RuncheyResearch

AI Absorption Ledger / AXP

American Express

AXP · Q2 2026 · reported 2026-07-24 · revenue $19.64bn

Assessment

Q2 2026 is the quarter American Express described where it has put AI to work. The CEO says coding cycle time is down to and calls it not a savings, because the capacity goes to a backlog; that servicing and travel representatives have AI-powered tools and their number has not grown with the business; that a new servicing portal with AI embedded is meant to cut handle time; that AI streamlines marketing campaigns; and that AI has long been used in credit, risk and fraud, with an LLM layer now being added. The 10-Q’s list of forward-looking factors adds AI usage to the technology costs it names, and a ChatGPT statement credit for business card members went live. No source gives an AI dollar, count or rate beyond the coding range.

New channels open for the AI usage bill, marketing production and credit, risk and fraud decisions. Every size shown is the ledger’s own estimate: internal build investment at and the usage bill at , both shares of revenue from the ledger’s peer readings, the ChatGPT credit at , marketing production at and programmer time at , against total expenses of . Card and travel servicing stay unsized because Q1 credited the same movement to fewer calls as well as AI; credit, risk and fraud and agentic commerce stay unsized because their money has not started.

The lines move for reasons the filing gives without AI. Total expenses rose on card member rewards, benefits and partner payments; salaries and employee benefits rose ; data processing and equipment rose on higher technology costs; marketing rose on customer acquisition. The CEO frames AI as capacity and future efficiency, and doubts that many companies have committed to profit-and-loss savings from it yet.

Sized channels against the income statement, Q2 2026

5 of 9 channels sized

Each blue mark is one channel's dollars for the quarter; a bar is the range of an estimate. Grey marks are the company's reported lines. The distance between them is the point: how large the AI channel is next to the line it sits in.

New money and old money, Q2 2026

2 new6 expanded1 relabelled

Each channel is tagged once for whether its money existed before language models, from the company's annual report and call at the start of the period. A bar splits one flow's sized dollars by that tag. The incremental total is the part that would not be there without the models: a new channel counts in full, an expanded one only for what AI added, a relabelled one at zero.

Paid for AI$15mn to $333mn sized

new $982k to $15mnexpanded $14mn to $319mn

Incremental total $982k to $333mnpoint $6.9mn$56mn in 2 channels has no traced baseline
Cost displaced by AI$0 to $47mn sized

expanded $0 to $47mnrelabelled not sized

Incremental total $0 to $47mnpoint $6.9mn

The sized total counts every channel the company credits to AI, including relabelled money that existed before language models and the ledger's own estimates for it. The incremental total counts relabelled channels at zero. A flow is split by layer where its dollars sit at more than one: end use, compute sold to builders, and hardware. The same dollar can be a buyer's spend, a cloud's revenue and a chipmaker's revenue, so the layers are never added together.

Paid for AI3 channels · $15mn to $333mn sized · $982k to $333mn incremental

engineering

Internal investment in AI capability

0.07% to 1.5% of the quarter’s revenue

Incremental total: counts at zero at the point and in full at the high end, with no traced baseline.

our inferencedisclosure: described· motive: exploratory· before LLMs: expanded

Management names agentic commerce as an investment added during the year and says it is investing in AI for credit, risk and fraud; the CFO ties AI to the company’s commitment to operating leverage. No amount is given. Salaries and employee benefits rose and data processing and equipment . The size shown is the ledger’s own estimate, , a share of revenue taken from the ledger’s readings of build investment at other buyers, as at JPMorgan, with Q1’s shares. Its change from Q1 is the revenue base moving, not a reading of more or less AI work.

Evidence: 7 quotes, 4 figures, 2 confounds, 4 from before coverage

Money the company spends building AI into its own products and work: the AI-powered products and capabilities the CEO says are under development, the agentic commerce initiatives he says required investment that was not in the plan at the start of the year, and the AI initiatives that the 10-Q’s list of forward-looking factors names among areas of spending. It sits in salaries and employee benefits (technology colleagues) and data processing and equipment, and neither is split. Separate from the bill for AI usage.

Why this motive

The CEO says agentic commerce required investment not in the plan (claim c11) and the company is investing in AI for credit and fraud (claim c7), while calling the work early innings with high hopes (claim c9): investing ahead of results, the exploratory tell, as in Q1.

Before LLMs: expanded

At the anchor the annual report already said the company must keep investing in technology across its business, machine learning and artificial intelligence among it, and that it and its partners used AI and machine learning; salaries and employee benefits were and data processing and equipment for FY2024, with no AI share of either. No quarter of the activity before AI can be traced. The size is the whole of an activity that existed before.

“In order to compete in our industry, we need to continue to invest in technology across all areas of our business, including in transaction processing, data management and analytics, machine learning and artificial intelligence, customer interactions and communications”
Filing, risk factors, 10-K periodic report, 2025-02-07
“effectively utilizing artificial intelligence and machine learning and increasing automation, including to address servicing and other business and customer needs”
Filing, mdna, 10-K periodic report, 2025-02-07
“Our and our partners’ use of artificial intelligence and machine learning is subject to various risks including flaws in models or datasets that may result in biased or inaccurate results”
Filing, risk factors, 10-K periodic report, 2025-02-07
“how we're driving efficiency, growth, and service through technology”
CEO, prepared remarks, earnings call, 2024-04-19

Figures

  • Salaries and employee benefits · 2026-CQ2
  • Salaries and employee benefits, growth year over year · 2026-CQ2
  • Data processing and equipment · 2026-CQ2
  • Data processing and equipment, growth year over year · 2026-CQ2

What else could explain it

  • line composition: Salaries and employee benefits rose on compensation and incentive costs, and data processing and equipment carries all technology costs; neither is split by purpose.
  • other: The CEO says the extra technology investment pulls a backlog of platform refreshes into the second half (claim c22), without naming AI.

Quotes

“Look, we have used AI in credit and risk and fraud for 15, 16 years. Now what we are doing is how do you implement sort of unstructured data agentic in that to help make the decisions a little bit better? Look, we are investing in it.”
c7 · CEO, qa, earnings call, 2026-07-24
“We are committed to doing that again. AI is going to play a critical role in that. It's part of the reason as well that we are investing so much in strategy.”
c10 · CFO, qa, earnings call, 2026-07-24
“You've seen, we've announced a number of things that we're participating in from an agentic commerce perspective, that requires investment as well, those things weren't on the docket at the beginning of the year.”
c11 · CEO, qa, earnings call, 2026-07-24
“its decisions regarding spending in such areas as technology, business and product development, sales force, premium servicing and AI initiatives”
c16 · Filing, mdna, 10-Q periodic report, 2026-07-24
“embedding AI into our business and increasing automation, including to streamline and improve internal processes and decision making, enhance our products, develop new capabilities and address servicing and other business and customer needs”
c19 · Filing, mdna, 10-Q periodic report, 2026-07-24
“Across a wide range of technology platforms, we're able to pull some of those investments into the second half of the year.”
c22 · CEO, qa, earnings call, 2026-07-24
“management’s ability to balance expense control and investments in the business, develop new capabilities, features and value propositions, effectively utilize artificial intelligence”
c24 · Filing, press release, 8-K earnings release, 2026-07-24

By quarter

  • Q1 2026described · our inference · exploratory · $13mn to $282mn
  • Q2 2026described · our inference · exploratory · $14mn to $293mn

vendor bill

Bill for AI usage

0.01% to 0.07% of the quarter’s revenue

Incremental total: counts in full.

our inferencedisclosure: described· motive: exploratory· before LLMs: new

A new channel. The Q2 10-Q adds AI usage to the increased technology costs named in its list of forward-looking factors that could affect operating expenses; the Q1 list did not. The list does not say the cost rose in the quarter or how large it is. Data processing and equipment, the line that would carry the bill, rose to , or more than a year earlier, on technology costs the 10-Q explains without AI. The line is a ceiling, not the bill. The size shown is the ledger’s own estimate, , a share of revenue () taken from the ledger’s readings of other buyers’ AI vendor bills, as at JPMorgan.

Evidence: 3 quotes, 4 figures, 2 confounds

What the company pays outside vendors for AI usage, which the Q2 2026 10-Q names among increased technology costs in its list of forward-looking factors that could affect operating expenses; the list does not say that the cost rose in the quarter. No vendor, model, seat count or amount is given; the bill would sit in data processing and equipment, inside other expenses. Coding tools, the AI-powered tools given to servicing and travel representatives and AI used in marketing would all draw on it, and no part is named, so it is one channel.

Why this motive

A cost the 10-Q’s list of forward-looking factors names among increased technology costs (claim c15) with no displaced line and no price set against it; the tools it pays for are described as early and aimed at capacity (claims c2 and c9): the exploratory tell.

Before LLMs: new

The anchor names no AI usage, model or token bill; the line that would carry it, data processing and equipment, was for FY2024. A bill for AI usage could not exist without LLMs, so the channel is new whatever the anchor says.

Figures

  • Data processing and equipment · 2026-CQ2
  • Data processing and equipment, prior-year quarter · 2025-CQ2
  • Data processing and equipment, growth year over year · 2026-CQ2
  • Data processing and equipment against the prior-year quarter · 2026-CQ2

What else could explain it

  • line composition: Data processing and equipment carries all processing, software and equipment costs; the 10-Q explains its rise as higher technology costs without naming AI (claim c20).
  • bundling: AI usage may be bought inside cloud and software agreements and never billed as AI.

Quotes

“increased technology costs, including AI usage and investments in technology innovations and system upgrades”
c15 · Filing, mdna, 10-Q periodic report, 2026-07-24
“The increase for the three month period was primarily driven by an increase in legal reserves and higher technology costs, partially offset by gains on Amex Ventures investments”
c20 · Filing, mdna, 10-Q periodic report, 2026-07-24
“Across a wide range of technology platforms, we're able to pull some of those investments into the second half of the year.”
c22 · CEO, qa, earnings call, 2026-07-24

pricing packaging

ChatGPT statement credit for business card members

0% to 0.13% of the quarter’s revenue

Incremental total: counts at zero at the point and in full at the high end, with no traced baseline.

our inferencedisclosure: described· motive: product-defensive· before LLMs: expanded

The CEO says a annual ChatGPT business statement credit was introduced in the quarter for U.S. Business Platinum and Gold card members. The credit is paid to those business card members against their charge for the subscription; the card members pay OpenAI, a private lab not on this ledger, and American Express earns discount revenue on the charge. In Q1 the CEO called the effect of the new commercial launches benign (claim c21), and the high end is narrowed for it. No enrolment or cost is given; Card Member services expense was against a year earlier. The size shown is the ledger’s own estimate, , built on Commercial Services cards-in-force () and the annual-plan band.

Evidence: 2 quotes, 4 figures, 2 confounds, 1 from before coverage

A statement credit for ChatGPT Business subscriptions, offered as a benefit on U.S. Business Platinum and Gold cards. The credit is paid to the company’s own business card members, against the charge on their card for the subscription; the card members pay OpenAI, a private AI lab that is not on this ledger, and American Express earns discount revenue on that charge like any other. The subscription itself is the card members’ spend, not the company’s. Named on the Q1 2026 call as part of the commercial roadmap and introduced in Q2 2026. The cost would sit in Commercial Services card member benefit costs; no enrolment, amount or partner funding is given. Counterparty: small and mid-sized business card members, the only members the sources name as eligible.

Why this motive

An AI subscription paid for inside the card’s benefits at no separate charge (claim c14), as in Q1: the product-defensive tell of a feature folded into an existing price.

Before LLMs: expanded

At the anchor the company already gave card members statement credits for purchases with partners as card benefits, with Card Member services expense of for FY2024. A credit for an AI subscription is a benefit budget redirected to an AI partner, read as expanded and sized as a level; no earlier quarter of that slot can be traced. The size is the whole of an activity that existed before.

“providing greater value to our Card Members (e.g., Amex Offers and statement credits for purchases with partners)”
Filing, business, 10-K periodic report, 2025-02-07

Figures

  • ChatGPT business statement credit per eligible card member, per year · FY2026
  • Commercial Services proprietary cards-in-force · as-of 2026-06-30
  • Card Member services expense · 2026-CQ2
  • Card Member services expense, prior-year quarter · 2025-CQ2

What else could explain it

  • bundling: The credit is one of the Business Platinum and Gold benefits; Commercial Services card member services rose on the new U.S. Business Platinum benefits as a whole (claim c21).
  • other: Partner statement credits are often funded in part by the partner; the sources do not say whether OpenAI pays any part of this one, which would shrink the company’s own cost.

Quotes

“For example, in the second quarter, we introduced a $300 ChatGPT business annual statement credit for our U.S. Business Platinum and Gold Card members”
c14 · CEO, prepared remarks, earnings call, 2026-07-24
“Card Member services expense increased for both the three and six month periods, primarily driven by the new U.S. Business Platinum benefits.”
c21 · Filing, mdna, 10-Q periodic report, 2026-07-24

By quarter

  • Q1 2026described · described, no size · product-defensive
  • Q2 2026described · our inference · product-defensive · $578k to $26mn

Cost displaced by AI5 channels · $0 to $47mn sized · $0 to $47mn incremental · 3 not sized

engineering · cheap to verify

Programmer productivity from AI in coding and testing

0% to 0.11% of the quarter’s revenue

Incremental total: counts in full.

our inferencedisclosure: direction only· motive: exploratory· before LLMs: expanded

The CEO puts the decrease in coding cycle time at to , a wider claim than Q1’s benefit in coding and testing, and says it is not a saving. The input is a cut in elapsed cycle time, not labour freed, so the estimate converts it to effort with an assumption of its own. The size shown is the ledger’s own estimate, ; its range starts at none, because the CEO says the gain is not a saving.

Evidence: 2 quotes, 4 figures, 2 confounds, 3 from before coverage

Technology colleagues’ time saved by AI in coding and testing, which the CEO measures as a benefit to programmers in Q1 2026 and a decrease in coding cycle time in Q2 2026. The CEO says it is not a saving, because the freed capacity goes to a backlog of technology projects, so the gain may show as more output rather than lower cost. The line it would show in is salaries and employee benefits.

Why this motive

The CEO measures a cycle-time decrease and says in the same breath that it is really not a savings because it lets the company do more (claim c2); the capacity goes to a backlog (claim c3). A measured result the company itself directs to output, with no displaced line, keeps the exploratory tell from Q1.

Before LLMs: expanded

At the anchor the company’s own technologists built and ran its platforms, paid inside salaries and employee benefits ( for FY2024), and the annual report already counted adoption of AI among the ways it might realize operational efficiencies; no coding tool or productivity rate was given. LLM tools in coding and testing change the cost per unit of that work, not its existence. The size is the change AI made, not the whole line.

“In order to compete in our industry, we need to continue to invest in technology across all areas of our business, including in transaction processing, data management and analytics, machine learning and artificial intelligence, customer interactions and communications”
Filing, risk factors, 10-K periodic report, 2025-02-07
“our ability to realize operational efficiencies, including through increased scale and automation and continued adoption of artificial intelligence technologies”
Filing, mdna, 10-K periodic report, 2025-02-07
“how we're driving efficiency, growth, and service through technology”
CEO, prepared remarks, earnings call, 2024-04-19

Figures

  • Decrease in coding cycle time from AI, per the CEO, low end · as-of 2026-07-24
  • Decrease in coding cycle time from AI, per the CEO, high end · as-of 2026-07-24
  • Salaries and employee benefits · 2026-CQ2
  • Salaries and employee benefits, growth year over year · 2026-CQ2

What else could explain it

  • other: The CEO says the gain is not a saving and goes to more projects (claims c2 and c3); more output is not lower cost.
  • line composition: Salaries and employee benefits holds all colleagues’ pay, which rose on compensation and incentive costs.

Quotes

“When we look at it, one of the first places that we have deployed it, and we have deployed it in multiple places in the business right now is technology. From a technology perspective, what we are seeing is anywhere from a 30%-40% decrease in cycle time from a coding perspective. That is really not a savings because what that does is allows us to do more.”
c2 · CEO, qa, earnings call, 2026-07-24
“If you remember early on, I mentioned that we have a large backlog of technology projects, we are getting to more things quicker.”
c3 · CEO, qa, earnings call, 2026-07-24

By quarter

  • Q1 2026direction only · our inference · exploratory · $0 to $25mn
  • Q2 2026direction only · our inference · exploratory · $0 to $21mn

customer support · cheap to verify

Card member servicing assisted by AI tools

Not sized

The CEO describes a flat representative count, not a number, and the same movement was credited in Q1 to fewer calls as well as AI tools; nothing separates AI’s part, so the channel is left unsized with no ballpark. The servicing portal had only just launched, so its money had not started in the quarter either. The ceiling is salaries and employee benefits ().

described, no sizedisclosure: direction only· motive: exploratory· before LLMs: expanded

The CEO says customer service representatives have AI-powered tools, that their number has not grown as the business grows and may fall through attrition, and that a servicing portal with AI embedded was just launched for card member service representatives to reduce handle time. No count, handle time or cost is given. The ledger reads this quarter’s account with Q1’s, which named fewer calls as a second cause, and leaves the channel unsized.

Evidence: 4 quotes, 2 figures, 2 confounds, 4 from before coverage

Payroll of card member service representatives avoided because they work with AI-powered tools, including the servicing portal with AI embedded that the CEO says was launched in Q2 2026 and is meant to reduce handle time. The CEO describes the number of representatives as flat while the business grows; in Q1 2026 he credited the falling ratio of representatives to volume to fewer customers wanting to call as well as to AI tools. The line it would show in is salaries and employee benefits. Separate from travel servicing.

Why this motive

The CEO credits a flat representative count to AI-powered tools (claims c4 and c5) and says the new portal will reduce handle time (claim c8), a result not yet measured; Q1 credited the same ratio to fewer calls as well (claim c8). The tells conflict across the quarters, so the less durable motive stands: exploratory, carried from Q1.

Before LLMs: expanded

At the anchor customer care professionals served card members by phone and digitally, paid inside salaries and employee benefits ( for FY2024, with about colleagues at year end), and the annual report already named AI, automation and card member self-service among the ways to address servicing. AI tools for representatives change the cost per contact, not the existence of the work. The size is the change AI made, not the whole line.

“our ability to realize operational efficiencies, including through increased scale and automation and continued adoption of artificial intelligence technologies”
Filing, mdna, 10-K periodic report, 2025-02-07
“effectively utilizing artificial intelligence and machine learning and increasing automation, including to address servicing and other business and customer needs”
Filing, mdna, 10-K periodic report, 2025-02-07
“Our customer care professionals, travel consultants and partners treat servicing interactions as an opportunity to bring the brand to life for our customers, add meaningful value and deepen relationships.”
Filing, business, 10-K periodic report, 2025-02-07
“our ability to innovate efficient channels of customer interactions and the willingness of Card Members to self-service and address issues through digital channels”
Filing, mdna, 10-K periodic report, 2025-02-07

Figures

  • Salaries and employee benefits · 2026-CQ2
  • Salaries and employee benefits, growth year over year · 2026-CQ2

What else could explain it

  • other: In Q1 the CEO named fewer customers wanting to call beside AI tools as the cause of fewer representatives per unit of volume (claim c8); nothing in Q2 separates the two.
  • operating leverage: A flat servicing workforce as volume grows is also ordinary scale.

Quotes

“From a servicing perspective, both from a travel and from a card servicing perspective, we have equipped our customer service and travel agents with AI-powered tools. What we have seen there is a lack of acceleration in hiring of more travel representatives, more customer service reps, even as the business continues to grow.”
c4 · CEO, qa, earnings call, 2026-07-24
“If you look at how many representatives that we had servicing our customers, that number really has not grown, and we would expect over time for that number to decrease. I think that will normally decrease from an attrition perspective.”
c5 · CEO, qa, earnings call, 2026-07-24
“We just launched a servicing portal for our CCPs, our card member service representatives, where we have embedded within, and it will reduce the handle time for them and enable them to give our customers a much better experience.”
c8 · CEO, qa, earnings call, 2026-07-24
“embedding AI into our business and increasing automation, including to streamline and improve internal processes and decision making, enhance our products, develop new capabilities and address servicing and other business and customer needs”
c19 · Filing, mdna, 10-Q periodic report, 2026-07-24

By quarter

  • Q1 2026direction only · described, no size · exploratory
  • Q2 2026direction only · described, no size · exploratory

customer support · cheap to verify

Travel servicing assisted by AI tools

Not sized

A flat travel representative count, with no number, credited in Q1 to fewer calls as well as AI tools; nothing separates AI’s part, so the channel is unsized with no ballpark. The ceiling is salaries and employee benefits (), beside travel commissions and fees of .

described, no sizedisclosure: direction only· motive: exploratory· before LLMs: expanded

The CEO says travel agents have AI-powered tools and that hiring of travel representatives has not accelerated even as the business grows. Travel commissions and fees were in the quarter. No count or cost is given.

Evidence: 3 quotes, 2 figures, 2 confounds, 4 from before coverage

Payroll of travel representatives (the consultants who book trips for card members) avoided because they work with AI-powered tools. The CEO names travel beside card servicing in both quarters and says hiring of travel representatives has not accelerated as the business grows. The line it would show in is salaries and employee benefits; the travel business earns travel commissions and fees. Separate from card member servicing.

Why this motive

Same tells as card servicing: a flat travel representative count credited to AI tools (claim c4), with Q1’s second cause unseparated; exploratory, carried from Q1.

Before LLMs: expanded

At the anchor travel consultants already served card members, the business earned travel commissions and fees of for FY2024, and the consultants’ pay sat inside salaries and employee benefits (). AI tools for travel representatives change the cost per booking, not the existence of the work. The size is the change AI made, not the whole line.

“our ability to realize operational efficiencies, including through increased scale and automation and continued adoption of artificial intelligence technologies”
Filing, mdna, 10-K periodic report, 2025-02-07
“effectively utilizing artificial intelligence and machine learning and increasing automation, including to address servicing and other business and customer needs”
Filing, mdna, 10-K periodic report, 2025-02-07
“Our customer care professionals, travel consultants and partners treat servicing interactions as an opportunity to bring the brand to life for our customers, add meaningful value and deepen relationships.”
Filing, business, 10-K periodic report, 2025-02-07
“our ability to innovate efficient channels of customer interactions and the willingness of Card Members to self-service and address issues through digital channels”
Filing, mdna, 10-K periodic report, 2025-02-07

Figures

  • Salaries and employee benefits · 2026-CQ2
  • Travel commissions and fees · 2026-CQ2

What else could explain it

  • other: Q1 named fewer calls beside AI tools (claim c8); nothing separates them.
  • mix shift: Travel bookings grew strongly in the quarter, so a flat travel workforce may also reflect booking channel and product mix.

Quotes

“From a servicing perspective, both from a travel and from a card servicing perspective, we have equipped our customer service and travel agents with AI-powered tools. What we have seen there is a lack of acceleration in hiring of more travel representatives, more customer service reps, even as the business continues to grow.”
c4 · CEO, qa, earnings call, 2026-07-24
“If you look at how many representatives that we had servicing our customers, that number really has not grown, and we would expect over time for that number to decrease. I think that will normally decrease from an attrition perspective.”
c5 · CEO, qa, earnings call, 2026-07-24
“embedding AI into our business and increasing automation, including to streamline and improve internal processes and decision making, enhance our products, develop new capabilities and address servicing and other business and customer needs”
c19 · Filing, mdna, 10-Q periodic report, 2026-07-24

By quarter

  • Q1 2026direction only · described, no size · exploratory
  • Q2 2026direction only · described, no size · exploratory

marketing · cheap to verify

Marketing campaign production streamlined by AI

0% to 0.13% of the quarter’s revenue

Incremental total: counts in full.

our inferencedisclosure: described· motive: exploratory· before LLMs: expanded

A new channel. The CEO says AI now streamlines marketing campaigns so they go out faster. Marketing rose to on customer acquisition, and management plans to raise it further in the second half. The stated effect is speed. The size shown is the ledger’s own estimate, , a share of the production part of the line, with a saving share that runs from none.

Evidence: 2 quotes, 3 figures, 2 confounds, 1 from before coverage

The cost and time of producing marketing campaigns, which the CEO says AI now streamlines so that campaigns go out faster. Who did the work before (staff or agencies) is not said. The line it sits in is marketing, which is mostly customer acquisition and brand spending.

Why this motive

The CEO describes faster campaigns and time to money (claim c6) with no measure and no line shown to move: AI named as the cause with no measure, the exploratory tell.

Before LLMs: expanded

At the anchor marketing was for FY2024, mostly the cost of promotional activities to attract, engage and retain customers; producing campaigns was existing work inside it, with no AI named. AI that speeds campaign production changes its cost and timing, not its existence. The size is the change AI made, not the whole line.

“our ability to realize operational efficiencies, including through increased scale and automation and continued adoption of artificial intelligence technologies”
Filing, mdna, 10-K periodic report, 2025-02-07

Figures

  • Marketing · 2026-CQ2
  • Marketing, prior-year quarter · 2025-CQ2
  • Marketing, growth year over year · 2026-CQ2

What else could explain it

  • line composition: Marketing is mostly customer acquisition offers and media, which rose on acquisition and growth initiatives (claim c23).
  • other: The CEO describes speed, which may raise marketing returns rather than lower cost.

Quotes

“From a marketing perspective, we are really using AI right now to streamline some of our marketing campaigns so that we are getting out the marketing campaigns a lot quicker, and that gets you time to money.”
c6 · CEO, qa, earnings call, 2026-07-24
“Marketing expense increased for both the three and six month periods, primarily driven by higher levels of spending on customer acquisition and other growth initiatives.”
c23 · Filing, mdna, 10-Q periodic report, 2026-07-24

back office · expensive to verify

Credit, risk and fraud decisions with AI

Not sized

The new layer has not been delivered: the CEO describes it as work under way, with no date, so its money has not started and the channel gets no ballpark. The models already in use predate LLMs and are read as relabelled. Card net write-offs () are the line it would show in.

described, no sizedisclosure: described· motive: exploratory· before LLMs: relabelled

A new channel, opened on the CEO’s answer that AI has been used in credit, risk and fraud for many years and that the company is now implementing unstructured data and agentic methods in it. Card net write-offs were against a year earlier, on balance growth, with a stable write-off rate. Unsized.

Evidence: 1 quote, 2 figures, 2 confounds, 3 from before coverage

Credit and fraud losses avoided by AI models in credit, risk and fraud decisions. The CEO says the company has used AI there for many years and is now working on adding unstructured data and agentic methods to improve the decisions. The line it would show in is card net write-offs (and fraud costs inside other expenses, which are not reported on their own).

Why this motive

The CEO says the company is working out how to add unstructured data and agentic methods to its credit, risk and fraud models to make decisions a little better (claim c7): work in progress with no measure, the exploratory tell.

Before LLMs: relabelled

At the anchor the company already used models and automation throughout its business to support decisions and manage risks, and listed machine learning and artificial intelligence among the technologies it invests in; the CEO says AI has been used in credit, risk and fraud for many years. The covered call names an LLM layer being implemented and no measured change, so the tie-break reads the channel as existing modelling under a new name.

“In order to compete in our industry, we need to continue to invest in technology across all areas of our business, including in transaction processing, data management and analytics, machine learning and artificial intelligence, customer interactions and communications”
Filing, risk factors, 10-K periodic report, 2025-02-07
“We use models and automation throughout our business, including to inform and support decision making, manage risks and estimate financial values.”
Filing, risk factors, 10-K periodic report, 2025-02-07
“Our and our partners’ use of artificial intelligence and machine learning is subject to various risks including flaws in models or datasets that may result in biased or inaccurate results”
Filing, risk factors, 10-K periodic report, 2025-02-07

Figures

  • Card balances net write-offs · 2026-CQ2
  • Card balances net write-offs, prior-year quarter · 2025-CQ2

What else could explain it

  • relabel: The CEO says AI has been used in credit, risk and fraud for many years (claim c7); the models predate LLMs.
  • other: Net write-offs move with balances and credit quality; the 10-Q credits the premium customer base and the reserve release to credit performance, not to models.

Quotes

“Look, we have used AI in credit and risk and fraud for 15, 16 years. Now what we are doing is how do you implement sort of unstructured data agentic in that to help make the decisions a little bit better? Look, we are investing in it.”
c7 · CEO, qa, earnings call, 2026-07-24

Revenue arriving through AI1 channel · 1 not sized

distribution

Card spending by AI agents (agentic commerce)

Not sized

The money has not started: the CEO places agentic commerce in the preseason, before its early innings, and gives no volume; the developer kit is free and the protection is a commitment, not a priced product. Discount revenue () is the line it would reach.

described, no sizedisclosure: described· motive: exploratory· before LLMs: new

The CEO says the closed loop and data from both sides of a transaction position the company in agentic commerce, which requires investment added during the year, and that the market is still in the preseason. The 10-Q’s list of forward-looking factors adds search and booking of Resy restaurants through AI platforms to the initiatives. No volume, partner or price is given; unsized.

Evidence: 5 quotes, 1 figure, 2 confounds

Discount revenue and card fees from purchases that AI agents make for card members, through the Amex Agentic Commerce Experiences developer kit (introduced April 2026), Amex Agent Purchase Protection for registered agent purchases, and the planned work with AI companies to make membership assets (travel, dining) discoverable and bookable on their platforms. The kit is free to developers and the protection is a commitment, not a priced product; management says it is too early to quantify. The protection may carry claims cost later, read here as part of the channel.

Why this motive

The CEO says the market is in the preseason (claim c13) and that internal AI is further along than agentic commerce (claim c9): pre-revenue, the exploratory tell in management’s words, carried from Q1.

Before LLMs: new

The anchor reports no purchases by AI agents and no product for them; the line such purchases would reach, discount revenue, was for FY2024. Card payments initiated by an AI agent could not exist without LLMs, so the channel is new.

Figures

  • Discount revenue · 2026-CQ2

What else could explain it

  • other: Any agentic purchases reach discount revenue beside all other card spending, which grew with billed business; no agentic share is given.
  • other: The CEO names fraud and hallucinations as risks of agentic commerce in the same answer, so the protection commitment may bring losses before volume brings revenue.

Quotes

“You've seen, we've announced a number of things that we're participating in from an agentic commerce perspective, that requires investment as well, those things weren't on the docket at the beginning of the year.”
c11 · CEO, qa, earnings call, 2026-07-24
“which is why we announced months ago now the agentic insurance product where I think we're going to have a huge advantage from a trust, service, and security perspective over our competitors in the marketplace because we have the data from both sides.”
c12 · CEO, qa, earnings call, 2026-07-24
“We're sort of in the preseason. We didn't even get to the early innings of the regular season yet.”
c13 · CEO, qa, earnings call, 2026-07-24
“advancing our agentic commerce initiatives, including embedding our payment capabilities in emerging AI ecosystems, such as through the Amex Agentic Commerce Experiences™ developer kit and Amex Agent Purchase Protection™, making Membership assets discoverable and actionable on AI platforms”
c17 · Filing, mdna, 10-Q periodic report, 2026-07-24
“enable the search and booking of Resy venues through AI platforms”
c18 · Filing, mdna, 10-Q periodic report, 2026-07-24

By quarter

  • Q1 2026described · described, no size · exploratory
  • Q2 2026described · described, no size · exploratory

Reported lines, year-over-year growth

Revenue +10.0%

Q2 2026. Growing slower than revenue: salaries and employee benefits (+8.9%), marketing (+6.1%). A displaced cost shows up as a line that stays under the dashed revenue line. These are the audited lines, as first reported; nothing here is attributed to AI by the filing.

Total expensesSalaries and employee benefitsMarketingData processing and equipmentRevenue
-5%0%5%10%15%20%Q1 2025Q2 2025Q3 2025Q4 2025Q1 2026Q2 2026Data processing and equipmentTotal expensesRevenueSalaries and employee benefitsMarketing
Reported values and filings
LineQ1 2025Q2 2025Q3 2025Q4 2025Q1 2026Q2 2026
Total revenues, net of interest expense$16.97bn$17.86bn$18.43bn$18.98bn$18.91bn$19.64bn
Total expenses$12.49bn$12.90bn$13.31bn$14.48bn$13.88bn$14.48bn
Salaries and employee benefits$2.12bn$2.15bn$2.24bn$2.50bn$2.48bn$2.34bn
Marketing$1.49bn$1.55bn$1.60bn$1.61bn$1.48bn$1.65bn
Data processing and equipment$705mn$720mnn/an/a$767mn$817mn