Last quarter the open question was whether the leverage that engineering already had would reach the rest of the company. This quarter it started to. Not everywhere, not evenly, but the non-engineering version of the coding miracle is no longer a promise on a vendor roadmap. It’s a scheduled task running on someone’s laptop in your finance team, producing a report every Monday, that nobody in IT knows exists.
That sentence is the quarter. The leverage is arriving, and it’s arriving through your side door.
This is the Q3 2026 briefing. Same assumptions as before: you have budget authority, no time, and a tolerance for opinions. It updates each quarter because the facts move that fast. The version you read in April is already wrong in three places. This one will be wrong by October. That’s the territory.
The shape of the market didn’t change. The bill did.
The six categories from last quarter still hold: chat interfaces, coding agents, desktop and knowledge-work agents, workspace AI, enterprise agent platforms, and direct API access. If a vendor is selling you something that doesn’t sit in one of those, ask harder questions. Nothing this quarter broke the taxonomy.
What changed is how you pay. The flat per-seat enterprise SKU is gone across the serious vendors, and the replacement is metered. You pay a small seat fee plus consumption at usage rates, against a token volume or annual spend you commit to up front.1 The headline isn’t the rate. It’s that your bill now scales with how much work your people actually push through the tools, not how many badges you bought.
This breaks the mental model most finance teams use. A seat license is a fixed cost you forecast once a year. A usage meter is a variable cost that climbs exactly as fast as adoption succeeds. The better your rollout goes, the bigger the bill. Leaders who ran a successful pilot in Q2 are getting Q3 invoices that surprised them, and the instinct to throttle usage is the wrong one. You don’t want to cap the thing that’s working. You want to see it.
The move this quarter is to manage token spend the way you’d manage any metered utility. Set per-team and per-user budgets. Watch consumption by workspace, not by headcount. The tooling for this exists now: spend limits per user and per org, consumption dashboards, and a per-session cost readout your power users can see while they work.2 Most of the actual savings sit in plumbing your engineers already know about: prompt caching cuts the cost of repeated context by an order of magnitude, and batch processing halves it for anything that doesn’t need an instant answer. You don’t need to become a FinOps shop. You need one person who reads the usage dashboard monthly and can answer “why did this double” before the CFO asks.
Anthropic stopped being the contrarian pick
A year ago, standardizing on Claude was a slightly brave call. It isn’t anymore. Anthropic now takes roughly 40% of enterprise LLM spend by the API meter, leads global LLM revenue, and holds more than half the enterprise coding market.3 Eight of the Fortune 10 are customers. Run-rate revenue crossed $47 billion in the spring. The average enterprise account spends an order of magnitude more per user than the consumer-brand competition, which tells you where the real work is happening.4
The reason matters more than the numbers. Anthropic won the layer above the model. Their Agent Skills format, the packaging that turns a model into a coworker that knows your finance close or your contract review, got opened as a standard this year, and the enterprise plug-in ecosystem grew up around it for legal, finance, accounting, and data science.5 When the integration story is this clean, the procurement story gets boring, and boring is what wins enterprise. The conservative choice and the capable choice are now the same choice.
This does not mean buy only Claude. Your people will still reach for ChatGPT by reflex and Gemini where it’s already in the Workspace they live in. It means the default for anything you’re building on, custom agents, internal tooling, the workflows you want to last, should be Anthropic’s stack unless you have a specific reason otherwise. Standardize the foundation. Let the chat tab be a preference.
The non-technical agent finally showed up
The thing last quarter’s briefing said was twelve months out arrived early for one specific use case: recurring work with a checkable output, run by someone who can’t code.
The mechanism is scheduled tasks. You write a prompt once, pick a cadence, and the agent runs it on its own.6 A weekly competitive scan. A Monday status brief assembled from five systems. A daily reconciliation that flags what didn’t tie out. No API, no engineer, no IDE. The person who used to spend Monday morning compiling the report now reads a draft that was waiting when they sat down. This is the proactive agent pattern, and it’s the first form of AI leverage that doesn’t require the beneficiary to be an AI champion. Someone designs the workflow once. The agent produces value on cadence after that.
Two caveats keep this honest. First, the integration wall is real: the most-cited blocker for agent deployments this year isn’t model intelligence, it’s secure, reliable access to the production systems the agent needs to touch.7 An agent that can’t reach the data is a clever demo. Second, the verifier constraint from earlier briefings still governs everything. Schedule the work where two people can mechanically agree the output is right. Don’t schedule the work where a senior person has to read carefully before anyone trusts it. The agent didn’t change that. It just made the question urgent.
The new risk: load-bearing work IT never approved
Here is the part of this quarter that should get your attention, because it’s the cost of the good news above.
Every one of those scheduled tasks is a small automation running in your environment that no one provisioned, inventoried, or owns. Shadow AI used to mean an employee pasting a deal summary into a free chat tab. That was a moment, a single risky action, and it was recoverable. A scheduled agent is different in kind. It’s durable. It runs on a clock. It accesses real systems on a cadence, it produces output other people now depend on, and it keeps doing all of that after the person who set it up changes teams or leaves.
You are accumulating a second workforce that doesn’t appear on any org chart. The industry data says most companies can’t see it. Roughly a quarter of organizations have full visibility into the agents operating inside them. A small fraction send agents to production with security or IT sign-off. Less than half of running agents are actively monitored.8 Gartner expects AI-related legal claims to clear two thousand this year, and the common thread is missing audit trails for automated actions in regulated contexts.9 This is no longer a hypothetical exposure. It’s a maintenance backlog and a compliance gap that’s already on your books, whether or not it’s on your dashboard.
The plan you’re on decides how bad this is.
On enterprise plans, the controls exist and most companies haven’t turned them on. You can disable unattended agents org-wide with a single admin toggle, or scope them down: which employees may run them, which connectors each role can touch, which branches code agents may write to.10 The Compliance API streams real-time logs of who ran what into your existing SIEM, so a scheduled job becomes an auditable event instead of a rumor.11 Managed settings pushed through your device management lock the configuration so it can’t be loosened locally. If you’re paying for enterprise, the spend already bought you the visibility. The task this quarter is to use it.
On Pro, Max, and Team plans, you have a problem you can’t fully solve with settings, and this is where most organizations will be caught out. These plans run the same scheduled tasks and routines, but the admin visibility and the compliance logging are thin or absent. Which means the load-bearing automation is most likely to be exactly where you have the least ability to see it: on the individual and small-team plans your best people expensed themselves, doing real work your IT department has never heard of. When that person leaves, the report stops, or worse, it doesn’t stop and nobody remembers why it exists. The honest answer for anything that’s become load-bearing is to move it onto a plan you can govern. The seat upgrade is cheaper than the breach, and far cheaper than the outage when an unowned job fails silently.
The four players, briefly
Anthropic (Claude). The enterprise default now, covered above. The bet to watch is whether the agent and skills ecosystem compounds fast enough to stay ahead of the platform vendors bundling “good enough” into tools companies already own.
OpenAI (ChatGPT, Codex). Still owns consumer mindshare and the chat tab your employees open by reflex. The most likely tool your people pick when given a free choice. Behind on enterprise procurement trust and on the integration ecosystem that’s deciding the agent layer.
Microsoft (Copilot). The distribution moat is intact and the “you already pay for it” sale still closes. Product preference still doesn’t match seat count. If you haven’t audited Copilot usage since last quarter, you’re very likely still funding shelfware. That hasn’t changed and it won’t until you look.
Google (Gemini). Strongest where it’s already embedded, Workspace and the data stack. The platform bet, not the chat bet. Worth watching for how aggressively it bundles agent capability into seats you already hold, which is the move that could pressure everyone’s usage pricing.
The four are stratifying, not consolidating. You don’t have to pick one. You do have to know which layer each is selling you.
Opinions worth holding this quarter
Your AI bill is now a usage report, so read it like one. The seat audit that mattered last quarter still matters, but the new instrument is consumption by team. A flat bill told you nothing. A usage meter tells you exactly where the leverage is landing and where seats sit idle. Don’t throttle it. Instrument it. (Evaluating Spend is the mechanics.)
Standardize your foundation on Anthropic. Let chat be a preference. Anything durable you build, custom agents, internal tools, scheduled workflows, should sit on the stack with the cleanest integration story and the strongest enterprise controls. The contrarian premium is gone. (Choosing Tools walks the decision.)
The knowledge-work miracle is real now, but only for checkable work. Recurring tasks with verifiable output, run on a schedule, are producing genuine leverage for non-engineers this quarter. Work that needs a careful human read before anyone trusts it is still waiting on a verifier nobody has built. The dividing line is the whole game. (Driving Adoption sits underneath this.)
Your shadow AI problem grew a clock. It used to be moments. Now it’s standing jobs. The exposure compounds because the automation persists and accumulates dependencies. Inventory it before it inventories you in an incident report. (Managing Risk has the controls.)
The skills gap is still a disposition gap, and now it’s a design gap too. Power users still pull most of the value, and training still doesn’t move that curve much. But the new lever is workflow design. One person who can spot a recurring, checkable task and schedule it creates leverage for a whole team without anyone changing a daily habit. Find those people. (Recognizing Leverage and Selecting Talent.)
The board conversation
Short version that survives the room. AI leverage is now reaching non-engineering roles through scheduled, checkable work, and the spend has shifted from fixed seat licenses to metered usage that scales with how much real work the tools do. Two things follow. The financial story is now a usage story: the bill rises when adoption succeeds, and the job is to instrument it, not cap it. The risk story is new: the same automation creating the leverage is accumulating outside IT’s line of sight, and the defensible posture is an agent inventory plus moving load-bearing jobs onto governed plans. If a board member wants a number, give the pair: consumption by team this quarter against last, and the count of scheduled automations you can actually see versus the count you suspect exist. Both move on their own evidence.
Three things worth pulling
The first is a usage report by team, not a seat count. Pull consumption for the last ninety days and sort it. The teams at the top are where leverage is landing. The seats at the bottom are the audit from last quarter, still waiting.
The second is an inventory of every scheduled task, routine, and standing agent running against your systems, starting with the personal and small-team plans your people expensed. You’re looking for the load-bearing job nobody owns. You will find at least one. Most leaders find several.
The third is the list of three people in your org who are designing these workflows rather than just using the tools. They’re the ones turning recurring work into standing leverage. If you can’t name them, that capability is happening without your input, which means it’s happening without your governance too.
The single most informative move between now and the next briefing is the agent inventory. The seat audit told you where the money was going. The agent inventory tells you what’s actually running your business when no one is watching. Most leaders can’t answer that yet.
Footnotes
-
Across the major enterprise vendors, flat per-seat AI pricing has been replaced by a low seat fee plus consumption billed at usage rates, committed against an annual spend or token volume negotiated up front (typically a 50-seat minimum). Usage billing generally cannot be disabled. Terms current as of mid-2026 and change frequently. ↩
-
Administrators can set per-user and per-workspace spend limits, monitor token consumption by user and project via console dashboards, and expose a per-session cost readout to end users. Prompt caching reduces the cost of repeated context by roughly 90%; batch processing is about 50% cheaper for non-interactive work. ↩
-
Anthropic run-rate revenue crossed roughly $47 billion in spring 2026, with 300,000+ business customers, 500+ accounts spending over $1M annually, eight of the Fortune 10 as customers, and the highest average revenue per user in the category. ↩
-
Anthropic’s Agent Skills format (packaged workflows that give a model task-specific competence) was opened as a standard in 2026, with an enterprise plug-in ecosystem spanning finance, legal, accounting, and data science, deployable under corporate IT controls. ↩
-
Claude Cowork scheduled tasks reached general availability April 9, 2026: write a prompt, pick a cadence (daily/weekly/monthly), no code required. Note the limitation that desktop-run tasks only execute while the machine is awake and the app is open. Enterprise tiers add an analytics API, OpenTelemetry traces, and role-based controls over who can run agents and which connectors each role may use. ↩
-
In 2026 enterprise agent surveys, integration with existing systems was the most-cited deployment challenge (~46%), ahead of model capability. The hard part of agentic workflows is secure, reliable access to production systems, not intelligence. ↩
-
2026 industry surveys found roughly a quarter of organizations have full visibility into the agents operating inside them, a small minority (~14%) send agents to production with full security/IT approval, and under half (~47%) of running agents are actively monitored. ↩
-
Gartner has projected AI-related legal claims will exceed 2,000 by the end of 2026, with recurring causes including missing audit trails for automated actions and inability to explain automated decisions in regulated contexts. ↩
-
On enterprise plans, unattended agents (Claude Code Routines and equivalents) can be disabled org-wide via an admin toggle or scoped by role, connector, and branch. Code agents are restricted to a “claude/” branch prefix by default. Managed settings pushed via device management lock configuration so it can’t be loosened on individual machines. ↩
-
The Compliance API (enterprise-only) provides programmatic, near-real-time access to usage and content logs, suitable for feeding an existing SIEM, so scheduled and unattended agent actions become auditable events. ↩