Newsletter · · Ashutosh Agarwal

Accenture Moves Inside Anthropic as AI Pricing Goes Metered - IT Services vs AI - Week of September 26, 2026

Accenture becomes Anthropic's first independent safety evaluator and the stock rallies, while metered, usage-based AI pricing and stalled pilots reshape how IT services get billed, for the week of September 26, 2026.

IT Services vs AI

Week of September 26, 2026: Accenture Moves Inside Anthropic as AI Pricing Goes Metered

TL;DR

  • Accenture found an AI story that isn't about being replaced. It will put a team of "embedded evaluators" inside Anthropic to test the company's models, with each side committing at least $1B over several years. The stock rose about 5% on the Monday after the deal was announced. On CNBC's podcast the hosts summed up the market's logic: if you're a company people think AI will disrupt, "you do a partnership with Anthropic." Accenture reports fiscal Q4 on Thursday, Oct 1.
  • AI is starting to be billed like electricity: you pay for what you use. Salesforce's CEO now lists six different ways to charge for AI. An agent seat costs roughly $500 a month. An Accenture survey says only $1 of every $5 companies spend on AI "tokens" shows up as a measurable financial result. When clients pay by usage but can't see the payoff, they push harder on the price of everything, and that includes consulting hours.
  • More evidence that the "AI replaces headcount" math is messier than promised. Consultants who get called in to rescue failing AI projects say only 10–20% of pilots ever reach production, and token bills have "tripled, quadrupled." Gartner expects half of the companies that cut customer-service staff for AI to hire people back. For offshore providers that is a real, if quiet, positive. India IT still got no direct podcast coverage, for the tenth week running.

What's new

Ranked by how much each one could matter for a portfolio.

1) Accenture becomes Anthropic's in-house safety inspector, and the market pays for it

Where: The Rundown: "Novo's Post-Ozempic Plan, Paramount Gets Closer to Warner Bros Merger" (Sep 21, host commentary). Also Squawk on the Street: 9AM Hour, 9/21/26 (Sep 21, CNBC anchors) and Elon Musk Podcast: "Anthropic delays $2 trillion IPO for safety" (Sep 20, host explainer).

What happened: Anthropic named Accenture its first independent safety evaluator. An Accenture team will "be sitting inside Anthropic and essentially act as a red team," meaning its whole job is to try to break Anthropic's safeguards and check that the AI behaves the way Anthropic says it does. The Elon Musk Podcast adds the structure: both companies are committing "at least $1 billion each to a multi-year venture." The evaluators get access to models while they are still being trained, not just the finished product. The hands-on work is done by Faculty, the AI firm Accenture acquired.

How the market read it: The Rundown's host said Accenture "really needed this win. Their stock was down 30% this year because everyone thought that AI was going to kill the consulting industry," and noted shares up about 5% that morning. On Squawk, one anchor gave the cynical version:

"If you're considered to be a disrupted software-type company, services company, you do a partnership with Anthropic."

Another anchor made the sharper point, which is the one to hold onto. The market hit services firms harder than software because "nobody likes paying these folks," and they had been "benefiting from just kind of creep of layering on of services." The same anchor also asked who will actually distribute Anthropic's products to big companies: "Anthropic has to have something in its S1 about who's going to distribute this stuff." (An S-1 is the filing a company makes before going public. A thefly headline the same day said the Anthropic IPO was pushed to November.)

Why it matters: Last week we wrote that AI labs and cloud giants are putting their own engineers directly into client companies, which cuts consultants out. This is Accenture's answer: become the lab's partner rather than its victim. The deal gives Accenture three things:

  • a new, high-trust line of work (AI assurance: testing, auditing and certifying AI systems);
  • a showcase relationship with the lab most associated with enterprise coding;
  • a story it can tell investors on Oct 1.

Caveats, flagged: The stock figures are host comments, not verified data. The Rundown said the stock was down 30% year to date; a Squawk anchor said it was down 50% "into the late part of May". The Elon Musk Podcast also raised a real conflict-of-interest problem: "You have the entity seeking approval paying the entity granting the approval." Anthropic itself calls direct funding of its evaluators a "stopgap" until pooled or government funding exists. So this revenue stream could shrink if regulators ever build that capacity. It is also safety work, not a big implementation project, so nobody should model it as a bookings driver.

2) Metered AI pricing arrives, and consulting's own survey shows the value is hard to see

Where: Business of Tech: "Vendors Build the Meter: MSP Risk Rises as AI Usage Moves to Consumption Pricing" (Sep 23, host-analyst Dave Sobel). Also Sources with Alex Heath: "Marc Benioff on the AI boom, SaaSpocalypse, and future of Slack" (Sep 22, Salesforce CEO, operator) and The Enterprise AI Show: "AI News for Mid-September 2026" (Sep 21, hosts).

The data points (from Sobel, who flags which ones come from vendors):

  • Accenture's "CIO's Guide to AI Tokenomics", a survey of 750 senior executives, found that only $1 in 5 of enterprise token spend shows up as a quantified financial outcome. Sobel's caveat: "that is a consultancy selling advisory work on this exact problem, so weigh it accordingly." (A token is the unit AI models charge by, roughly a piece of a word.)
  • Gartner projects that by 2028, consumption-based pricing will make up more than 35% of net new corporate legal-tech spending.
  • Global AI spending reaches $2.7 trillion in 2026, up almost 50% year over year, most of it on infrastructure.
  • BambooHR surveyed 1,600 US desk workers. About 35% of their AI time goes to productive work and 42% to fixing errors and rewording prompts, which works out to roughly 20 working days a year per person. BambooHR sells HR software, so this is vendor research.
  • A benchmark of top coding agents on private, production-grade code found the best of them failing more than 60% of the time, mostly from "integration errors and missing requirements."

Sobel's core argument: under a per-seat license, a tool that takes six tries instead of two is the vendor's retention problem. "Under a meter, that vendor has a better month." And "the only complete record of what was consumed sits with the party sending the invoice."

Benioff confirms it from the vendor side. Salesforce will charge on six dimensions: "named users… named agents… usage… consumption… transaction outcomes… and business outcomes. I saved you so much money. I want a percentage of that." He added, "We're going to do more than $50 billion next year… We cannot be limited by one pricing model." The Enterprise AI Show hosts put a price on the new Claude-powered Salesforce agent seat: "it's like $500 per seat."

Why it matters for services: When software is billed by outcome and by usage, the hourly rate card starts to look old-fashioned. Clients who are learning to question every token will question every billable hour too. There is an opportunity here as well: someone has to build the "meter you control," meaning cost governance, token monitoring and FinOps (financial operations for cloud and AI spend). The fact that Accenture published a tokenomics guide tells you it sees this as a new advisory line of business.

3) The AI-pilot hangover: 10–20% reach production, and token bills have "tripled, quadrupled"

Where: Strategy Simplified: "S24E24: 80% of AI Projects Fail. These 2 Firms Explain Why." (Sep 25). The guests are Hugo and Damien, operators at consulting and systems-integration firms. Damien previously "ran IT for one of the big banks." The show is produced with Management Consulted's industry research.

What they said:

  • The client question has changed. Two years ago it was "where should we use AI?" Over the last 18 months it has become "these pilots haven't produced anything measurable."
  • The industry figure the host cited: "only 10 to 20 percent of enterprise AI pilots end up making it into production."
  • Damien on cost: "the cost… in terms of what they've been paying over the last couple of years, is now tripled, quadrupled." Hugo: companies "chipped out on CapEx… But those gates have turned into floodgates," with compute and token spending having "ballooned out of control."
  • Why projects stall: governance (security, MLOps, which means the tooling for running models in production, plus incident and change management) and, above all, "the data pipelines are not ready." For banks running old mainframe systems, building proper APIs "takes a year, year and a half."
  • Their firms often work as "the fire squad who gets called in to fix failed projects, often by the biggest name." Fixes usually mean "the build itself, the model, the tooling, the pipelining, most of it often has to go and be replaced."

Why it matters: This is the strongest bull argument of the week for integrators, and it comes from people who do the work. The hard part of enterprise AI is still the old, boring part: data plumbing, governance and change management. That is exactly what large services firms sell. It also backs up last week's point that clients are getting tighter with AI spending (the Ramp data showed the heaviest AI spenders cutting 10% in August). The catch is that the rescue work is going to specialist boutiques, and they say the "biggest name" firms are the ones whose projects fail. That is a brand risk for the large firms, not a guaranteed revenue win.

4) UiPath's CEO: coding agents are the "giant leap", and process discovery is being turned into a product

Where: 20VC: "Why AI Cannot Replace Humans in Enterprise… with Daniel Dines, UiPath" (Sep 21, UiPath founder/CEO, operator).

What he said:

  • "Coding agents has been the most major giant leap that we were seeing in the past year… since the invention of ChatGPT."
  • An important distinction: "Deploying AI agents is not getting easier today than it was two years ago… But deploying automation has become much easier. Because I can create these automations with AI." AI writes predictable, testable software once, and that software then runs the business, rather than having a probabilistic agent improvise every time.
  • UiPath is launching "cartography," which includes a "cartographer agent" that interviews subject-matter experts while they work and builds the "map of work": every workflow, exception and procedure.
  • On headcount: UiPath has "around 4,000" people and "more than a thousand engineers." Dines says he is "not doing any mass extinction for the pretense of AI." He also said: "If the work, the quality or better of a human can be done by a machine, I will hire today a machine even if it's more expensive than a human." Host Harry Stebbings relayed that Mercor's CEO said Mercor spends "3x the spend of human salaries on inference." That is a secondhand claim and unverified.

Why it matters: Interviewing experts, mapping how work actually happens and redesigning the process has been the front end of every big transformation project for consultants for 30 years. A software vendor now sells an agent that does it. That is a direct example of Theme 8: a software platform absorbing work that systems integrators used to do. Dines's "automation got easier, agents did not" point also suggests where services demand will hold up (agent deployment, governance) and where it will shrink (building routine automation).

5) Customer-service AI is rehiring humans, which is quietly good for offshore providers

Where: touch point podcast: "TP506 - Who Hears the Patient Now" (Sep 23, healthcare-marketing hosts plus a guest from an AI call-listening company; pundit and industry commentary).

What they said: The hosts walked through how Gartner's forecasts have changed:

  • Mar 2025: agentic AI would resolve 80% of customer-service issues on its own by 2029.
  • Jun 2025: half of organizations planning AI-driven customer-service cuts would abandon those plans.
  • Jan 2026: the cost of generative AI would exceed $3 per customer issue. The host's point: "that's more than a cost to have, let's say, you know, an offshore human person respond to the call."
  • Feb 2026: half of companies that cut service staff for AI would rehire, "but under different job titles."

They cited Klarna, whose CEO admitted cost had been "too predominant a factor" and which then started hiring human agents again. The guest's observation from monitoring AI-to-human conversations: "it's not a pretty picture yet."

Why it matters: The business-process outsourcing arms of Infosys, Wipro, TCS and Cognizant were supposed to be the first victims of AI. If AI costs $3 per issue and an offshore human costs less, the substitution slows down, and the human work that comes back often lands with outsourcers. The Gartner figures are relayed secondhand by the hosts. We have not checked them against Gartner directly.

The debate

Bull: AI grows the services pie and fattens margins

  • The hard part is still plumbing, and plumbing is billable. Integrators at the coalface on Strategy Simplified say 80–90% of pilots stall on data pipelines and governance. Fixing a legacy bank's APIs takes "a year, year and a half." That is multi-year implementation work.
  • New categories are opening up. Accenture's Anthropic deal (The Rundown; Elon Musk Podcast) creates AI assurance as a line of business. Metered pricing creates demand for cost governance (Business of Tech).
  • The labs need distributors. "Anthropic has to have something in its S1 about who's going to distribute this stuff" (Squawk on the Street).
  • Human work comes back. Gartner's rehire forecast and the "$3 per issue" figure (touch point) suggest the headcount collapse is slower than feared.
  • Productivity is real for mature teams. A guest from a large services company on ChinaTalk says: "really mature teams, I am seeing like large uplift in productivity." If you sell fixed-price work, that uplift is margin.

Bear: AI absorbs 5–25% of billable work and deflates the model

  • The hourly rate card is under attack from the customer side. The same ChinaTalk guest supports "any efforts to go away from the time and materials sort of like incentive structure," because "if you incentivize a company to bill more hours… they're just going to bill more hours." For now, "most spending is time and materials, hourly." That is the revenue at risk, not revenue that has already been lost.
  • Price cuts are already happening in a related industry. On Agency Giants, Kelly Allison, founder of the UK agency KVA (operator), sells "half price hours" on retainer for anything AI can do. That covers creative, copywriting, content and websites, while strategy work stays at full price. The host: "AI has to drive cost down… that's where the market goes eventually… in like the next five years." Her agency's revenue fell from £300k a month to £20k in two weeks when client budgets were cut. This is a marketing agency, not an IT services firm, but the pricing dynamic is the same.
  • The pyramid gets narrower at the bottom. Allison: "it's really important that you don't just put a junior team with your agentic workforce." She keeps senior staff running the agents and "would never be proud" of a 200-person headcount. That is the fresher-hiring problem in miniature.
  • Platforms take the implementation work. UiPath's cartographer agent productizes process discovery (20VC). Salesforce's "headless" model moves the interface into Claude and Slack (Breaking Analysis). Labs are sending in their own engineers and "creating also consulting companies" (Everyday AI, Ep 866).
  • Clients resent the fee. "Nobody likes paying these folks," and services firms were riding a "creep of layering on of services" (Squawk). Once clients have a metered alternative, that resentment turns into procurement pressure.

Our read: This week pointed toward a split rather than a single winner or loser. Demand for the hard, messy, regulated work (data, governance, assurance, rescue projects) looks durable. Routine, hours-billed build work and junior-heavy delivery are where prices fall. The test for Accenture on Oct 1 is whether its mix is moving toward the first bucket faster than the second one shrinks.

Stocks in play

Accenture (ACN): directly discussed

  • Bull: The Anthropic embedded-evaluator deal gives Accenture a credible "AI beneficiary" story. It covers safety, assurance and distribution for the lab most closely tied to enterprise coding (The Rundown; Squawk). Its own tokenomics research (Business of Tech) positions it to sell AI cost governance. Integrators on Strategy Simplified confirm that data and governance work, Accenture's core, is where AI projects live or die.
  • Bear: The Anthropic deal is a small, safety-focused contract with a conflict-of-interest overhang (Elon Musk Podcast), not a bookings engine. The specialists on Strategy Simplified say they are often cleaning up failed projects led by "the biggest name." The move to metered and outcome-based pricing (Sources/Benioff) puts pressure on the hours-based model.
  • Next catalyst: Fiscal Q4 FY26 results, Thursday Oct 1, before the open (company reporting calendar; last week's issue wrongly said Wednesday). Investor day is expected Oct 14. Watch new bookings and book-to-bill, the GenAI bookings run-rate, headcount versus revenue, any comment on pricing for AI-assisted work, and how the Anthropic economics are described.
  • Non-podcast context (newswire, no podcast source, for orientation only): Wire reports this week showed sell-side price targets rising into the print after last week's downgrades. JPMorgan went to $200 (Overweight), BMO to $200 (Market Perform, noting "demand will remain muted heading into 2027"), and Jefferies to $190 (Hold). Accenture also announced an AWS collaboration for its mid-market Accenture Edge unit and launched Accenture Construct, a capital-projects business.

IBM (IBM): not directly discussed on IBM Consulting

  • Nothing this week covered IBM Consulting, watsonx or Consulting Advantage. The only IBM mention was Building AI Boston: "The Open Accelerator… Guest Stefanie Chiras" (Sep 21). It covered a Boston startup accelerator run with Red Hat and IBM Ventures as a funding source. That is ecosystem building, not consulting economics.
  • Bull (read-through): IBM's hybrid-cloud and Red Hat platform sits in the governed, auditable middle layer. Dines (20VC) and the Strategy Simplified integrators both argue that is where enterprise AI value accumulates. The mainframe-to-API problem they describe is IBM's home ground.
  • Bear (read-through): IBM Consulting is exposed to the same time-and-materials pressure described on ChinaTalk, plus federal procurement's push away from hourly rate cards.
  • Next catalyst: Q3 results Oct 21, after the close (company reporting calendar). Watch consulting signings and the size of the AI book of business.

Infosys (INFY): not directly discussed, tenth straight week

  • The only mention was a passing reference to an Infosys–Chainlink partnership on The Wolf Of All Streets (Sep 24), a crypto show. It is not relevant to the thesis.
  • Bull (read-through): Gartner's "$3 per customer issue" figure and its rehire forecast (touch point) support offshore business-process work lasting longer than feared. The pilot-rescue and data-pipeline work described on Strategy Simplified suits large offshore delivery.
  • Bear (read-through): "Half price hours" for AI-doable work (Agency Giants) and "senior, not junior" agent teams point straight at the bottom of the Indian labor pyramid.
  • Next catalyst: Q2 FY27 results Oct 23, before the open (company reporting calendar). Watch fresher hiring, headcount versus revenue, and large-deal total contract value.

Wipro (WIT): not directly discussed

  • There was no Wipro podcast coverage in the window. The analysis rests on the same read-through as Infosys.
  • Bull: Customer-service rehiring and the cost of AI per ticket favor offshore operations (touch point).
  • Bear: Wipro has more exposure to commoditized application maintenance and hours-based work, which is where pricing is falling (ChinaTalk; Agency Giants).
  • Next catalyst: Q2 FY27 results Oct 15, before the open (company reporting calendar). It is the first Indian large-cap print of the season, so watch discretionary-spend commentary and guidance.

Read-throughs

TCS, Cognizant, Capgemini, EPAM, HCLTech, Tech Mahindra, LTIMindtree. None were discussed directly on English-language podcasts this week. The relevant read-through is the same as for Infosys and Wipro. Offshore support work is getting some relief from the rising cost of AI per interaction (touch point). Routine build and content work is seeing price cuts (Agency Giants). Upcoming reports: Cognizant Oct 28 and EPAM Nov 5 (company reporting calendar).

Salesforce / Agentforce (CRM).

  • On Breaking Analysis with Dave Vellante (Sep 19; Vellante and George Gilbert, industry analysts), Qualitate interviewed 20 Salesforce customers:
    • most "remain an agent force pilot or proofs of concept";
    • they allocate "between 5% and 15% of Salesforce spend to agent force";
    • there is "no evidence yet… of agent force adoption detracting from the Salesforce budget";
    • among customers who have modeled "headless" access, where AI tools such as Claude reach Salesforce without its own screens, "75% expect Salesforce spending to increase."
  • Benioff on Sources with Alex Heath: Claudeforce, Slack Force and Slackbot are all powered by Claude, and "I bet we'll have a dozen interfaces" within a year.
  • The AI Files #117 (Sep 19) notes a new Agentforce reasoning model built on NVIDIA's Nemotron.
  • Services read-through: When the interface moves into Claude and Slack, there is less screen-customization work for implementation partners. The value moves to the data and business-logic layer, which Gilbert calls "the most important piece of real estate in enterprise software for the next 15 years." Integrators that follow it there keep the work. Those tied to user-interface configuration lose it. Small sample size: n=20.

ServiceNow (NOW) / Workday (WDAY). Not discussed in depth. The Enterprise AI Show hosts (episode) described agent pricing settling into three tiers: pay-as-you-go, bundles and negotiated enterprise deals. They also mentioned the Workday pattern of "write up a short summary and put this in Workday." Agents fill in the system of record instead of people doing it. That read-through affects both vendors' seat counts and the partners that train users.

SAP / Joule. On Everyday AI Ep 866 (Sep 21, host Jordan Wilson, pundit), SAP was cited as having launched an "autonomous enterprise" with "50 plus domain assistants orchestrating 200 plus specialized agents." Caution: the episode is a replay of a series the host says was recorded around May 2026, so the figures may be dated.

Microsoft / GitHub Copilot and coding agents. On The Pragmatic Engineer (Sep 23), a member of the GitHub Next team described agents taking on implementation work, with the remaining bottleneck being team collaboration tools. Dines called coding agents the biggest leap since ChatGPT (20VC). The counterweight is Sobel's benchmark: the best agents fail more than 60% of the time on private production code (Business of Tech). Squawk's anchors also noted enterprises "having to clean up code that is just generated and not working that well" (Squawk). Cleaning up AI-written code is services work too.

Build-vs-buy and in-house AI.

  • Wilson argues the 2023 "build or buy" choice is now four choices (build, buy, partner or wait) across four layers: model, workflows, data and business software. He also says labs are becoming consultancies: "these big AI companies are now putting out… engineers into actual companies. And they're creating also consulting companies at the same time" (Everyday AI Ep 866).
  • He cites GDPval, a benchmark that pits AI against human experts on economically valuable tasks: the best model "can tie or do better than 85% of experts." That is a host-cited benchmark and may be dated.
  • On Bisnow Reports (Sep 25), JLL's President of Technology Solutions, Reeves Davis, discussed how commercial real estate clients are using AI themselves. The episode title asks "So What's Left To Sell?", which is the same question every services firm is now asking.

What changed vs last week

Compared with the Sep 19 issue:

  • Accenture sentiment turned around. Last week the sell-side had turned cautious into the print, and a Guggenheim downgrade knocked the stock down 4.1% to $182.41 on Sep 18. This week the Anthropic deal lifted shares about 5% on Sep 21 (The Rundown), and wire reports show price targets rising into Oct 1. Correction: the print is Thursday Oct 1, not Wednesday as we wrote last week.
  • The "labs as consultants" thread got an integrator response. Last week Jay McBain put numbers on hyperscalers funding their own engineers inside client companies: Microsoft $2.5B, AWS $1B, Google $750M. This week Accenture's move shows one defense: partner with the lab instead of competing with it (Squawk). Everyday AI repeats the idea that labs are standing up their own consulting arms (Ep 866). The thread is now two-sided: disintermediation versus alliance.
  • Consumption pricing went from one voice to a chorus. Last week McBain described "base plus consumption" contracts. This week Benioff laid out six pricing dimensions, Salesforce's agent seat came in around $500, and Sobel described a metered model where "the unit of payment stops being the person and becomes the attempt." That confirms and strengthens last week's point.
  • The demand-caution signal was reinforced. Last week Ramp data showed the top 1% of AI spenders cutting 10% in August. This week integrators reported token bills "tripled, quadrupled," only 10–20% of pilots reaching production (Strategy Simplified), and just $1 in 5 of token spend tied to a measured outcome (Business of Tech). Buyers are getting more disciplined, not spending more freely.
  • Headcount decoupling met counter-evidence. Last week Eight Sleep's CEO described "engineers stopped coding" and "250 people making a billion." This week brought the other side: Klarna-style rehiring and Gartner's rehire forecast (touch point), and Dines keeping 4,000 staff with "no mass extinction" (20VC). Both can be true. Engineering-heavy startups are decoupling fast, while customer-facing and regulated work is not.
  • Quiet threads: There was no update on the Cognition/Devin funding round, and the Cursor/OpenAI thread has now been quiet for three weeks. No new Accenture Edge acquisitions came up on podcasts (only the AWS tie-up, reported on newswires). There was still no direct India IT coverage (week ten) and nothing on IBM Consulting.