Newsletter · · Ashutosh Agarwal

IBM's Full Print Lands and a Third of Slipped Deals Already Return - IT Services vs AI - Week of July 25, 2026

IBM's CEO makes the case that Q2's collapse was a one-quarter capex reprioritization, while Salesforce's outcome-based Agentforce pricing and the big consultancies' own AI-tool adoption sharpen the bull-versus-bear debate for IT services. Coverage window: the week of July 25, 2026 (episodes dated July 17 to 23).

IT Services vs AI

Week of July 25, 2026: IBM's Full Print Lands and a Third of Slipped Deals Already Return

The week the numbers arrived. Last week IBM warned; this week its CEO sat down twice to explain what actually happened, and to argue, with a straight face, that a $70-billion-plus wipeout was a rounding error caused by a few big customers moving their spending around for one quarter. Whether you believe him is the whole ballgame.

Subject Line

IBM's full print lands: a third of slipped deals already back.

TL;DR

IBM's CEO reframed the collapse as temporary, and put numbers on it. Arvind Krishna told two shows this week that "a few tens of large capex deals" slipped out of Q2, that about a third have already come back in three weeks, and that he expects to recover roughly three-quarters of them by year-end. He also drew a sharp line: only about 2% of IBM's software is the kind AI can easily replace. Bull-case if true; a very convenient story if not.

The bear thesis got a concrete new data point, from Salesforce itself. Salesforce launched an outcome-based "pay-per-resolution" price for its Agentforce service agent ($2 per resolved issue, not per conversation). An IT-negotiation advisor says the switch happened because customers balked at consumption pricing and it was slowing Agentforce sales cycles, the first real corroboration of last week's "Agentforce is stalling" worry.

The big consultancies are now heavy users of the very tools that threaten them. The creator of Claude Code said on Odd Lots that Deloitte and Accenture are among Anthropic's large customers writing ever more of their code with AI, and that inside Anthropic, roughly 90% of code is now written by Claude Code. The productivity gains are real; the question is who captures them.

What's New

The five developments that actually move numbers this week, ranked by how actionable they are.

1. IBM's CEO did a two-stop media tour to un-spook the market, and it half-worked. Following last week's rare pre-announcement, Arvind Krishna sat for interviews on Bloomberg Talks, "IBM CEO Arvind Krishna Talks Infrastructure Deals" (Jul 23) and Squawk on the Street, "10AM Hour: Google Cloud CEO, IBM CEO, Tesla's Worst Day in More Than a Year" (Jul 23). This is operator commentary, the CEO himself, and it is the most important thing an IT-services investor heard all week. The gist: the miss was "a few tens of large capex deals" that a subset of Fortune-100 clients pushed out to prioritize spending on AI servers, storage and memory late in the quarter. Krishna's key claim: "one third of what didn't happen has already come back," and on Squawk he added that this recovery happened in three weeks, that IBM typically recovers about three-quarters of slipped deals over the following six months, and that a third coming back this fast is "a stronger signal than would typically be the case." He cut full-year revenue guidance only modestly, to 4-5% from "5%-plus," while maintaining the roughly $1 billion of free-cash-flow growth that analysts cared about. He also said IBM has not seen the June capex-diversion behavior continue into July, but warned memory prices (up "three to four times over the last 18 months") and networking gear (up "60 to 80 percent") could keep pressuring customer budgets. Why it matters: this is the single cleanest read on whether the "discretionary IT spend is deteriorating" narrative is a one-quarter air pocket or the start of something structural. The next two prints will settle it.

2. Krishna drew the line on which software AI eats, and put Starbucks on the record. On Squawk, Krishna quantified the vulnerable slice precisely: "about 2% of our software falls into the category which I'll say could be easily replaced with a new application built on AI." The tell is the interface: software whose value is mostly a human clicking through a pretty screen is exposed; software that frees up data, secures it, or moves it across hybrid environments is a tailwind. He used the widely-reported Starbucks story as his exhibit: Starbucks is a ~$2 million-a-year software client replacing IBM's TriRiga real-estate lease-management tool, which Krishna noted was "almost 10 years old" and hadn't been upgraded. His framing: "software which is largely interaction-based… can be easily replaced by AI and AI agents." Why it matters: this "interaction-heavy = at risk, infrastructure = safe" test is exactly the lens the market is now applying to every enterprise-software name (Salesforce, ServiceNow, Workday, SAP). Krishna is essentially arguing IBM sits on the safe side of that line. It's also the clearest CEO-level admission this quarter that AI is actively replacing some enterprise software, just not, in his telling, the parts that matter to IBM.

3. Salesforce moved to pay-per-outcome pricing on Agentforce, because the old model was slowing sales. On Insights for IT Negotiations, "Outcome-Based Pricing: What Salesforce's Pay-Per-Resolution Model Really Means" (Jul 23), Adam Mansfield, who leads the Salesforce practice at IT-sourcing advisory UpperEdge (an advisor voice, not a Salesforce operator), walked through the new "Agentforce help agent" pricing: customers pay $2 per resolution (a fully resolved support issue) rather than $2 per conversation, and pay nothing when the agent escalates to a human. His read on why is the actionable part: customers had been spooked by consumption-based pricing horror stories (he cited enterprises that turned on AI tools and "all of a sudden have a bill for $4 million they weren't expecting"), and that friction was slowing Salesforce's Agentforce deals: "I think that has slowed some of their sales cycles. Customers are like, what about the ROI?" The fine print still bites: there's a consumption element (you pre-buy packs of a minimum 1,000 resolutions), and a "resolution" is defined partly by a count of actions within a 10-minute window. Why it matters: this is the first hard corroboration of last week's unconfirmed "Agentforce POC-stall" thread. It cuts two ways: a smart move to unstick demand, but also an admission that customers were not convinced the product paid for itself at the old price.

4. The creator of Claude Code said the big SIs are now his customers. On Odd Lots, "The Creator of Claude Code on The Hottest Piece of Software in the World" (Jul 20), Anthropic's Boris (Claude Code's creator, an operator at the disruptor) named names: "our customers range in size… there's like Airbnb and Ramp… and then also like the biggest companies, there's like Salesforce and Deloitte and Accenture. Like all these very big companies also use Claude Code. And they're saying the same thing. A bigger and bigger percent of the code is being written by Claude Code." Inside Anthropic, he said, roughly 90% of code is now written with Claude Code, with only ~2% (config files) still done by hand, and "I think this is going to go away really fast." He also described roles fragmenting away from "just write code" into prototypers, builders, maintainers, scalers and "sweepers." Why it matters: the productivity multiple that threatens the labor-pyramid model is not theoretical. The pyramids themselves (Accenture, Deloitte) are the power users. Whether that shows up as margin expansion (they keep the gains) or revenue deflation (clients demand the savings back) is the entire bull/bear debate below.

"A bigger and bigger percent of the code is being written by Claude Code."

Boris, creator of Claude Code, on Odd Lots

5. ServiceNow's product chief put a number on adoption, and on why most AI projects fail. On The Product Podcast, "ServiceNow President & CPO on Why the Market Can't Tell AI Winners From Losers" (Jul 22), Amit Zavery (President & CPO, an operator) said ServiceNow's AI business is on track for $1.5 billion this year, up from an original $1 billion plan, with all products now "AI-native." But the more useful data was on the demand side: he described a market where AI spend is rising but most companies "struggle to show real ROI," splitting into a small group of "pace-setters" who reorganize for AI versus a majority taking a "spare-part approach" that "lands up failing more often." He cited 71% of companies struggling with data accuracy, access and management as the number-one blocker. Why it matters: this is the read-through for every systems integrator whose bread-and-butter is enterprise-agent implementation. If most enterprises can't get ROI on their own, that's more implementation and data-plumbing work (bull for SIs), or it's a signal that the whole agent wave is over-hyped and budgets get cut (bear for everyone).

The Debate

Both sides got fresh ammunition this week. Here's the steel-man of each.

Bull: AI grows the pie and the survivors expand margins.

  • The demand isn't gone, it's reprioritized. Krishna's "a third of the slipped deals already came back in three weeks" is the bull's exhibit A: these were real deals, verified with clients, not pipeline fiction (Bloomberg Talks, Jul 23).
  • Most enterprise software, and by extension the implementation work around it, sits on the safe side of the AI line. Krishna claims only ~2% of IBM's software is easily AI-replaceable; the rest gets clients ready for AI (Squawk on the Street, Jul 23).
  • Enterprises can't do it themselves. ServiceNow's Zavery says 71% struggle with their own data and most AI projects fail without a "platform-first" approach and expert help, which is precisely what integrators sell (The Product Podcast, Jul 22).
  • Tech keeps taking share of the enterprise budget. Krishna's long-range call: technology spend rises from ~5-6% of a company's budget today toward 10% by 2035 (Bloomberg Talks, Jul 23).

Bear: AI absorbs the billable work and structurally deflates revenue.

  • The pricing model is flipping to outcomes, and it's happening because the old model wasn't selling. Salesforce's move to pay-per-resolution is a vendor conceding it must guarantee results to close deals, and outcome pricing plays to whoever owns the model, not to the hours-based integrator (Insights for IT Negotiations, Jul 23).
  • The productivity gains are enormous and already in the hands of the SIs' own clients, and of the SIs themselves. If ~90% of a code base can be AI-written, the "put 200 engineers on it" business model is under pressure (Odd Lots, Jul 20).
  • The customer is changing from a human to an agent, and agents don't buy the way humans do. On Avory - Markets and Investing, "Humans Buy Platforms. AI Buys Capabilities." (Jul 21), CIO Sean Emory (a pundit/investor voice) argued software winners will shift from distribution, brand and slick UI toward "reliability, latency, accuracy, cost per call." His acid test for any software (and its implementation ecosystem): "if AI disappeared tomorrow, would customers still buy this platform?" If the answer is no, "you're selling a wrapper."
  • Growth is decelerating regardless of the story. On Modern Value Investing with Sven Carlin, "Accenture ACN Stock Has Been Fairly Punished!" (Jul 20), value investor Sven Carlin (pundit) framed the whole sector risk in one line: the business "might change from growing 10% to 3%. And that is what makes all the difference."

The swing factor: who keeps the productivity gains. If integrators pocket AI savings, margins expand (bull). If clients claw them back through outcome-based and fixed-price deals, revenue deflates (bear). This week's evidence, Salesforce guaranteeing outcomes and clients demanding ROI proof, leans bear on pricing power, even as the demand-recovery data (IBM's returning deals) leans bull on volume.

Stocks in Play

IBM (IBM): discussed heavily and directly (operator commentary from the CEO).

  • Bull: Deals are coming back fast (a third in three weeks; ~three-quarters expected by year-end), free-cash-flow growth intact, and the software mix is 80% recurring/annuity growing ~8%. Mainframe demand is strong. Krishna says 85% of mainframe clients are increasing or holding capacity, and the current Z17 cycle is running 130% ahead of the prior cycle (which itself was ~120% ahead of the one before). Software lags hardware by about a year, so that installed base is a coiled spring (Squawk on the Street, Jul 23).
  • Bear: Consulting is guided to just 1-3% growth. Krishna flatly said "consulting will be in that 1%, 2%, 3%. It is not going to be in the double-digit growers" (Bloomberg Talks, Jul 23). Memory and networking cost inflation could keep diverting client capex. And a $200B-market-cap IBM faces competitors collectively spending over a trillion dollars on compute, and one Squawk panelist asked point-blank whether IBM is "subscale."
  • Next number to watch: the two-thirds of slipped deals that haven't returned yet. Krishna says he'll know "before the end of the year." If the recovery rate beats his own three-quarters baseline, the "one-quarter air pocket" story holds; if deals keep slipping, the structural-decline bears win the argument.

Accenture (ACN): discussed directly, but through a pundit/valuation lens, not fresh operator commentary.

  • Bull: The stock has been hammered. Sven Carlin notes the P/E has fallen from ~42 two years ago to ~10-11, a ~66% de-rating, while earnings are still growing and the dividend yield sits near 4.5-4.75%. At a conservative ~6% growth assumption he pegs fair value around $148 versus a ~$147 price, roughly a 10% expected return. CEO Julie Sweet remains "all-in" on embedding AI for clients (Modern Value Investing, Jul 20).
  • Bear: Same source, harsher read: ~40% of revenue is billable-hours consulting (consultants billing $250-$1,250/hour), the model is heavily acquisition-fueled, and reported free cash flow is flattered by buybacks that mostly offset stock comp rather than shrink the share count. His warning: growth could downshift from 10% to 3%, and "it works until it doesn't." Note the ecosystem risk he flags: ~60% of Accenture revenue rides on its top-10 partners (Adobe, AWS, IBM, Microsoft, Google, Oracle, Salesforce, SAP, ServiceNow), and "these big businesses… they are investing to get rid of Accenture." Separately, Claude Code's creator confirmed Accenture is a heavy user of AI coding tools (Odd Lots, Jul 20), productivity upside, but also the disintermediation risk in one name.
  • Next catalyst: Accenture's own bookings and "GenAI new bookings" disclosure at its next print, plus any commentary on whether AI-augmented work is being repriced. Watch the book-to-bill.

Infosys (INFY): NOT discussed this week. No direct English-language podcast coverage surfaced, despite dedicated searches. This is a persistent, chronic gap (see What Changed vs Last Week below). The only India-IT-adjacent voice was C.P. Gurnani, former CEO of Tech Mahindra, on DGTL Voices with Ed Marx, "Innovate With Constraints (ft. C.P. Gurnani)" (Jul 23), and his comments were philosophical rather than about Infosys specifically. Read-through, clearly labeled as such: Gurnani's central AI view is that "no tool alone can transform a business" and that AI "is not replacing people… it's reshaping how the work gets done," a measured, augment-not-replace stance from a veteran India-IT operator that cuts against the most aggressive deflation narratives. He also framed India's edge as "talent, data, and the capability to innovate with constraints." Useful color on the mindset; not a substitute for hearing from Infosys itself. Next number to watch: Infosys's actual quarterly deal TCV, headcount and fresher-hiring numbers, which land during India IT's mid/late-July earnings window, none of which reached the podcast feed this week.

Wipro (WIT): NOT discussed this week. No direct coverage surfaced. Same India-IT English-podcast gap. The relevant read-through is the same labor-pyramid and pricing-pressure evidence running through the enterprise-software and AI-coding episodes above: if AI compresses implementation cycles (Claude Code) and pricing shifts to outcomes (Salesforce), the offshore, headcount-linked model that Wipro exemplifies is squarely in the crosshairs, but we have no Wipro-specific operator data point this week to confirm or refute. Next catalyst: Wipro's bookings, attrition and utilization on its next print, and any commentary on AI-driven productivity give-backs.

Read-throughs

Salesforce / Agentforce (CRM). The pay-per-resolution launch is the week's most concrete enterprise-agent development (Insights for IT Negotiations, Jul 23). Bullish spin: removing pricing friction should re-accelerate deals. Bearish spin: you only guarantee outcomes when customers won't pay for effort, and Salesforce is itself a heavy Claude Code user (Odd Lots, Jul 20), i.e. leaning on a rival's model. For the SIs, note the implementation read-through: outcome pricing that "plays into the hands of the model owner" is the same worry raised in prior weeks.

ServiceNow (NOW). Zavery's $1.5B AI run-rate (raised from a $1B plan) and "AI-native across all products" framing is a genuine positive data point (The Product Podcast, Jul 22). But his "71% struggle with data" and "most AI projects fail" candor is a double-edged read-through: more integration work for those who can deliver, more risk of budget disillusionment for those who can't.

Workday (WDAY) and SAP / Joule (SAP). No named, standalone adoption commentary this week. They surface only as part of the "enterprise software under AI-disruption pressure" basket that Krishna was asked to distinguish IBM from on Squawk. Apply Krishna's own test (how much of each vendor's value is interaction/UI-based and at risk, versus data/infrastructure and safe) as the working framework until we get vendor-specific evidence.

TCS, Cognizant (CTSH), Capgemini (CAP.PA), EPAM. Zero direct coverage again this week; dedicated searches came up dry. The read-through is identical to Infosys/Wipro: the labor-pyramid and outcome-pricing pressures are visible everywhere except in these names' own voices. Treat all India-IT/European-SI conclusions this week as inference, not testimony.

Microsoft / GitHub Copilot (MSFT). No dedicated hit naming Copilot specifically this week; it appeared only inside generic "Codex vs Claude" comparisons. The relevant MSFT-adjacent signal was a teaser on Squawk on the Street (Jul 23) that Google Cloud revenue grew 82% year-on-year and accelerating, a positive read-through for cloud-implementation demand broadly, and a reminder that the hyperscalers keep gaining share of the enterprise-AI budget the SIs also chase.

In-house AI build-vs-buy. Two vivid data points this week. First, Starbucks replacing IBM's decade-old TriRiga with an in-house AI build, Krishna's own admission that interaction-heavy, un-maintained software is now a build-it-yourself target (Squawk, Jul 23). Second, on the small-company end, the a16z interview with Replit CEO Amjad Masad (The a16z Show, "Amjad Masad on Going Direct, Building Replit" (Jul 17)) included the claim that Saastr founder Jason Lemkin now "runs his entire… multimillion-dollar business with him, another employee, and like 10 agents based on Replit agents." One anecdote, clearly labeled, but it's the labor-pyramid-breakdown thesis in miniature: fewer humans, more agents, real revenue.

What Changed vs Last Week

Last week's issue (Week of Jul 18) was dominated by IBM's shock pre-announcement, a wave of "billable-hour deflation" and "SI-implementation-compression" evidence, and a contrarian Jevons cluster. Here's what moved.

IBM Q2: the open thread closed (mostly). Last week we flagged "watch the FULL Q2 print." It's in, and the CEO's framing is now on record across two shows: capex reprioritization, not structural collapse; a third of slipped deals already back; guide trimmed only to 4-5%; free cash flow held. This updates, and partially rebuts, last week's fear that IBM was the leading edge of a broad discretionary-spend deterioration. The bears aren't refuted (consulting still guided to 1-3%, memory inflation persists), but the "one-quarter air pocket" story now has real numbers behind it. Verdict: bull case strengthened on volume, bear case intact on pricing/consulting.

Agentforce POC-stall: now corroborated. Last week we noted the KeyBank Agentforce downgrade had "no follow-up… neither confirmed nor refuted." This week it was effectively confirmed from the sell side of the negotiating table: UpperEdge says consumption pricing was slowing Agentforce sales cycles, which is why Salesforce switched to pay-per-resolution. Verdict: last week's unconfirmed thread is now supported.

Neoconsulting / outcome-based pricing: the thesis kept building. Last week's sharpest bear evidence (the Data Exchange's ex-IBM AI head on clients demanding 30-50% rate cuts, and outcome-based models favoring hyperscalers) got a real-world instance this week: a flagship vendor (Salesforce) actually shipping outcome-based pricing. The direction of travel is consistent.

ServiceNow adoption: the gap partially filled. Last week we had "no named ServiceNow/SAP/Workday agent adoption coverage (only IBM sell-off collateral)." This week ServiceNow's own CPO gave us a run-rate ($1.5B) and an ROI-reality-check. SAP and Workday remain uncovered.

The SI-uses-the-disruptor's-tools thread: reinforced. Last week's strongest disintermediation data point was operator-sourced (Deltek cutting integration cycles ~50-60% with Claude Code). This week the disruptor confirmed it from the other side: Deloitte and Accenture are named Claude Code customers writing more and more of their code with it.

Still quiet / still a gap: Direct India-IT coverage (Infosys, Wipro, TCS, HCL, Tech Mahindra, LTIMindtree): zero again, for the second week running, right through India IT's earnings season. Cognizant, Capgemini and EPAM, also zero. Accenture Song, no update. This is now a persistent structural blind spot in the English-language podcast feed, not a one-week fluke.