The State of AI: Q4 2026

Last quarter’s briefing said AI leverage was arriving through your side door. It did, and the invoice came with it.

GitHub moved Copilot onto metered credits in June. Microsoft will switch on pay-as-you-go billing by default for new Copilot Business licenses on December 1. Accenture’s Lan Guan told Fortune that clients ready to scale are hitting “this unexpected cost wall.” And the 2027 budget your CFO is drafting right now, if Gartner’s survey is any guide, raises AI spend by about 15% and expects most of it to pay for itself inside a year.

That’s the quarter. More money, a meter running underneath it, and a shorter clock on proving it was worth it.

This is the Q4 2026 briefing, written in October while 2027 budgets are still soft. Same assumptions as before: you have budget authority, no time, and a tolerance for opinions. It opens by grading last quarter’s calls, because a forecast is only worth reading if someone checks it afterward. The previous briefing is The State of AI: Q3 2026.

How the Q3 calls held up

Q3 made four calls. Two held. Two held in part, and the parts that missed change what you should do next.

The metered bill held, and arrived faster than expected

Q3 said seats would still anchor the price while agent work landed on a meter, and that a successful rollout would produce a surprising invoice. That is what happened.

GitHub replaced premium requests with token-billed AI Credits on June 1. Seat prices didn’t move. Business includes $19 of credits per seat and Enterprise $39, and once they run out the cheaper fallback model is gone.1 Microsoft bills Copilot Cowork and its other agent work in Copilot Credits on top of the license. From December 1, new Copilot Business licenses bought through resellers come with pay-as-you-go switched on and a default limit of 4,000 credits per user per month.2 Accenture’s analysis found one AI tool’s usage grew 113-fold in about ten weeks, with roughly a fifth of users driving four-fifths of its spend.3

What Q3 underplayed is that unit prices kept falling. Anthropic priced Opus 5.5 below Opus 5 two months after launching Opus 5, and dropped a planned Sonnet 5 price increase.4 OpenAI’s CFO told shareholders that enterprises had “moved from tokenmaxxing to focusing on cost per unit of intelligence.”5 Cheaper tokens still produced bigger bills. Gartner expects the inference cost per agentic workflow to rise more than fivefold through 2028, because more capable workflows consume tokens faster than price cuts take them off the bill.6

So the levers Q3 emphasized, caching and batch discounts, matter less than the question of which workflows run and how often. That is a budget question, so it sits with you.

Anthropic stayed the default, but the lead narrowed

Anthropic’s revenue run-rate passed $65 billion by the end of July, according to an investor update reported by Bloomberg and Reuters, up from $47 billion in May.7 It still led Ramp’s count of paying businesses in July, at nearly 44% against OpenAI’s nearly 40%. The gap is closing. Ramp’s economist credited OpenAI’s GPT-5.6 Sol, and said Anthropic’s Fable 5 underperformed on uptake, which he tied to its price and its data-retention requirements.8 OpenAI’s enterprise revenue passed its consumer revenue this summer, earlier than the company itself forecast.5

The standardization advice held. A four-point lead in paying businesses is still a lead, and the integration and admin story still favors Anthropic for anything durable you build. One line from Q3 was too clean: that the conservative choice and the capable choice were the same choice. For the top model this quarter, data-retention terms made the capable choice the harder procurement conversation. A narrowing lead still supports a default, as long as your contracts leave room to move.

Non-technical agents arrived in volume. The value is still unproven.

The adoption numbers moved the way Q3 expected. KPMG’s Q3 survey of large US companies has 62% building or deploying agents, up from 53% a quarter earlier, and 44% reporting significant workforce adoption, up from 23%.9 Microsoft reported more than 30 million paid Microsoft 365 Copilot seats, up from about 20 million a quarter earlier, and about 40 million agents registered in Agent 365.10

The value side is thinner. The evidence is vendor counts and self-reported surveys, and I found no public numbers on scheduled tasks for non-engineering work producing outcomes anyone measured. Accenture found only 23% of C-suite leaders reporting widespread, sustained value from AI.3 Q3’s dividing line still looks right: schedule the work where the output can be checked. No one has published the results that would prove it yet.

Shadow workflows: the exposure grew, but regulators gave everyone another year

The risk Q3 described got worse. KPMG found the share of companies building controls into their agents fell to 30%, from 43% two quarters earlier, while the share deploying agents rose.9 Forty million agents in one vendor’s registry is what an inventory problem looks like at scale.

Q3 leaned on legal exposure to make the case urgent, and the legal calendar slipped. The EU’s AI Omnibus took effect on July 27 and pushed the high-risk obligations for hiring, credit, and similar uses to December 2, 2027, and to August 2028 for AI built into regulated products.11 Colorado repealed its AI Act before it ever took effect and replaced it with a narrower transparency law for automated decisions, effective January 1, 2027.12 Gartner’s projection of more than 2,000 AI legal claims runs to the end of this year, so it can’t be graded yet.

The case for an agent inventory this quarter rests on cost and continuity. Regulators won’t be the ones forcing it.

Metered AI spend needs its limits set before December 1

The meter has defaults now, and the vendor chose them. Microsoft’s 4,000-credit default arrives on December 1 for new Copilot Business licenses. Its spending policies also apply automatically to agents and services Microsoft adds later, unless an admin turns that setting off.13 GitHub’s included credits now run out instead of falling back to a cheaper model.1 On these surfaces the limit is usually set by whoever configured the account, and often read by no one.

The decision this quarter is who owns each limit and what happens when it’s hit. A cap without an owner either stops useful work in the middle of a busy week or gets raised quietly every time someone complains. An AI spend cap needs a response plan: who gets the alert, who decides whether the work deserves more, and when it stops.

Then build the 2027 AI budget as two lines. Seats are a fixed cost you audit for idle licenses. Metered work is funded by named workflow, each with an owner, an expected volume, and a ceiling. Evaluating Spend has the mechanics for both. A single “AI tools” line is how the Q3 surprise invoice happens again in Q2 2027.

The sentence for the board: “Our AI spend is now two lines: seats we audit for idle licenses, and metered work we fund by named workflow, each with an owner and a ceiling.”

The 2027 AI budget comes with a twelve-month payback clock

The money is there. Gartner’s survey of CFOs, published in late September, has AI budgets rising 15% to 16% on average for 2027, with 92% planning an increase. The terms are tight. 88% want full return within two years, and 55% within twelve months. 54% are funding AI by cutting SG&A headcount and costs.14 Layoffs are already being attributed to AI. Challenger counts AI as the most cited reason for announced US job cuts so far this year, about 21% of the total, though its share fell to about 9% in September.15

A twelve-month clock pulls budgets toward the easiest thing to measure, which is back-office cost cutting. Gartner itself warns that cutting staff before the AI workflows mature could create bottlenecks.14 The bigger risk is booking the savings first. A headcount reduction taken on the strength of a pilot leaves the work with nowhere to go when the pilot stalls.

The decision this quarter is to fund fewer workflows, each with a named owner and a measure you can read within two quarters, and to take headcount savings only behind a workflow that has already run in production. The measure should be what the workflow produces: cycle time, volume handled, work that didn’t happen before. Hours saved is the weakest number you can bring to a CFO on a twelve-month clock. Measuring Returns is the instrument.

The sentence for the board: “We’ll report payback per workflow inside the year, and we take headcount savings only behind workflows already running in production.”

Agents are outrunning their controls

Deployment is rising and controls are falling. That’s the KPMG trend line, and it matches what Q3 described as shadow work: scheduled jobs and standing agents that nobody provisioned, inventoried, or owns. Only 49% of KPMG’s respondents have defined the high-risk uses where an agent may not decide on its own.9 With the regulatory deadlines moved to 2027 and beyond, nothing outside your company will force this work in Q4.

The decision this quarter is to finish the inventory Q3 asked for and add the column most inventories skip: who notices if this stops. Every scheduled task and standing agent gets an owner, the systems it touches, and the decisions it isn’t allowed to make alone. Managing Risk has the controls by plan.

Hold off on buying an agent-governance platform until the inventory exists and Microsoft Ignite has happened. Ignite runs November 17 to 20, and Microsoft has said agent governance is a main theme.16 Whatever gets bundled into licenses you already hold will reset the price of every standalone tool you’re being pitched.

The sentence for the board: “We know which agents run against our systems, who owns each one, and which decisions they aren’t allowed to make alone.”

The four players, briefly

Anthropic (Claude). Still the default for anything durable you build. It shipped Opus 5, Opus 5.5, Sonnet 5.5, and Haiku 5.5 between late July and early October, and it retires older models on short notice: Sonnet 4.5 leaves the API on November 30.4 Workflows built on a specific model now need an owner who handles migrations. At renewal, ask for deprecation notice periods and price cuts passed through to committed spend.

OpenAI (ChatGPT, Codex). The quarter’s momentum. Enterprise revenue overtook consumer, and Ramp’s data shows business adoption closing on Anthropic.58 Still the tab your people open by reflex. Now a credible second source for work you don’t want depending on one vendor.

Microsoft (Copilot). Paid seats grew by about half in one quarter, and the meter is now the story.10 The December 1 default is the most consequential pricing change for most readers this quarter, because it applies to the license your company is most likely to already hold.

Google (Gemini). Strong where it’s already embedded. Alphabet says nearly 90% of the Fortune 100 use Gemini Enterprise, and Google Cloud revenue grew 82% in the second quarter.17 Nothing this quarter changes the advice: buy it where your people already live in Workspace, and don’t build on it by default.

The board conversation

The short version. AI spend for 2027 is going up, and finance expects it to pay back within a year. The bill now has two parts: seats, which we audit, and metered agent work, which we fund by named workflow with an owner and a ceiling. We take savings only behind workflows already running in production. On risk, regulators moved their deadlines to 2027, so the controls are on us: we know what agents are running, who owns them, and what they can’t decide alone. If a board member wants numbers, give three: metered spend by workflow against its ceiling, payback on the workflows funded last year, and the count of agents with a named owner against the count you’ve found. What Your Board Is Going to Ask About AI covers the rest of the meeting.

Three things worth pulling

The first is a list of every AI meter your company is on, with its current limit and the name of the person who can raise it. Copilot Credits, GitHub AI Credits, Claude Enterprise usage, API keys on corporate cards. The two things worth looking for are a meter with no limit and a limit nobody has looked at since setup.

The second is your draft 2027 AI budget split into seats and named workflows. Any line that can’t name its workflow, owner, and measure is the line that will be hardest to defend at the twelve-month review.

The third is the agent inventory from last quarter, with the new column. For each scheduled task and standing agent: who owns it, what it touches, and who would notice if it stopped tomorrow.

The single most informative move before the next briefing is the meter list, before December 1. It tells you how much of your 2027 AI spend is already committed by defaults you didn’t choose.

Footnotes

  1. GitHub, “GitHub Copilot is moving to usage-based billing,” April 27, 2026: AI Credits replaced premium requests on June 1, 2026; Business includes $19 and Enterprise $39 of credits per user per month; seat prices unchanged; fallback to lower-cost models removed once credits are used. Last verified October 9, 2026. ↩ ↩2

  2. Microsoft Partner Center, “October 2026 announcements,” October 1, 2026: usage-based billing on by default for new Microsoft 365 Copilot Business licenses purchased through CSP from December 1, 2026 (moved from November 2), with a default limit of 4,000 Copilot Credits per user per month that admins can adjust. Initial rollout excludes several markets, including Australia, France, Germany, and India. Last verified October 9, 2026. ↩

  3. Sheryl Estrada, “CFOs are hitting a ‘cost wall’ on AI,” Fortune, July 29, 2026, quoting Accenture’s Lan Guan and Accenture research: one internal tool’s usage rose 113-fold in about ten weeks, roughly 19% of users drove about 80% of its spend, and 23% of surveyed C-suite leaders report widespread, sustained value from AI. ↩ ↩2

  4. Anthropic API release notes: Opus 5 on July 24, 2026 at $5/$25 per million input/output tokens; Opus 5.5 on September 22 at $4/$20; Sonnet 5.5 on September 28; Haiku 5.5 on October 7. Sonnet 5’s introductory $2/$10 pricing became standard on August 10, cancelling a scheduled increase. Sonnet 4.5 retires from the Claude API on November 30, 2026. Last verified October 9, 2026. ↩ ↩2

  5. “OpenAI’s enterprise revenue overtakes consumer,” The Next Web, August 14, 2026, reporting CFO Sarah Friar’s shareholder remarks, including the “tokenmaxxing” quote and a run-rate of about $40 billion. ↩ ↩2 ↩3

  6. Gartner, “Gartner Predicts AI Inference Costs Per Agentic Workflow Will Increase More Than Fivefold Through 2028,” press release, August 17, 2026. ↩

  7. “Anthropic revenue run rate tops $65 billion, source says,” Reuters via The Star, August 18, 2026, following Bloomberg; the figure comes from an investor update, not a public Anthropic statement. A run-rate annualizes recent revenue and is not audited annual revenue. ↩

  8. “OpenAI is gaining on Anthropic with business users, new data indicates,” TechCrunch, August 20, 2026: Ramp data for July 2026 shows nearly 44% of its paying business users on Anthropic and nearly 40% on OpenAI, with comments from Ramp economist Ara Kharazian. ↩ ↩2

  9. KPMG, “Q3 2026 AI Pulse Survey,” September 24, 2026: 314 US leaders at companies with $1 billion or more in revenue, surveyed July 24 to August 25. 62% building, deploying, or developing agents (53% the prior quarter); 30% building controls into agents (43% two quarters earlier); 49% have defined high-risk uses where agents cannot decide autonomously; 44% report significant workforce adoption (23% the prior quarter); 43% have token or usage budgets. ↩ ↩2 ↩3

  10. Microsoft FY26 Q4 results as summarized by Office 365 for IT Pros, July 30, 2026: more than 30 million paid Microsoft 365 Copilot seats, up from about 20 million the prior quarter, and about 40 million agents registered in Agent 365. ↩ ↩2

  11. “EU Digital Omnibus on AI enters into force,” National Law Review, July 31, 2026: Regulation (EU) 2026/1744 entered into force July 27, 2026. Stand-alone high-risk obligations apply from December 2, 2027 and product-embedded high-risk obligations from August 2, 2028. Most transparency obligations still apply from August 2, 2026. ↩

  12. Norton Rose Fulbright, “Colorado enacts revised AI law”: SB 26-189, signed May 14, 2026, repealed and replaced SB 24-205 with a narrower automated-decision law effective January 1, 2027, enforced only by the Attorney General. ↩

  13. Microsoft Learn, “Usage-based billing and cost management for Copilot Credits”: Cowork and other services billed in Copilot Credits; spending policies’ “Auto-apply new services” setting is on by default. Last verified October 9, 2026. ↩

  14. Gartner, “2027 AI Budgeting Assumptions: CFOs Misjudge Cost Volatility and Stifle ROI,” September 28, 2026 (abstract): average AI budget increases of 15% to 16%, 92% planning increases, 88% requiring full ROI within two years and 55% within twelve months, 54% funding AI by cutting SG&A headcount and costs. ↩ ↩2

  15. Challenger, Gray & Christmas, “September 2026 job cuts report,” October 1, 2026: AI cited for 3,961 of 43,281 September cuts and 120,136 cuts year to date, about 21% and the leading cited reason for 2026. ↩

  16. Microsoft Security, “Secure what’s next: your guide to Microsoft Security at Microsoft Ignite 2026,” September 30, 2026: Ignite runs November 17 to 20 in San Francisco, with a “Secure AI” track on seeing, governing, and protecting agents. ↩

  17. Alphabet Q2 2026 earnings release, July 2026: nearly 90% of the Fortune 100 using Gemini Enterprise; Google Cloud revenue up 82% to $24.8 billion. ↩