Revenue Strategy
What Anthropic's S-1 Tells Us About AI ROI

Anthropic’s S-1 tells an ambitious story about AI: software that changes how work gets done, across businesses with very different customers, costs, and measures of success.
The most useful way to read that story is to follow the work. What changed? What did people do with the improvement? Did it produce additional output, better service, avoided spending, or more profit?
We reviewed the 20 dedicated customer panels, covering 21 named organizations, in Anthropic’s October 26, 2026 filing. The final panel includes two businesses. Below, we group them into ten industries so the economics are easier to compare. These are our classifications, not Anthropic’s revenue segments. S-1, pp. 208–227.
There is substantial evidence of useful change. There are also different kinds of evidence: adoption, time recovered, faster delivery, commercial results, and a reported financial return. They cannot be added together into an average customer ROI.
Capacity is valuable. Capacity is not automatically cash value. The distinction matters to the buyer deciding whether to expand a deployment and to the supplier whose valuation rests on customers continuing to find it worth buying.
Design and remodeling: put the time back into what you sell
KBF Design Gallery offers the clearest human-scale example. The company reports cutting seven hours from writing a contract scope. Across 70 projects a year, that adds up to 490 hours of capacity for one employee. The lead designer puts that time back into design work. The same panel reports 32% sales growth since adopting Claude in 2025, without separating AI’s contribution from other factors. S-1, p. 227; one employee’s annual capacity estimate and company sales growth.
KBF Design Gallery · Design and remodeling
490 hours
Recovered each year
7 hours per scope × 70 annual projects. One employee’s estimated capacity.
Back into design work
Reported use of the time
The lead designer spends the recovered time designing.
32%
Reported sales growth
Since adopting Claude in 2025. The panel does not isolate AI’s share.
Inferred route to value: more design capacity could support more or better projects, with additional revenue and contribution if that work is sold and delivered.
What the filing leaves out: the extra projects, revenue, or profit attributable to the recovered time, and the full AI cost. The dashed connection is an unquantified link, not a measured conversion of hours into sales.
Source: Anthropic S-1, p. 227, October 26, 2026. The 490 hours are derived from the panel’s inputs; redeployment and sales growth are reported by the panel. No payroll saving or causal revenue estimate is asserted.
This gets beyond an empty calendar slot. The recovered time has a named destination: design, the work KBF sells. More design capacity could support additional projects, better projects, or a less overloaded designer. Additional profitable work would be one route to a financial return.
The filing gives us both redeployment and commercial growth. It does not quantify how much of that growth came from the recovered hours, or the contribution after delivery costs. That is the next measurement. Multiplying the hours by a wage rate would value capacity; it would not establish a payroll saving or explain the sales result.
The industry map
One dot per named organization. Our industry classification of 21 organizations in 20 selected case panels, not revenue mix or market share.
See the organizations by industry
- Software
- Legora · Mozilla · Salesforce · SAP
- Commerce platforms
- DoorDash · Mercado Libre · Rakuten · Shopify
- Financial services
- AIG · Jane Street · PayPal · Travelers
- Life sciences
- Bristol Myers Squibb · Lilly
- IT and consulting services
- Accenture · UST
- Telecommunications
- Verizon
- Retail
- Walmart
- Hospitality
- Shy Bird
- Design and remodeling
- KBF Design Gallery
- Humanitarian services
- International Rescue Committee
Legora is grouped as software; its study concerns law firms. SAP is software; its featured beneficiaries are consultants. Accenture’s featured broadband client is a separate telecommunications beneficiary.
The cases below all come from the October 26 filing unless a supplemental source is identified. Figures are reported by the filing or its cited sources; we have not independently audited the deployments. Where the panel omits a sample or comparison period, the result should be read as a bounded example. Missing public detail does not mean the customer has no private measurement.
For each industry, the useful question is what connects the operating improvement to the outcome the organization wants.
Software: more work completed, with different routes to value
Legora, serving legal services, supplies one of the most explicit bridges from time to revenue. The filing cites 4.3 recovered nonbillable hours per lawyer each week and $6.9 million in annual billing per 100 lawyers. The underlying report draws on interviews at 31 customer firms across 14 countries and describes the billing as potential, using 46 working weeks. It is not a ledger of collected fees. S-1, p. 214. Legora/Ari Kaplan report, p. 6; undated, retrieved October 26.
The opportunity depends on demand and contracts. A law firm with work waiting can turn available hours into billable matters. A fixed-fee team may improve its margin. An in-house team may avoid outside counsel spending. Each route requires a different calculation; none makes all recovered time equivalent to cash.
Salesforce reports 79% more pull requests merged per developer and a 151% rise in its separate “Effective Output” measure. Its supplemental engineering account dates the comparison to April 2026 versus April 2025 and explains that Effective Output is a machine-learning-based score. These are engineering-output measures, not revenue or profit. S-1, p. 220. Salesforce, May 27, 2026; detailed task sample not supplied.
The next question is what the additional accepted code accomplishes: a useful feature delivered sooner, maintenance completed, or work that previously sat in a backlog. Code volume alone cannot tell us which happened.
SAP’s Joule for Consultants panel reports 25 hours saved per consultant per month and up to 14% faster project delivery, without a consultant sample or measurement dates. The beneficiary is the consulting ecosystem, not simply SAP’s own payroll. Contract terms decide whether faster work becomes lower client bills, room for another engagement, or stronger fixed-fee margins. S-1, p. 221.
Mozilla shows another kind of output. It shipped fixes for 22 previously unknown vulnerabilities found during a two-week evaluation, then for an additional 271 found by a later model. The fixes reached Firefox 148 and 150 respectively. This is completed security work, even without a dollar estimate. Valuing it financially would require evidence about exposure and expected harm, not a presumed breach for every bug. S-1, p. 217; two distinct batches.
Commerce platforms: faster delivery has to meet customer demand
DoorDash reports that major-product launch timelines fell by more than half. The panel does not identify the number of launches or establish that their scope was comparable. Even so, the proposed route is recognizable: customers can use a product sooner. The financial question is whether that brings forward profitable demand, creates additional demand, or simply changes the release date. S-1, p. 211; Claude Code rollout in 2025, Cowork rollout in 2026.
Mercado Libre reports that, in the second quarter of 2026 versus a year earlier, code submissions rose 110% and deployments nearly 75%. It also reports a one-third reduction in the monthly financial close. These are useful operating signals, but the code figures are not normalized into a matched productivity measure per developer. A faster close can improve management’s information without removing an equivalent fraction of finance payroll. S-1, p. 216; underlying task and staffing samples not supplied.
Rakuten reports reducing new-feature time to market from 24 to five working days. More useful than the headline alone is its described practice of measuring each agent job for task completion and cost per task. That links spending to finished work. The missing pieces for a reader are task volumes, quality, rework, and the value of the features delivered. S-1, p. 219; comparison window and feature sample not supplied.
Shopify places the benefit further downstream, with merchants. Its panel says daily active merchants using Sidekick more than tripled year over year in 2026, and describes new merchants reaching a first sale in days instead of weeks. It does not disclose the first-sale cohort or a matched comparison. Claude is one of the models powering Sidekick, so the result belongs to a broader product and workflow. S-1, p. 222.
Across these platforms, the commercial prize is useful products reaching real demand sooner. Faster production can help, but distribution, adoption, and contribution after costs still decide the return.
Financial services: throughput matters when the business is attractive
AIG’s Lexington Middle Market Property business reports 30% more quoted submissions, 55% shorter time to quote, and 40% more submissions bound, as of March 31, 2026. Those measures reach beyond document processing into the underwriting funnel. S-1, p. 209; baseline period and submission counts not supplied.
For an insurer, more business is valuable when it is good business. The economic bridge runs through incremental premium, expected claims, and expenses. A 40% increase in bound submissions is not a 40% profit increase. Faster review may help produce the extra business; counting both a wage value for that time and the full financial benefit of the resulting business could double count the same improvement.
PayPal reports an average 45 minutes saved per week for users of a custom sales plugin introduced in August 2025. That figure concerns sales-plugin users, not all of the company’s more than 12,000 monthly active Claude Enterprise users. The sales sample and measurement method are not supplied. The next step is to trace the time into useful selling work and then measure its result. S-1, p. 218.
Travelers reports a 50% reduction in time to merge code. That may reduce waiting and let improvements reach production sooner. Elapsed merge time, however, is not the same as active engineering labor. Without that distinction, turning the percentage into payroll savings would create precision the panel does not support. S-1, p. 223; merge count, baseline duration, and measurement interval not supplied.
Jane Street offers a more exploratory example: one researcher estimates doing about 50% more analysis than a year earlier. That is a report from one person, not a firm-wide productivity result or a trading return. Its potential lies in expanding the set of ideas investigated; the economic test is whether the added analysis changes decisions for the better. S-1, p. 213.
Life sciences: faster answers can improve decisions
Bristol Myers Squibb reports reducing the time to generate an answer to a real-world-data question from up to two months to five minutes or less. Its company-wide platform was selected in May 2026; the panel does not give a question count, quality comparison, or active staff effort. S-1, p. 210.
The opportunity is substantial: a question that was impractical to ask may become routine, and a team may reach a decision sooner. But two months of waiting is not two months of labor. Nor does a faster answer automatically advance a drug launch. The useful connection is whether a reliable answer removes a real decision bottleneck.
Lilly illustrates adoption at scale. More than 50,000 employees have access to Claude Enterprise. More than 85% of full-time employees in five heavily using functions actively engage; that is not a company-wide activity rate. Company-wide access arrived by April 2026, while the activity window is unspecified. S-1, p. 215.
That engagement suggests people find uses for the tools. It does not identify which use creates the largest return. The next useful unit of analysis is the decision or workflow, with its quality requirements and cost of delay.
IT and consulting services: a reported return, and who receives it
Accenture’s panel includes an unnamed U.S. broadband client reporting a sevenfold return on AI investment in the first year, alongside an 86% reduction in lead-validation cost and 55% faster speed to market. This is an explicit reported financial outcome. The filing does not expose the investment amount, benefit calculation, or whether “return” means gross benefit or net gain. It cannot safely be translated into a net ROI percentage. S-1, p. 208; client sample counts and calendar dates not supplied.
The result belongs to the broadband client, not Accenture’s own operations. Accenture separately describes expanding internal security coverage. Keeping those scopes distinct matters: a partner’s platform, implementation, and the client’s changes can all contribute to a result. The panel should receive credit for reporting a return while leaving readers clear about what they cannot reconstruct.
UST reports chip-validation cycle times falling 50–70%, including standard turnaround moving from four days to two. The panel does not disclose staffing, cycle counts, or the amount of hands-on work. Faster validation could support more iterations or remove a downstream delay. It does not establish that every engineer recovered two working days, or that a chip reached market sooner. S-1, p. 224.
Services economics also raise a distribution question. Who keeps the benefit: the provider through better margins, the client through lower prices, or both through more work completed? A faster process alone cannot answer that.
Retail: a bigger basket needs an incremental-profit test
Walmart reports that orders involving its Sparky assistant have a 35% higher average order value than orders without it. This is closer to a transaction than a time-saving estimate. The panel does not disclose order counts, the measurement window, or a matched comparison between shoppers. S-1, p. 226.
The promising interpretation is that the assistant helps people find and buy more of what they need. Testing that requires separating its effect from the characteristics of shoppers who choose to use it. Then the financial calculation needs merchandise margin, fulfillment, returns, and any purchases shifted from another channel. A bigger basket can create value; the full basket is not the benefit attributable to AI.
Hospitality: faster service creates an opportunity to serve
Shy Bird reports reducing brunch ticket times at one location by more than 60%, to 12 minutes, following a kitchen redesign informed by Claude and implemented with cooks. The filing supplies neither a ticket sample nor a comparison window. S-1, p. 227.
That is a meaningful service improvement. It can make the guest experience better and may let the restaurant serve more people. Additional sales still require demand, seats, and staff; additional profit requires contribution after the extra costs. The people who changed the kitchen are part of the explanation. The value comes from an operating change that was put into practice.
Telecommunications: broad adoption is a starting signal
Verizon’s panel says more than 33,000 engineers build with Claude Code and describes cybersecurity work with Anthropic beginning in May 2026. It does not disclose a task-level time saving, quality result, or financial return. S-1, p. 225.
The scale makes the deployment worth studying. It cannot supply the missing economics by itself. Different engineering workflows may create value through faster releases, better reliability, or new work becoming feasible. The next measurement is one of those outcomes, attached to a defined workload and full cost.
Humanitarian services: the destination can be better care
The International Rescue Committee reports 464 staff hours returned each week and an expansion from a 300-user pilot to roughly 2,400 seats between December 2025 and May 2026. The panel does not identify the population behind the weekly hours, so dividing them by the seat count would invent a measured average. S-1, p. 212.
Here, better service can be the intended return. Staff may spend more time supporting people or making better-informed decisions. That should be measured in the quality and reach of the service, alongside its cost. The organization’s overall reach is not a count of additional people served because of AI. A mission outcome can justify spending without being dressed up as payroll removed.
The common thread: decide what happens after the task gets faster
Work changes
Less time, more capacity, better quality
What improved in the workflow?
Someone captures a benefit
Useful output, avoided spending, or more contribution
What will you do with the improvement?
The benefit covers the full cost
AI fees, implementation, review, and operation
Is the result worth what it takes?
A conceptual framework, not measured financial results. Count each benefit once.
The industries change the arithmetic, but the discipline carries across them. Establish the operating change, identify where the improvement goes, and test the resulting benefit against the full cost.
Recovered hours may support more sold work. A shorter cycle may bring a useful product to customers sooner. A better answer may change a decision. A vulnerability fixed may reduce exposure. Those are different routes, and they need different evidence.
Count the economic result once. If freed time produces additional profitable work, do not add the wage value of those same hours to the profit. Include implementation, review, training, and ongoing operation alongside AI fees. And keep improvements in capacity or service visible even when their cash conversion remains unknown.
Why this matters to Anthropic’s valuation
Anthropic announced a $965 billion post-money valuation in its May 28, 2026 financing. That is a historical private-financing equity reference, not an IPO target or a valuation derived from these cases. Anthropic’s financing announcement.
Supporting an ambitious valuation over time requires more than impressive activity. Customers need compelling reasons to keep paying and expand their use. Additional economic output, avoided spending, better decisions, and valuable service improvements can supply those reasons.
Anthropic then has to earn a share of that benefit while covering delivery costs, research, and continuing investment. Competition can pass more of the benefit to customers. Heavy spending can absorb growing revenue. Customer ROI therefore helps support the cash flows investors value; it does not determine the company’s worth on its own. Valuation framework: Aswath Damodaran, NYU Stern.
The strongest reading of the S-1 is that AI is expanding what organizations can do. The work ahead is to turn that expanded capability into outcomes worth their full cost, repeatedly. Some cases already point toward that conversion. Others show the capacity or adoption from which it may develop.
For a buyer, the question to carry into the next review is simple: What will we do with the improvement, and what evidence will show that it created additional value? For Anthropic, customers’ answers are part of the foundation beneath its own value.
What this looks like with Serre
The S-1 panels make the point finance teams make in every budget review: a time saving is the start of a value argument, not the end of it. The case has to name where the improvement goes, put a range on what it is worth, and show the full cost beside it.
That is the discipline Serre builds into every deal. Serre keeps one durable value model for each account and projects each business case from it, with conservative, expected, and upside ranges and the assumptions in the open, so hours recovered are never presented as cash saved. No value engineer required.
Sources and companion data
This review covers the October 26, 2026 filing and the dated supplemental sources identified above. The cases are selected illustrations, not a representative success-rate sample. The complete source register preserves each panel’s measures, scope, and limitations.
Serre has no affiliation with Anthropic or the organizations named here. This article is commentary on public disclosures and is not investment advice.
Turn every deal into a defensible case for change.
Value engineering for every seller. A person approves what enters the library.

