Value Selling
What Anthropic's S-1 Case Studies Are Actually Worth

Anthropic filed its S-1 on October 26, 2026. Twenty pages of it, 208 to 227, are customer case studies: hours saved, cycle times cut, code shipped, orders grown. In our first read of the filing we went through what those panels say and how to read them. This piece asks the next question, the one a finance team asks of any case study a seller puts in front of them: what is it worth?
We priced 14 of the cases. The answer isn't one number. It's two numbers that shouldn't be added together, and a range that's wide for a reason worth understanding before you write your own.
What these numbers are, and aren't
Every dollar figure below is our projection, not a figure from the filing and not a figure from Anthropic or any customer. The operating metrics are quoted from the S-1. The financial inputs come from the companies' own annual reports or from public wage data. Everything else is an assumption, labeled as one, with a conservative, expected and upside value. This is an exercise in reading business cases. It isn't investment advice and says nothing about what Anthropic is worth.
The situation: good metrics, missing denominators
The panels are well-evidenced operating metrics. Salesforce reports that Slackbot, built on Claude, "is driving 8.1 million hours of annualized productivity gains." SAP reports "25 hrs saved per consultant per month." AIG's Lexington Middle Market Property business reports "+40% submissions bound." Walmart reports a "+35% increase in average order value via Sparky vs. non-Sparky orders."
Each of those is a numerator. To turn any of them into money you need things the panels weren't written to supply: how many people the hours apply to, what those people cost, what the business did with the time, what share of the change belongs to the product rather than the market, and what margin the extra revenue carries.
That isn't a criticism of the filing. A case-study panel exists to show that something changed. A business case has to show what the change is worth to one company, and the gap between the two is where most AI value claims fall apart in front of a finance team.
The diagnosis: capacity isn't cash
The first thing to separate is capacity from cash.
Capacity value is time priced at what it costs: hours recovered multiplied by a loaded hourly rate. It tells you how much work was freed up. It isn't money anyone banks.
Realized value is what reaches the income statement: costs actually avoided, revenue actually added, margin on that revenue. Time becomes realized value only through a named action. Someone stops hiring, cuts contractor spend, bills the hours, or uses them to win business that wouldn't otherwise close.
Most of the S-1's metrics are capacity metrics or speed metrics. A few, like AIG's bound submissions and Walmart's order value, sit on a revenue line. Treating them the same way is how a case-study page becomes a number nobody in finance will sign.
A worked example: AIG's +40% submissions bound
AIG's panel (S-1, p. 209) reports that in Lexington's Middle Market Property business, Underwriting by AIG Assist, which AIG built on Claude, delivered "+30% Quoted submissions," "-55% Time-to-quote" and "+40% Submissions bound," as of March 31, 2026. It adds that, where deployed, the tooling "cut submission review time by more than 90%."
The tempting reading is "40% more business." Here's what it takes to put a dollar figure on it.
Illustrative projection, built from labeled assumptions
From the S-1: +40% submissions bound. From AIG's FY2025 10-K: North America Commercial ran a combined ratio of 86.8, so each premium dollar left about 13.2 cents of underwriting margin. Assumed: the business's premium before Claude (not disclosed): $250M conservative, $400M expected, $700M upside. Share of the uplift that belongs to Claude rather than market growth and pricing: 25%, 50%, 75%. Underwriting margin on the new business: 5%, 13.2%, 18%. Investment income on the extra premium: 0%, 3%, 5%. We assume premium per bound policy didn't change.
Run the arithmetic. Forty percent on a $400 million base is $160 million of extra premium, and that's the number a quick reading produces. Attribute half of it to Claude and you have $80 million. At a 13.2% margin that's $10.6 million of underwriting profit, plus $2.4 million of investment income: about $13 million a year, expected. The conservative case is $1.3 million and the upside is $48 million.
The underwriter time saved doesn't appear in the total. It's the mechanism that let the team review every submission and bind more of them, so counting it again would double-count. Valued on its own, it's under $1 million a year of capacity.
Two lessons carry over to every case. First, the expected value is about 8% of the naive reading. Second, the biggest swing input isn't in the filing. The premium base and the attribution share move the answer more than the +40% does.
The totals: two numbers you can't add
We ran the same discipline across 14 cases: S-1 metric, then the value chain from metric to money, then labeled assumptions at three levels. Thirteen are recurring annual values. Mozilla's is a one-time value from vulnerabilities found and fixed, so it's kept out of the annual totals.
$1.25B a year
Capacity value across the 13 annual cases: time recovered, priced at loaded cost. This is not a cash figure, and it isn't added to the realized range below.
Serre projection from S-1 case-study metrics (pp. 209-227), company annual reports and BLS wage data. Expected-case inputs. Walmart has no time component.
$132M to $2.4B a year
Projected realized value across the same 13 cases: conservative $132M, expected $622M, upside $2,415M. The upside is about 18 times the conservative case.
Serre projection. Every figure that is not in the S-1 comes from a company report or public wage data, or it is an assumption.
Reading the two side by side tells you more than either number alone.
Where the only route is time, most of the capacity doesn't become cash. For Salesforce, SAP, PayPal, Travelers, the International Rescue Committee, KBF Design Gallery and UST, the expected case keeps about $221 million of roughly $942 million of capacity, under a quarter. Salesforce and PayPal keep about 15% each in the expected case. We assume most of a few saved hours a week is absorbed into the day rather than banked, and we discount Salesforce's own estimate of its hours by a quarter. SAP keeps 40%, because consultants can bill some of their saved time.
Where the route is revenue or speed, the hours barely matter. AIG, Bristol Myers Squibb, DoorDash and Walmart account for about $276 million of the expected $622 million, from more premium, earlier decisions, earlier product launches and larger orders. Their combined capacity value is under $8 million. An hours-only lens would have missed almost all of it.
The range is wide because the denominators are missing. The conservative-to-upside spread isn't noise. It's the honest width of what can't be known from outside the company: how many consultants use a tool, what share of e-commerce orders run through an assistant, what share of a year's growth came from new products.
The 14 cases at a glance
Capacity is shown at expected-case inputs. Realized values are conservative / expected / upside, in millions of dollars a year unless noted.
- DoorDash (p. 211). "Launch timelines for major products cut by more than half." Capacity $4.0M. Realized $39M / $132M / $280M, mostly from new-product revenue arriving sooner.
- SAP (p. 221). "25 hrs saved per consultant per month." Capacity $306M. Realized $25M / $123M / $589M, depending on how many consultants use it and how much saved time becomes billable.
- Mercado Libre (p. 216). Merged code up more than 100% year over year, and Cowork "shortened the monthly financial close by a third." Capacity $234M. Realized $24M / $107M / $379M.
- Salesforce (p. 220). "8.1 million hours of annualized productivity gains," which the panel labels as estimated. Capacity $605M. Realized $29M / $91M / $193M.
- Walmart (p. 226). "+35% increase in average order value via Sparky vs. non-Sparky orders." No time component. Realized $10M / $86M / $419M, because shoppers who choose an assistant may already buy more, so only part of the gap is caused by it.
- Bristol Myers Squibb (p. 210). Time to answer a real-world-evidence question cut "from up to two months to five minutes or less." Capacity $2.9M. Realized $1.5M / $46M / $364M, almost all of it the value of earlier decisions on the few questions that sit on a program's critical path.
- Rakuten (p. 219). New-feature time to market from 24 to 5 working days. Capacity $63M. Realized $2.0M / $18M / $72M.
- AIG (p. 209). "+40% submissions bound." Capacity $0.7M (not added). Realized $1.3M / $13M / $48M.
- Travelers (p. 223). "-50% reduction in time to merge code." Capacity $25M. Realized $0.2M / $6.4M / $64M.
- PayPal (p. 218). "45 min average time saved per sales team member per week." Capacity $4.1M. Realized $0 / $0.6M / $2.1M.
- International Rescue Committee (p. 212). "464 staff hours returned per week." Capacity $0.56M. Mission value $0.24M / $0.42M / $0.63M, redeployed to programs rather than banked.
- UST (p. 224). Chip validation cycle time cut 50-70%. Capacity $0.54M. Realized $0.02M / $0.27M / $4.8M.
- KBF Design Gallery (p. 227). "490 hrs per year recovered for a single employee," which is 7 hours a project across 70 projects. Capacity about $25,000. Realized $13,000 / $19,000 / $25,000.
- Mozilla (p. 217), one-time. Twenty-two vulnerabilities found by Claude Opus 4.6, 14 of them high-severity, and 271 more found by Claude Mythos Preview. Capacity $2.0M. Realized $0.9M / $5.9M / $33M once, from avoided incident cost and bounty-equivalent value.
We didn't price six panels. Accenture's "7x return on its AI investment" and Legora's "$6.9 million in additional annual billing per 100 lawyers" are already in money terms, with no base disclosed to rebuild them from. Eli Lilly and Verizon report adoption. Shopify and Jane Street report metrics with no disclosed link to revenue or cost. Shy Bird's cut of more than 60% in brunch ticket times has no revenue base to value it against.
A case study shows that something changed. A business case has to show what the change is worth to one company, and who'll still call that number fair a year later.
A checklist for reading any AI case study, including yours
The same questions apply to the case studies your own team publishes, and to the business case you put in front of a buyer.
- What's the denominator? "25 hours per consultant" is worth nothing until you know how many consultants. Ask for the count, or state the one you're assuming.
- Is it time, speed or revenue? Each converts to money differently. Don't value a speed metric as if it were hours, or hours as if they were cash.
- What's the capture action? Recovered hours turn into cash only when someone names what they replace: a hire, a contractor, overtime, billable work.
- How much belongs to the product? If the market grew too, or the users self-selected, say what share you're attributing and why.
- Is it recurring or one-time? A migration finished early is a one-time saving. A faster weekly process recurs.
- Is it stated or estimated, and as a ceiling or an average? "Up to" is a ceiling. "Estimated" means a method you can't see.
- What would make the number wrong? Name the one input that moves it most, and show the answer at conservative and upside values of that input.
Where this helps, and where it doesn't
This method is good at one thing: telling you how big a claim could plausibly be once you've stated what it depends on. It turns "8.1 million hours" into a range finance can argue with, which is more useful than a headline they'll quietly discount.
It doesn't tell you what any of these companies realized. Only they know their premium base, their user counts and what they did with the time. Our assumptions could be wrong in either direction, and that's why every one is labeled and why the range is as wide as it is. Read the totals as a test of the method, not a verdict on the customers or on Anthropic.
The next step: do this for one buyer, not 14
Pricing 14 public case studies took a spreadsheet, a stack of annual reports and a lot of labeled guesses. Doing it for one prospect is easier, because the buyer can supply the denominators the S-1 couldn't. It's also more repetitive, because it has to be done again for every deal.
That's the repeated work Serre is built to take off a seller's desk. Guided Discovery gives reps a starting hypothesis: built-in research drafts the account's likely priorities, the context for each buyer role and the solution value, and the rep tests it with the buyer, filling in real counts where we had to assume them here. The arithmetic is deterministic, and every number shows where it came from. Capacity is kept apart from cash. Outcomes are presented as conservative, expected and upside projections, with the assumptions open to finance review. One case, with one set of totals, serves the whole buying group. No value engineer required: the guidance is in the product.
For what the filing says case by case, read our first look at the S-1. For why finance teams discount ROI models, see why your CFO doesn't trust your ROI model. Get started or contact sales.
Key takeaways
- 1The S-1 case studies report operating metrics. Turning them into money needs denominators the panels don't disclose.
- 2Across 13 annual cases, recovered time is worth about $1.25B a year at cost. That's capacity, not cash.
- 3Projected realized value is $132M conservative, $622M expected and $2,415M upside a year. The width comes from undisclosed inputs, not from noise.
- 4Where the only route is time, under a quarter of capacity becomes cash in the expected case. Where the route is revenue or speed, hours barely matter.
- 5Ask seven questions of any case study: denominator, metric type, capture action, attribution, recurrence, stated or estimated, and what would make it wrong.
Turn every deal into a defensible case for change.
Value engineering for every seller. A person approves what enters the library.

