The code miracle worked. The market knows what can automate code. The open question now — and it is the question shaping every product release, every pricing change, and every vendor pitch landing in your inbox this quarter — is whether the same kind of leverage extends to the rest of the company before the engineers who already have it build it themselves.
Your AI budget is going up again. Some department head is forwarding you a podcast. A vendor wants forty-five minutes “to walk through the agent platform.” Your CFO wants a number. Your board wants a strategy. None of those conversations get better if you read another McKinsey deck.
This is the Q2 2026 briefing. It assumes you have budget authority, no time, and a tolerance for opinions. It updates each quarter because the facts move that fast. The version dated April 2026 will be wrong by July. That’s the nature of the territory.
Most executives are reading this market wrong
The picture you’re getting comes from three places: vendor pitches, McKinsey-style strategy decks, and the tech-press coverage that gets blasted into your LinkedIn feed. All three have the same defect. They describe AI as a thing happening to your industry. A wave. A transformation. A roadmap.
This framing makes you a passenger. It justifies the 30-slide deck and the consulting engagement. It doesn’t help you decide whether to buy ChatGPT Business or Claude Team for your operations group.
The useful frame is operational. AI is a small set of products, made by four serious vendors, sold to your company under specific pricing terms that changed last quarter and will change again next. The interesting questions are: which product, which vendor, which pricing tier, which roles, what to measure. “What does AI mean for our industry” isn’t on that list.
If you’re still in the strategy-deck conversation in Q2 2026, you’re nine months behind the people who started buying and measuring last summer.
The market in six categories
Chat interfaces. Claude.ai, ChatGPT, Gemini. Your people open a tab and type. Good for drafting, analysis, brainstorming. Limited by session memory and by whether the user knows how to set up persistent context. About 30% of users ever do.
AI coding agents. Claude Code, Codex, GitHub Copilot CLI, Cursor. Engineers in a terminal, AI writing and shipping code alongside them. This is the proven force multiplier in the market. Claude Code alone is at $2.5B annualized revenue. Engineering teams using these tools are seeing 2-3x throughput increases that are measurable and repeatable.
Desktop and knowledge-work agents. Claude Cowork. The Codex desktop app for non-coding tasks. The attempt to bring the coding miracle to roles that do not write code. Cowork hit GA in April. Real product, not a demo. Still around a 50/50 success rate on complex cross-application work per independent testing. Promising but early.
Workspace AI. Microsoft 365 Copilot. Gemini in Google Workspace. AI embedded in the tools your people already pay for. Lowest friction to deploy. Lowest ceiling on what it does. Excel and document drafting features are real and produce gains. The chat experience is consistently the worst of the major options.
Enterprise agent platforms. Gemini Enterprise Agent Platform, ChatGPT Workspace Agents, Microsoft Copilot Studio, Salesforce Agentforce. All shipped or majorly updated in April 2026. This is where vendors are racing for the next layer: the place your company builds and runs custom AI agents under your security controls.
API access. Direct model access for engineering teams who build internal tooling. Highest capability, most flexible, requires technical talent. This is where the engineers who already have leverage build leverage for everyone else.
If a vendor is selling you something that does not fit one of these six categories, ask harder questions.
Five things changed this quarter
In order of how much they should affect your spending.
1. Anthropic moved enterprise pricing to usage-based. In April, Anthropic dropped the flat $200/seat enterprise SKU and replaced it with $20/seat plus consumption at standard API rates. The change removes the volume discount and asks customers to commit to estimated monthly spend up front. For heavy users, total cost can double or triple. For lightly-used seats, cost drops to almost nothing. The implication is straightforward: you are now paying in proportion to how much value you extract. Idle seats stop subsidizing power users. Power users get billed for the leverage they generate. Your seat audit just became urgent.
2. The agent platform race opened on the same day. April 22 saw three of the four major vendors announce enterprise agent platforms: OpenAI Workspace Agents, Google Gemini Enterprise Agent Platform, Salesforce Agentforce with Google partnership. Microsoft expanded Copilot Studio the same week. This is a land grab. None of these platforms are mature. All of them want to be the layer your company builds custom agents on for the next decade. Pick a pilot. Do not pick a platform.
3. Cowork went GA. Anthropic’s Cowork moved out of research preview. Real enterprise features now exist: role-based access, granular MCP permissions, scheduled tasks, phone-based control via Dispatch. The product is real. Whether it eats Microsoft 365 Copilot’s lunch in non-coding knowledge work is the live question of the next two quarters.
4. The Copilot adoption numbers got published. Microsoft has 16.1 million paid M365 Copilot seats. That’s 3.9% of their commercial M365 base, two years into the product. When users at the same company get a choice between ChatGPT, Copilot, and Gemini, they pick ChatGPT 76% of the time. Copilot accuracy NPS sits in the negative twenties. If you’re paying for Copilot seats and haven’t audited usage in the last quarter, you’re very likely funding shelfware.
5. The code miracle is no longer up for debate. Claude Code at $2.5B ARR. Codex shipping desktop and computer-use updates monthly. GitHub Copilot at 26 million users. The productivity gains are large, measurable, and accelerating. The only remaining argument is whether the same gains will reach non-engineering roles. The answer is yes, but slower than the marketing suggests.
What didn’t change: the 6x engagement gap between AI power users and everyone else, the skills gap as the number-one stated barrier, and the 74% of organizations that still can’t demonstrate measurable ROI. These three numbers have been stable for three quarters. They’re structural, not transitional.
The four players
Anthropic (Claude). $14B annualized revenue, $380B valuation, 80% of revenue from enterprise. Claude Code alone is $2.5B ARR and is the single best evidence in the market that AI delivers force multiplication. The strategic bet, and the one to watch: Cowork has to extend that leverage to every other knowledge worker before the engineers Anthropic just armed with Claude Code build the same thing internally and skip Cowork entirely. Best technical reputation. Most enterprise-friendly procurement. Pricing model now favors heavy use.
OpenAI (ChatGPT, Codex). 800M users, $13B revenue, $17B projected cash burn for 2026. Owns consumer mindshare and the default chat brand. Workspace Agents (April 22) is their answer to Cowork: cloud-based shared agents that integrate with Slack and keep working when nobody is watching. Stronger in image generation and multimodal breadth. Slightly behind Anthropic in coding benchmarks and enterprise procurement trust. The most likely tool your employees will pick if you give them a choice.
Microsoft (Copilot). 16.1M paid M365 Copilot seats, 26M GitHub Copilot users, presence in essentially every enterprise. The “you already pay for it” sale. Distribution moat is real. Product preference isn’t. The all-in Copilot strategy is a bet on procurement convenience over outcomes. That bet has worked before in enterprise software. It’s working less well here.
Google (Gemini). Just relaunched its enterprise stack as Gemini Enterprise Agent Platform, including a low-code Agent Studio, multi-day agent runtime, and persistent agent memory. Also supports Anthropic’s Claude models alongside Gemini, which signals where Google thinks the puck is moving. Strongest in infrastructure, data integration, and Workspace-native users. Distant third or fourth in coding tools. The platform bet, not the chat bet.
These four are stratifying, not consolidating. You don’t have to pick one. You do have to know which tier each is selling you, and why.
Opinions worth holding this quarter
A handful of opinions. State them, defend them, change them next quarter when the facts change.
The skills gap is a disposition gap. Power users aren’t better trained. They have a temperament: they delegate, iterate, and treat the model as a capable but literal collaborator. The other 80% treat it as a search engine or ignore it entirely. Training programs don’t move this curve. Hiring does. So does manager modeling. Your org’s “AI training plan” is probably the wrong intervention.
Microsoft Copilot is the most expensive free trial in enterprise history. 3.9% paid penetration after two years. Negative accuracy NPS. 44% of lapsed users stopped because they don’t trust the answers. Copilot has its uses. The all-in Copilot strategy isn’t justified by the data. Audit usage. Kill idle seats. Redirect to tools your people choose to open.
Agents are about 12 months from being real. The April platform race was a land grab, not a maturity signal. Today’s agents fail 40-50% of the time on complex multi-app workflows. State management is fragile. Security implications of autonomous AI acting on enterprise data are unsolved. The right posture is to pilot now, with low-stakes workflows, so you have specification and oversight muscle when the platforms mature.
The code miracle is real. The knowledge-work miracle isn’t. Yet. Engineering productivity gains from AI tools are 2-3x and accelerating. The same gains don’t exist for marketing, finance, ops, HR, or legal. The reason is structural: code has tight feedback loops, verifiable output, and self-selecting users. Knowledge work has slow feedback loops, subjective quality, and users with wildly varying disposition. The gap will close. It hasn’t closed.
The real leverage is organizational, not personal. “Claude helped me write this report 3x faster” is the small version. The consequential version is the engineer who used Claude Code to rebuild your reporting pipeline in a weekend and eliminate twenty hours of weekly manual work for the ops team permanently. Or the PM who built an internal tool that replaced a $50K SaaS contract. Force multiplication is one person’s output removing bottlenecks for fifty others. Optimize for that.
Shadow AI is your active policy problem. While legal drafts a 40-page governance framework, your people are pasting client data into free-tier ChatGPT because the approved tool is worse than the free one. Policy documents don’t govern behavior. Available tools do. Give people something good enough that they don’t need to go around you.
74% can’t show ROI because they’re measuring the wrong thing. Cycle time compression, capability expansion, eliminated toil. Anthropic’s own internal data shows 27% of AI-assisted work consists of tasks that wouldn’t have happened without the tool. A P&L line designed for cost savings can’t measure that. If your CFO is asking for a traditional ROI model, the honest answer is that the instrument doesn’t measure this. Build a different one.
The board conversation
The version of this briefing that survives a board meeting is short. AI leverage is proven in engineering and in a handful of knowledge-work roles. The investment is in the people who demonstrate that leverage and the tools that enable it. The risk this quarter is no longer spending too much; it is spending without measuring whether the leverage actually arrives. If a board member pushes for a number, the defensible pair is the seat audit from your most recent quarter and a target percentage of seats producing observable output. Both move on their own evidence. Neither is theatre.
Three things worth pulling
The pricing terms of every AI vendor your company currently pays now read differently than they did in March. The headline number is rarely the actual number, especially after April. Your own seat-level usage data for any AI tool licensed for more than ninety days will tell you more than any vendor’s quarterly review — pull it yourself. And the names of the three people in your org getting the most leverage from AI right now are worth knowing. If you can’t list them, the rest of the program is hard to steer.
The single most informative move available between now and the next briefing is a seat-level usage audit on whichever AI tool you have the most seats of. Sort by activity, look at the bottom third, and decide what to do about it. The Anthropic pricing change makes the conversation defensible. The Copilot adoption data makes it overdue. That audit will tell you more than the next vendor pitch deck.