← Back to Insights

Insight

Ten Thousand Agents, One Door

Ariel Agor
Updated
The cover illustration with subtle ambient motion, for "Ten Thousand Agents, One Door".

Key facts

FactValueSourceAs of
Fortune 100 companies using Gemini Enterprise, per Googlenearly 90%techcrunch.com
Custom agents deployed by SOMPO on Gemini Enterprisemore than 10,000futurumgroup.com
Companies reporting revenue gains from AI, Deloitte survey28%deloitte.com
Companies letting AI act without human approval, at least moderately40%deloitte.com
Date Anthropic's unified Chat and Cowork turns on by default for existing Enterprise orgsNovember 16pymnts.com

Listen · Read by Ariel

0:00 / n/a

Somewhere in your company there is a spreadsheet of AI use cases. I have seen dozens of them, and they look alike. One row per idea. A column for estimated value, a column for feasibility, a column for owner, and a column called status that mostly says "pilot". At some point a steering committee met, plotted every row on a two by two grid, picked the ones in the top right corner, and funded them. That spreadsheet was a sensible document for 2024. In October 2026 three things happened, inside twenty two days, that turned it into a museum piece. The people still updating it are doing careful work on an object the vendors have quietly stopped building for.

My claim is simple and I will defend it for the rest of this essay. The "use case" was never a natural unit of business value. It was a packaging format, invented because early generative AI tools were narrow, and buyers needed a way to match a narrow tool to a narrow task. The products are no longer narrow. When the tool can take any task an employee hands it, the list of generative AI business use cases stops being a plan and becomes a census. You cannot steer a company with a census.

Three announcements in three weeks

On October 8, 2026, at its Gemini at Work event, Google introduced what it calls "a single, universal agent for work." According to TechCrunch, Google Cloud CEO Thomas Kurian said the agent can be given "objectives," rather than step by step instructions. It plans its own work, loads custom skills, connects to Google Workspace, Microsoft 365, Slack, Jira, Confluence, Git, BigQuery, Databricks, Postgres and Snowflake, and talks to Model Context Protocol servers inside or outside a company's network. It picks a model for each task by default, and users can choose third party models, starting with Anthropic's Claude. It gets its own Workspace account and email address. Sundar Pichai said nearly 90% of Fortune 100 businesses use Gemini Enterprise.

Keith Kirkpatrick at Futurum put the headline plainly: Google collapsed its enterprise AI into one agent. His reading is that Google is betting enterprises want one front door, one memory and one governance layer in place of a drawer full of assistants, copilots and point agents. Persistent coworker agents get an address at @agents.company.com, storage and a Workspace account. A Google representative told Futurum those coworker agents will not carry a per seat license.

On October 23, Anthropic did the same thing from the other direction. PYMNTS reported that Anthropic unified Claude chat and Cowork for Enterprise customers across web, desktop and mobile, so users don't have to pick one service before starting. A conversation begins as chat. When the task needs code or file work, it moves into an isolated, Anthropic managed cloud sandbox, then comes back with finished work for review. The setting is off by default in existing organizations and turns on by default on November 16.

On October 29, Deloitte's AI Institute published its State of AI in the Enterprise results, based on 3,500 director to C-suite leaders in 25 countries. The numbers are a mixed bag. 81% report faster decision making and execution. 28% report revenue gains from AI, up from 20% in Deloitte's January report. 40% let AI execute decisions or take actions without human approval, at least to a moderate degree. The line that matters most came from Jim Rowan, Deloitte's US head of AI: "Leaders should look beyond individual use cases and ask whether the underlying work still makes sense."

Read those three together. Google removed the choice of which tool. Anthropic removed the choice of which mode. Deloitte, the firm whose consultants built a good share of the use case spreadsheets in the first place, told its clients to look past them.

Where the use case came from

The use case was a reasonable answer to a real constraint, and the constraint is worth remembering precisely, because once you see it you see why it no longer binds.

In 2023 and 2024 each generative AI product did roughly one thing. A summarizer summarized. A contract tool reviewed contracts. A support copilot drafted replies inside one help desk. Each one needed its own integration, its own data pipeline, its own security review and its own vendor contract. Deployment cost was high and per tool, so every deployment had to justify itself alone. The use case was the unit of justification. It answered one question for the CFO: what does this specific tool do, for this specific team, and what is it worth?

That logic produced a whole management apparatus. Intake forms. Prioritization matrices. Centers of excellence whose main output was a ranked backlog. Vendor bake-offs per use case. Value tracking per use case. Every piece of it assumed that capability arrived in small, separately priced boxes, and that the hard part was choosing which boxes to open.

The grid assumed a fixed supply

The value versus feasibility grid has a hidden premise. It assumes that feasibility is a property of the task. Some tasks are easy to automate, some are hard, and the committee's job is to find the easy and valuable ones. That premise held when each task needed its own build. It falls apart when one agent, already connected to your mail, your files, your tickets and your warehouse, can attempt any task an employee describes. Feasibility becomes a property of what the agent can reach and what it is allowed to do. Those are governance settings. The committee used to discover feasibility. Now it sets it, usually without noticing.

Say your grid has 60 rows and the committee funded the top 12. Under the old model, the other 48 sat waiting for budget. Under a universal agent, an employee with access can try any of the 48 next Thursday, without asking, because nothing in the product distinguishes a funded use case from an unfunded one. The backlog has already shipped. Nobody approved it.

What SOMPO's ten thousand agents tell you

The Futurum piece includes a customer detail that deserves more attention than it got: SOMPO deployed more than 10,000 custom agents. Futurum notes that most of the customer examples reflect earlier Gemini Enterprise deployments, before the unified agent. So SOMPO's number is the use case model carried to its logical end. Every team builds the agent for its task. The count climbs. At some point the count passes the number of processes any human can name.

I do not read the 10,000 as a failure. I read it as the point where the spreadsheet stops being able to describe the company. You cannot plot 10,000 agents on a two by two grid. You cannot run a value review on each. Nobody can tell you which 300 overlap, which 40 contradict each other, or which ones touch customer data. Google's answer, a week later, was to fold all of that into one agent with one identity system and one set of spend controls. That tells you how the vendor sees the problem. They looked at customers with thousands of narrow agents and concluded the fragmentation itself was what needed fixing.

The other Futurum examples point the same way. Bloomberg Media reported a 63% lift in SQL query accuracy during initial development. Commerzbank cut manual document review from 20 hours to one. Both are classic use case wins, and both were measured the old way, one task at a time. They are real. They are also the kind of result that becomes invisible once the same agent that reviews Commerzbank's documents can also draft the follow up email, update the case file and flag the exception to compliance. The win moves from the task to the chain of tasks, and the spreadsheet only ever had rows for single tasks.

Generative AI business use cases after the catalog

If the use case is going away as a planning unit, something has to replace it. Executives still need to decide where to point money, attention and risk. Here is what I tell clients to govern instead, in rough order of how much damage it does to skip each one.

Standing objectives

Kurian's word was "objectives." Take it literally. An objective is an outcome with an owner and a time horizon: close the books in four days, clear every supplier invoice dispute within a week, answer every inbound RFP with a compliant draft before the deadline. A use case describes what the tool does. An objective describes what the business needs to be true. When the agent can plan its own steps, the steps stop being yours to specify, and the outcome becomes the only thing left that only you can define.

The practical change is in who writes the document. Use case lists are written by the AI team and reviewed by the business. Objective lists have to be written by the business, because only the business knows which outcomes matter and what "done" looks like. A finance leader who says "use AI for invoice matching" is still thinking in use cases. A finance leader who says "no supplier waits more than five days for a dispute answer, and I want to see the exceptions" has written something an agent can be pointed at and a person can be held to.

The permission budget

Under the old model, risk was managed by limiting which use cases got built. Under the new model, the agent can attempt anything, so risk sits in what it can reach and what it can do without asking. Google's launch was full of these controls: cryptographically attested agent identity, OAuth propagated permissions, action level audit trails, an Agent Sandbox, and an Agent Gateway that Google calls an AI network firewall, per Futurum. Anthropic's design routes agentic work into an isolated sandbox and returns finished work for review.

Deloitte's 40% figure is the one that should keep a board awake. That many companies already let AI act without human approval at least moderately. And 23% say accountability is unclear when an AI system produces a negative outcome. Those two numbers together describe companies that granted permissions before they designed them. Deloitte's own recommendation is that oversight should match a system's autonomy and potential impact, so systems that act need more controls than systems that only recommend.

I call the replacement artifact a permission budget. It is a short document, owned by an executive, that says which systems an agent may read, which it may write to, which actions need a named human, and what spend triggers a pause. Futurum describes Google's real time, per project spend caps that pause an agent when triggered. That is a permission budget expressed as a product feature. Most companies have the feature available and no budget to put in it.

Unit cost per objective

Costi Perricos, Deloitte's global GenAI business leader, said many firms track their overall AI bill, but "far fewer understand the unit cost underneath it." Deloitte's numbers back him: 71% track total AI inference and API spend, 51% track token volume, and 44% track per unit inference cost.

Use case accounting at least had a natural denominator. You knew what the contract tool cost and how many contracts it reviewed. A universal agent with no per seat license and smart routing across Gemini and Claude models has no natural denominator at all. The bill arrives as one number. The only way to make it legible is to tag spend to objectives, because objectives are the one unit that maps to something the business values. Suppose your agent spend for the quarter is a single seven figure line. Tagged to objectives, it becomes a statement like "the invoice dispute objective cost this much per resolved dispute". A CFO can argue with that sentence. Nobody can argue with an undivided bill.

The work underneath the work

Rowan's full point in the Deloitte release goes further than retiring the spreadsheet. He asked "whether the underlying work still makes sense." Deloitte found most organizations apply AI to existing workflows, speeding them up without changing the work. Only 25% say they use AI to deeply transform the business, and 15% report a significant shift in decision making authority.

This is the trap the use case model built, and it is subtle. A use case takes the current process as given and asks which step to speed up. It never asks whether the step should exist. The approval chain gets a faster drafter. The monthly report gets a quicker summarizer. The handoff between two teams gets an AI to write the handoff note. Each row on the spreadsheet scores well. The process they sit in is the same process you had in 2019, now running faster in places, with the same queues between the fast parts.

A universal agent handed an objective behaves differently, at least in principle. If the objective is "dispute resolved in five days," the agent has no loyalty to the seven step process your shared services center designed. It will do whatever reaches the outcome inside its permissions. That is the opportunity and the danger in one design. The opportunity is that the agent shows you which steps were real and which were ceremony. The danger is that it skips a step that existed for a reason nobody wrote down, like a segregation of duties check that lived only in one manager's habits.

So redesigning the work is the price of giving agents objectives safely. You have to know which steps carry legal, financial or safety weight before you let something optimize around them. Deloitte reports 60% have redesigned jobs around AI capabilities to at least a moderate degree. I would ask each of those companies a narrower question: did you redesign the process, or did you redesign the job descriptions of the people stuck inside it?

Why the vendors want you to stop counting

There is a commercial reason the use case is fading, and it is worth naming so you do not mistake it for neutral progress. A use case is a natural point of competition. Each row on your spreadsheet was a chance for a specialist vendor to win a contract. A single front door, with one memory and one governance layer, removes those openings. Futurum's analysis says Google's multi model support "removes a reason to build a separate orchestration layer," and that dropping seat licensing shifts the cost burden to how Google meters agent activity.

None of that makes the unified agent bad for buyers. It does mean the decision has moved. Under the use case model you made dozens of small vendor choices. Under the front door model you make one large one, and the metering terms of that one choice will matter more to your AI economics than any individual use case ever did. Futurum lists coworker agent metering and pricing as the first thing to watch, and notes that pricing and general availability dates were mostly not disclosed. You are being asked to standardize on a door before you know what walking through it costs.

What to do before November 16

Anthropic's date is a useful forcing function even if you are a Google shop, because it is concrete. On November 16, existing Claude Enterprise organizations that have not touched the setting will find chat and agentic work merged by default. Every employee with access will be able to move from a question to a sandboxed task without choosing a product first. Whatever your use case spreadsheet says, that is the day the spreadsheet stops describing what your people can do with Claude.

Here is the short list I would work through this month.

First, retire the status column. Stop reporting AI progress as the count of use cases in production. A count rewards fragmentation, and SOMPO's 10,000 shows where fragmentation leads. Report progress against named objectives instead, with the owner beside each one.

Second, write the permission budget before the defaults write it for you. For each platform you run, Google, Anthropic, Microsoft, OpenAI, decide in writing which systems the agent can write to and which actions need a named person. Then check that the admin settings match the document. In most companies I see, nobody has compared the two.

Third, tag spend to objectives from the first invoice. If the platform offers per project caps or departmental chargeback, set them up around objectives, even crudely. A rough allocation now beats a forensic reconstruction next year.

Fourth, pick two processes and redo them from the outcome backward. Choose ones where the objective is clear and the risky steps are known. Map which steps carry real weight. Then hand the objective to an agent inside a tight permission budget and watch which steps it treats as optional. You will learn more about your own operations in a month than the use case review taught you in a year.

Fifth, negotiate the door. If you are about to standardize on one universal agent, the metering terms are the contract. Ask how agent activity is counted, what a paused agent costs, and how you leave. Vendors who have not disclosed pricing are still writing those terms, and buyers who ask early get to shape them.

Architect the door, then walk through it

Every executive I talk to wants a list of the best generative AI business use cases for their industry. I understand the impulse. Lists feel like control. A list worked when tools were narrow and every one had to earn its place on its own. In October the largest vendors showed you that era ending. Google folded its products into one agent that takes objectives. Anthropic stopped making users choose between talking and doing. Deloitte, surveying 3,500 leaders, told them to look past the use case to the work itself.

Buying the universal agent is the easy part. Every company in the Fortune 100 will have one within a year, and most already have the license. The hard part is the architecture around it: the objectives that give it direction, the permission budget that tells it where to stop, the cost model that makes its bill legible, and the redesigned processes that keep it from optimizing away a control you needed. Those four things are specific to your company. No vendor ships them, because no vendor knows which of your steps are ceremony and which are the reason you have not been sued.

That is the work Agor AI Advisory does. We sit with your operators and turn the spreadsheet into objectives, write the permission budgets with your risk and finance leads, set up cost tagging that survives the first quarter, and run the outcome backward redesigns on the processes where the payoff is largest. The companies that do this in the next few months will run one door with clear rules. The rest will run ten thousand agents with nobody able to say what any of them are for.

Sources

Google removed the tool choice. Anthropic removed the mode choice. Deloitte said to look past the use case.

This checks the post's claim that three October announcements point the same way: away from choosing tools and use cases one at a time. In 15 seconds the reader sees three different actors each taking a different choice off the table.

  • Three actors, three weeks, one direction: stop choosing tools one use case at a time.
  • When the agent can take any task, feasibility stops being discovered and starts being set by permissions.
  • Only one of these three is a product launch; the other two are a default switch and a warning.
DateWhat happenedWhat it took off the table
GoogleOne front door and one governance layer, but pricing and general availability dates were mostly not disclosed.2026-10-08Introduced a single universal agent for work that takes objectives, plans its own work, and picks a model per task by defaultWhich tool to use
AnthropicOff by default in existing organizations, then on by default November 16.2026-10-23Unified Claude chat and Cowork for Enterprise; a conversation moves into an isolated cloud sandbox when the task needs it, then returns finished work for reviewWhich mode to use (chat or agentic work)
DeloitteIts US head of AI advises asking whether the underlying work still makes sense, which means redesigning processes rather than only speeding them up.2026-10-29Published survey results from 3,500 leaders: 28% report revenue gains from AI and 40% let AI act without human approval, at least moderatelyThe use case as the planning unit (advised, not shipped)

Source: TechCrunch (2026-10-08), Futurum Group (2026-10-09), PYMNTS (2026-10-23), and Deloitte press release (2026-10-29), as cited in the post · verified · as of 2026-11-07

Want this kind of automation working for your business?

Agor AI designs and ships the systems these posts describe, scoped in weeks, not quarters.

Book a Free Strategy Call