Microsoft’s FY25 annual report shows that GitHub Copilot surpassed 20 million users. After the launch of Coding Agent in May 2025, developers only needed to assign issues to the agent, which then autonomously spins up environments, modifies code, and submits pull requests in the background. The faster developers write code, the more commits, branches, and pull requests accumulate in the repository, directly driving up backend usage.
Yet within the M365 ecosystem of 450 million commercial users, the financial statements tell the opposite story. Public data compiled by Windows Latest shows that M365 Copilot’s paid penetration rate among these customers is around 3.3%, with third-party industry reports estimating 4.5%. Among purchased seats, the proportion of active weekly users is only between 20% and 30%. In December 2025, Reuters cited a report by The Information stating that multiple Microsoft divisions cut sales growth targets for certain AI products by up to 50%; Microsoft denied that overall sales quotas for its AI products were lowered. Business lead Andreou admitted in an internal memo that Copilot must “earn the right to exist”.
Microsoft did not favor one over the other in marketing; both were pushed with full force. The real divergence began with product design before launch. GitHub Copilot launched in 2022 at $10 per month for anyone. Individual developers could try before buying, cancel anytime, and pay voluntarily. A year later, M365 Copilot launched with a completely reversed design: enterprise-only, $30 per seat per month, requiring a base E3 or E5 subscription, and featuring a 300-seat minimum purchase threshold that completely locked out individuals and small teams (a requirement not lifted until January 2024). The same company introduced two opposing go-to-market strategies one year apart: one lowered the barrier to entry to a minimum and relied on voluntary spending by individual users; the other raised the barrier extremely high and allowed only enterprise procurement departments through the door. Marketing efforts were comparable on both sides. The difference lay in the business model: from day one, these were two distinct businesses.
Subsequent moves followed this design, with price hikes moving along two tracks. On the enterprise track, Copilot remained an add-on on top of E3/E5 subscriptions: a base E5 subscription costs $57 per seat per month without Copilot; adding Copilot costs another $30, bringing the total to $87 per seat. On December 4, 2025, Microsoft announced that the base suites themselves would see price increases in July 2026, with E3 rising from $36 to $39 per month and E5 rising from $57 to $60 per month, under the nominal justification of incorporating AI capabilities into the suite. The consumer track was even more direct: on January 16, 2025, consumer M365 underwent its first price increase in twelve years, rising by $3 per month and bundling Copilot directly into the plan across 84 million subscribers, with media reports describing it plainly as “whether they asked for it or not”. The commonality between both tracks is that the price sheet has always had only one format: per-seat markups. Customers cutting seats because agents handle their work and switching to paying based on actual outcomes has never been an option on Microsoft’s table.
Both products are built on the same generation of foundational large models and share similar infrastructure. One lowered the barrier to the minimum and expanded organically, while the other, facing hundreds of millions of users, could only sustain its numbers through price hikes and forced bundling. The root cause of this divergence does not lie in technical capabilities.
Once embedded into actual workflows, the two businesses connect to financial mechanisms moving in entirely opposite directions.
In the GitHub ecosystem, developers write code using GitHub Copilot and invoke Coding Agent to run tests and integrate interfaces. The faster code is produced, the more commits, branches, and pull requests accumulate in repositories, directly driving up GitHub Actions build minutes and package distribution storage. While Microsoft collects a fixed per-seat fee, compute and cloud storage charges continue to accrue in the background. The better the tool does the job, the more usage fees the parent company earns. Efficiency gains and underlying resource consumption always expand in the same direction.
In the M365 ecosystem, however, the financial logic works in reverse. If an office agent can cross-reference spreadsheet data, automatically audit invoices, and produce professional financial reports, the enterprise procurement manager’s first response is to eliminate entry-level positions and cut M365 E5 seat allocations in the next procurement cycle. At a price of $57 per seat per month (soon to rise to $60), the more complete the output delivered by the agent and the more work hours saved, the fewer software seats the enterprise purchases.
When evaluating agent products from big tech companies, the core judgment comes down to one question: the better this agent performs, does the company make more money, or less?
This explains why big tech companies consistently tend to turn agents into a disguised surtax: charging an extra fee on top of existing subscriptions and enforcing bundling. Because the success of an agent directly erodes existing revenue, incumbent giants will not allow it to expand through standalone outcome-based pricing. Instead, they force it onto existing customers to protect the per-head price of legacy seats. Surtaxes also have a ceiling: customers paying separately requires standalone product value, yet the only pricing format Microsoft can offer is a $30 per-seat markup. An E5 seat costs $57 per month, and Copilot adds $30. Once the agent handles half a person’s workload, the seat costs saved by the customer exceed the markup, making seat cuts inevitable. The closer Copilot gets to its advertised capabilities, the sooner that day arrives. This sales approach recovers at most the premium of an assistant. A true outcome-based billing model never appeared on enterprise price quotes from the start.
Some might offer an alternative explanation: the two user groups are inherently different. Developers are accustomed to trying new tools, and if an agent makes a mistake, they can simply rerun it. Office workers handle contracts, reports, and invoices, where a single wrong number carries a high cost; M365 Copilot might sell slowly simply because users are conservative. This argument holds some merit and indeed explains part of the friction. Yet it cannot explain the structure of the price sheet: if Microsoft truly believed agents would replace large amounts of routine office work, enterprise price sheets should have offered outcome-based billing options long ago, allowing early adopters to pay for delivered results. The continuous absence of this option shows that the constraint is not on the user side, but on the seller’s own ledger.
Microsoft management actually sees the direction of technological evolution very clearly. In December 2024, during an appearance on the Bg2 Pod, CEO Satya Nadella stated plainly that traditional SaaS business applications will collapse (Nadella used the word “collapse”), describing their essence as merely “CRUD databases with a bunch of business logic”. Management does not lack forward-looking judgment. The real obstacle is that within a massive organization, no business leader has the authority to let core revenue voluntarily decline.
This revenue structure constraint exists widely across tech giants. To demonstrate technological progress without hurting core income statements, each giant has evolved its own implementation strategy.
Google dismantled agents into search. Google’s profit pillar is search and advertising. In 2025, Alphabet’s total revenue exceeded $400 billion, with year-over-year growth in Google Search and related advertising revenue accelerating from 10% in Q1 to 17% in Q4. To secure its traffic entry point, Google pays Apple approximately $20 billion annually to lock in the default search position (disclosed in court filings), and was permitted to keep the deal in a September 2025 antitrust remedy ruling. If an independent agent directly compares prices, books tickets, and completes orders for users in the background, users no longer need to browse search result pages, let alone click sponsored links. The more complete the tasks an agent performs for users, the more Google loses its most critical ad impressions. Faced with an indispensable core search business, independent agents lost room to survive. After 17 months online, the standalone browser agent project Project Mariner was officially shut down on May 4, 2026, with the decommission page stating “its technology voyaged to other Google products”. In our May 2026 analysis, we pointed out that standalone browser agents struggle to maintain an independent form amid a lack of session ownership and anti-scraping conflicts. Google subsequently distributed relevant agent capabilities into existing product lines: Mariner’s ticket booking capability went to AI Mode, automated browsing moved to Chrome’s Auto Browse, and Shopping took on agentic checkout. These split features had to conform once again to the monetization rules of search ads. According to Ahrefs data, the click-through rate for the top result on search result pages fell 58% over two years, yet Google’s earnings calls confirmed that the monetization efficiency of AI Overviews has matched traditional search. AI Mode, which features end-to-end capabilities, logs about 1 billion monthly queries, accounting for less than 0.3% of total Google search volume. Entering Q2 2026, Google Search growth slowed for the first time in six quarters. Constrained by these financial statements, any standalone agent that could reduce time on page and click-through traffic faces intense internal scrutiny.
Meta placed monetizable agents into WhatsApp. Meta expects capital expenditures of $130 billion to $145 billion in 2026; as of June 30, 2026, its stock price had pulled back about 29% from its August 2025 peak. Its core revenue comes from display ads in Facebook and Instagram feeds. Although AI recommendations already account for about one-third of Facebook feed content and extend user dwell time, deploying a personal agent that makes decisions for users in core feeds would reduce feed refreshing and directly cannibalize impressions that belong to display ads. Meta One Advanced continues to sell impression weight at a $49.99 monthly subscription fee. Agents that can smoothly generate revenue were placed on WhatsApp, free from advertising baggage. The merchant-facing Meta Business Agent has reached 1 million weekly active businesses on WhatsApp and Messenger, while Business AI weekly conversations disclosed earlier by Meta stood around 10 million. Its current billing combines subscription and usage-based models, with plans to transition gradually to pay-for-performance. Mark Zuckerberg stated on the earnings call: “just like the ad system, effectively, we will get paid when we deliver results for those businesses”. Yet consumer personal agents for the general public face a different reality, with Zuckerberg acknowledging on the Q2 2026 call that the team, “we haven’t done that yet”. As long as personal agents lack a monetization path that avoids damaging display advertising revenue, they can only remain in a marginal, free testing state.
Amazon had the best positioning, but lagged in execution. Amazon’s core revenue comes from e-commerce Gross Merchandise Volume (GMV) and fulfillment logistics. In e-commerce scenarios, the better an agent is at comparing prices, selecting products, and placing one-click orders for users, the more Amazon’s merchandise volume and logistics revenue expand. From a revenue logic perspective, Amazon has natural incentives to push autonomous transaction agents, and making agents powerful does not cannibalize its existing revenue base at all. Yet clean business logic does not guarantee delivering a usable product. Alexa+, originally slated for 2024, was delayed by roughly two years and saw extremely low active usage after being made available for free to 200 million Prime members. In January 2026, Amazon chose to force Alexa+ on users; according to user accounts on AFTVnews, frustrated users could only roll back to the legacy version via the voice command “Alexa, exit Alexa+”. This defines the boundaries of our analysis: understanding the revenue ledger is merely a necessary condition that determines whether a company has the incentive and room to sell agents well, but it can never substitute for solid execution in underlying engineering and interaction experience.
Outside the tech giants, willingness to pay for outcome-based agents has already been validated in standalone products. Aggregated data from multiple sources shows that Claude Code reached an annualized run-rate revenue of $2.5 billion in February 2026; in our August 19, 2026 article, we estimated Cursor’s annualized revenue approaching $4 billion. As long as a product delivers end-to-end outcomes and prices by value, legacy software paradigms do not block the expansion of demand.
When business models shift, large enterprises repeatedly downgrade innovative products into surtaxes on legacy businesses because breaking old revenue models exacts an enormous organizational toll. Looking back at historical cases that successfully navigated business model transitions, a company must achieve four things simultaneously: the top executive possesses decision-making authority shielded from short-term quarterly earnings pressures; evaluation metrics for legacy businesses are restructured to shift focus from defending old assets to expanding the new model; management and the board build consensus to willingly endure the financial pain of substantial net profit declines over one to three consecutive years; and a mechanism is established to proactively migrate existing customers and legacy revenue, even allowing the new business to cannibalize the old.
When expanding into streaming, Netflix accounted for its legacy DVD mail-rental business separately and allowed it to decline naturally until shutting it down entirely in 2023, demonstrating the execution path for such proactive migration.
Adobe’s subscription transition laid this cost bare on its financial statements. Between 2011 and 2013, Adobe halted sales of boxed perpetual-license software and transitioned completely to Creative Cloud subscriptions. Adobe’s net income dropped from $833 million in FY2012 to $290 million in FY2013, a 65% decline. More than 50,000 users signed a joint petition opposing the elimination of perpetual licensing, but management withstood the pressure. After weathering 18 months, Adobe reached 3.4 million subscribers, locking in the expansion of long-term recurring revenue.
In contrast, management at Kodak, Blockbuster, and Intel (which became bogged down in its transition to advanced nodes) similarly saw industry trends clearly in the early stages of technological transformation. The real dilemma facing the organization is that no business entity has legitimate authority to let core profitable operations bleed intentionally.
Yet the distinction between Adobe’s historical transition and today’s agent transformation must be made clear. Adobe shifted from perpetual seats to recurring subscription seats, keeping the billing baseline anchored to employee headcount. Today’s agents, by contrast, are shifting the center of value from per-head seats that assist human operators to work outcomes delivered autonomously. The leap from seats to outcomes delivers a far more violent shock to existing revenue structures than the shift to subscriptions ever did.
Among the four requirements, today’s incumbent giants have only met the first; the remaining three require someone to sign off on making core businesses look bad. This is not a failure of vision. Adobe looked bad in its time, and Netflix let DVDs decline. The difference is that the leaders of Adobe and Netflix decided to pay that cost, while today’s giants have not. So when watching any big tech agent launch event, there is no need to get swept away by demo videos and benchmark parameters. Asking a single question is enough: the better this agent performs, does the company make more money, or less? Once that ledger is clear, whether they will genuinely push to sell it becomes obvious.