype C, the 28% credit share and 50% AI-product gross margin in archetype D, and the services conversion path in archetype E are modeling assumptions to be replaced with an acquired company's own. The case study's $630,000 of working capital is 17 days × $13.55M ÷ 365. The bridges were computed in one script so that the lines add; the script is part of the working files.
Glossary
| Term | Definition used | Source |
|---|---|---|
| Lever | An intervention an owner can make that moves a specific line by a mechanism that can be stated in a sentence and checked in a quarter | This playbook |
| Deploy / Reshape / Invent | Proven purchasable tools on existing workflows; reorganizing a function because the tools changed its shape; new revenue from AI products, pricing and rails | Kunal's framing, borrowed from consultancy usage |
| Workflow vs agent | Workflows orchestrate models and tools through predefined code paths; agents dynamically direct their own process and tool use | Anthropic, "Building effective agents" |
| True resolution vs containment | Resolution: the customer confirms or does not return within a window; containment: no human handoff. Containment runs about 20 points above resolution | Intercom; Lorikeet |
| Net take rate | Gross profit from payments as a share of processed volume, in basis points; distinct from gross take (revenue as a share of volume) | Toast Q3 2025 prepared remarks |
| Payment facilitator | A platform registered with the card networks that sets merchant pricing and pays the processor cost-plus, as opposed to referring merchants for a revenue share | Stripe Connect pricing; Stax; Tilled |
| Gross transaction volume | The total value of customer transactions the platform enables, whether or not it processes them; a superset of processed volume | ServiceTitan filings |
| Business associate | A person who creates, receives, maintains or transmits protected health information on behalf of a covered entity | 45 CFR 160.103 |
| Safe Harbor de-identification | Removal of the eighteen identifier classes at 45 CFR 164.514(b)(2); expert determination is the alternative | 45 CFR 164.514(b) |
| Risk analysis | "An accurate and thorough assessment of the potential risks and vulnerabilities to the confidentiality, integrity, and availability of ePHI" | 45 CFR 164.308(a)(1)(ii)(A) |
| Article 50 transparency | The EU AI Act duty, in force from August 2, 2026, to tell people they are interacting with an AI and to label AI-generated content | EU AI Act; Digital Omnibus on AI |
| Excessive agency | The vulnerability enabling damaging actions from unexpected or manipulated model outputs; root causes are excessive functionality, permissions, autonomy | OWASP LLM06:2025 |
| Evaluation set | A test set drawn from real cases with a passing threshold, re-run on any model, prompt or tool change; three levels: assertions, trace review, A/B | Hamel Husain |
| Change failure rate | Share of deployments causing a failure in production | DORA 2025 |
| FinOps | An operational framework and cultural practice maximizing the business value of technology through engineering, finance and business collaboration | FinOps Foundation |
| DSO | Average days to collect payment: average receivables ÷ (annual sales ÷ 365) | APQC |
| NRR / GRR | Second-period revenue from first-period customers ÷ first-period revenue; GRR caps each customer at its first-period revenue | SaaS Capital |
| Rule of 40 | Revenue growth rate plus profit margin ≥ 40% | Bessemer |
| Quality of earnings | The buyer's accounting review of the target's reported earnings, revenue recognition and adjustments | Standard practice |
| Data rights | The contractual permission to use customer data for a stated purpose, including product improvement and model training | This playbook; P2 |
| Agent register | The list of every production agent with its owner, permission scope, last passed evaluation, retention term, cost and measured value | This playbook; O13 |
| Kill date | The date an initiative is stopped if its target metric has not moved from its baseline | This playbook |
Boring AI
AI for manufacturers, operators and service businesses — not startups chasing hype. Every other week.
Your address is used only to send this newsletter. No sharing, no selling, no tracking pixels. Unsubscribe from any issue.