Read the Verbs

In the previous piece, I gave the advanced student the first cut for evaluating an AI company: does this corporation believe AI is a tool, or does it believe AI is an entity? There is another way to detect exactly the same thing, and it costs nothing. Read the verbs.

Do not begin with the model. Do not begin with the benchmark. Do not begin with the word agent. Certainly do not begin with how many times the annual report mentions artificial intelligence.

Go to the website. Read the product description. Watch the demo. Read the templates the company hands its customers. Then pay very close attention to the verbs. They will tell you what the company actually believes, whether or not the company has noticed what it believes.

01Tool verbs

Tool-oriented AI has a recognizable vocabulary. Help. Assist. Augment. Enhance. Enable. Empower. Accelerate. Support. Recommend. Suggest. Surface. Simplify. Improve.

These are not bad words. They are revealing words. Consider the sentence:

Who has the job? The salesperson. Who updates the CRM? The salesperson. What does the AI do? It helps.

The AI is subordinate to the human action. The human remains the operating entity. The AI is a capability attached to that entity. That is software. You could replace AI with almost any previous generation of enterprise technology and preserve the structure of the sentence exactly:

Nothing fundamental has changed. The technology got better. The ontology did not.

02Entity verbs

Now change the sentence.

Everything changed. The AI now owns the verb. It updates. And once it owns one verb it tends to acquire the rest of them:

It reads the call transcript. It determines whether the opportunity changed. It updates the CRM. It schedules the follow-up. It drafts or sends the appropriate communication. It checks whether the prospect responded. It escalates the account to you when judgment is required.

Now read those sentences again. They sound like a job description. That is exactly the point. If an AI product is truly being conceived as an entity, its description begins to resemble the job description of a human being.

03The job description test

This may be the easiest practical test I can give you. Take the company’s description of its AI and imagine placing it on a job board. Would it make sense?

Figure 01 The same capability, posted two ways

Reads like a job board

Responsible for monitoring the sales pipeline, maintaining accurate CRM records, following up with prospects, identifying stalled opportunities, coordinating next steps with account executives, and escalating high-value exceptions to the VP of Sales.

Describes responsibility

Reads like a brochure

Provides AI-powered suggestions that help sales professionals work more efficiently, surface customer insights, accelerate CRM data entry, and improve follow-up productivity.

Describes functionality

You would immediately understand that the first paragraph describes a job. You might call the person a Sales Operations Manager. Now imagine that no human occupies that position. The AI occupies it. That is entity thinking.

The second paragraph does not read like a job description. It reads like a software brochure. That distinction is enormous. The first describes responsibility. The second describes functionality.

04Features have functions. Entities have responsibilities.

This is the deeper distinction. Software has features and functions. Entities have responsibilities.

A CRM feature might summarize a sales call. A sales-operations entity is responsible for ensuring that the CRM accurately represents what is happening with every opportunity. Those are not the same thing. The feature waits to be invoked. The entity owns a condition of the world.

The feature waits to be invoked. The entity owns a condition of the world.

That is where I would push the advanced student beyond simply looking for action verbs. Look for responsibility verbs.

Figure 02 Three classes of verb, and what each one admits

Subordinate

The human keeps the job. The software is a capability attached to the human.

• help
• assist
• augment
• enhance
• enable
• empower
• accelerate
• support
• recommend
• suggest
• surface
• simplify
• improve

Responsibility

Something has been entrusted. A condition of the world is now somebody’s to hold.

• monitor
• maintain
• prepare
• reconcile
• research
• schedule
• contact
• follow up
• publish
• deploy
• resolve
• coordinate
• escalate

World-changing

Evidence is left outside the conversation. Something is different afterward.

• send
• file
• update
• book
• publish
• deploy
• purchase
• reconcile
• register
• assign
• close
• contact

Those verbs imply that something has been entrusted to the entity. And then look for an even more important phrase: until complete. That is the dividing line. A tool helps you perform a task. An entity can become responsible for the completion of the task.

05Look at who owns completion

Watch what happens to a single sentence as it climbs.

Figure 03 The escalation ladder, and what is acquired at each rung

1. Rung one · tool

   “Our AI helps accountants reconcile invoices.” The accountant owns the verb, the task and the completion. Nothing has moved.

2. Rung two · action

   “Our AI reconciles incoming invoices and flags discrepancies for accountants.” Acquired: the verb. The AI now does the reconciling. But the exception is handed back undifferentiated, so completion is still the accountant’s.

3. Rung three · entity

   “Every night, the Accounts Payable Bot reconciles incoming invoices against purchase orders and receipts, resolves ordinary discrepancies according to company policy, updates the accounting system, and contacts the controller only when an exception exceeds its authority.” Acquired: a responsibility, a schedule, an operating environment, an authority, a definition of normal, an exception condition, and an escalation path.

That is basically an employee. Not legally. Not biologically. Not consciously. Operationally. And operationally is what matters to the corporation.

06Grok Bot makes the philosophy hard to miss

I recently started using Grok Bot, and I find it particularly interesting because its philosophy is exposed almost immediately. Grok Bot launched on 11 August 2026 under SpaceXAI — the entity formerly known as xAI, acquired by SpaceX earlier this year and renamed in July — and its own headline calls the Bots “AI teammates you can give real work to.” SpaceXAI, x.ai/bot and the Grok Bot launch announcement

Read that sentence as a grammarian. Not a tool you can use. Not an assistant that helps you. A teammate, and real work, and the verb is give — the verb of handing something over.

The documentation is more revealing still. It describes persistent Bots with names, jobs, and context that compounds over time. You create one, you describe its job, you grant it access, and you let it perform multi-step work. Bots work in parallel. They message one another. They pass ownership of tasks. They remember role-specific context. They run recurring responsibilities on a schedule or on an event, without waiting for a new prompt.

That is not accidental interface design. That is ontology expressed through product design. You name them. They have jobs. They have memory. They have routines. They have a workplace. They send you messages. They hand work to one another. One can coordinate several others — SpaceXAI’s own design writeup notes that some users made a Chief of Staff Bot responsible for coordinating several specialists.

And then there is the design decision I find the most philosophically honest thing any of these companies has published. The team wrote that the more prominent they made the Bot’s computer, the more the product encouraged users to supervise it — so they decided it should remain the Bot’s workspace, with the interface offering levels of access rather than a cockpit. SpaceXAI, “Designing Grok Bot for a world of persistent agents”

That last point is profound. The old software assumption is: you operate the computer. The emerging entity assumption is: the AI operates the computer, and you operate at the level of intention, responsibility and judgment.

They did not hide the computer to simplify the interface. They hid it to stop turning the human back into the operator.

07“Your job is…”

Listen to how naturally your own language changes when the architecture supports entities. You stop saying:

You start saying:

That is an extraordinary difference. And then you might continue: every morning at 3:00, review my inbox. Determine which conversations require follow-up. Handle the routine ones yourself. If something requires my judgment, message me. Otherwise I don’t want to hear about it.

Notice what you are not doing. You are not designing every step. You are not programming an automation. You are not describing which database field gets written first. You are not building a decision tree. You are describing a job.

Your job is… Those three words are one of the clearest signals of entity computing. Grok Bot’s own onboarding reflects it almost word for word: create a Bot, describe the job, grant access as it asks. Routines then give that Bot a standing responsibility that activates on a schedule or when an event occurs. That is very different from configuring software. You are staffing a function.

08IBM shows the other philosophy just as clearly

IBM remains useful as the contrasting example because its philosophical position is unusually explicit. IBM’s first Principle for Trust and Transparency states that the purpose of AI is to augment human intelligence. Its ethics guidance defines augmented human intelligence as AI that enhances human intelligence rather than operating independently of, or replacing it, and says that AI augmenting human intelligence maintains human responsibility for decisions even when supported by an AI system. IBM, Principles for Trust and Transparency; IBM, What is AI Ethics?

I am not criticizing those principles as irresponsible. I am pointing out what they reveal. Read the verbs. AI augments. AI enhances. AI supports. AI assists. Humans remain responsible. Humans make the decisions. Humans perform the jobs. AI makes humans better at performing them.

That is a coherent philosophy. It is also very nearly the perfect definition of the tool model.

Here is the interesting part. On IBM’s product pages the grammar is already somewhere else. watsonx Orchestrate sells agents that complete multistep tasks, route work and update systems, and that act independently, with no prompt needed at every step. And IBM’s own explanation of how its agents work says that high-level agents break tasks into subtasks and delegate them to other agents. IBM watsonx Agents product and explainer pages

Note carefully where that word sits. IBM is not telling you to delegate to an agent — that phrasing circulates in partner blogs, not in IBM’s own copy. IBM is describing agents delegating to each other. Which is, if anything, the more striking sentence. A company whose first principle is that humans hold the responsibility has shipped a product in which one non-human entity hands a subtask to another non-human entity and expects it to come back done.

There is a smaller tell in the same direction. The 2018-vintage principle says the purpose of AI is to augment human intelligence. IBM’s Responsible Technology page, restated in May 2026, now renders the first Principle of Trust as the purpose of technology being to augment human capabilities. I would not build a thesis on one noun. But the sentence is being quietly loosened, and it is being loosened in the direction the products already went.

That tension is exactly what the advanced student should notice. The world is moving underneath the corporate philosophy.

09Do not merely count the words

There is an important trap here. Do not build a childish checklist where augment equals bad and execute equals good. Marketing departments learn vocabulary quickly. A company can sprinkle autonomous, agent, execute and delegate all over a website while still building ordinary software.

You need to examine the grammar underneath the vocabulary. Ask three questions:

“Our AI helps you prepare the report.” You are the entity. “Our AI prepares the report for you.” Interesting. “Our Financial Analyst reviews the books every Friday, prepares the report, investigates unusual movements, distributes the completed analysis and contacts you only when something requires judgment.” Now I know what I am looking at. The AI has crossed from functionality into responsibility.

10Why grammar is harder to fake than vocabulary

It is worth saying plainly why this test works at all, because the obvious objection is that copywriting is cheap.

It is cheap — but not uniformly. An adjective is a cheap artifact. You can put agentic on a homepage in an afternoon and nothing downstream has to be true. A verb with an object and an escalation path is an expensive artifact. You cannot write it updates the CRM after every customer conversation unless it updates the CRM after every customer conversation, because the sentence names a condition a customer can check on Monday. Grammar of that kind is downstream of the product in a way adjectives never are.

And this is the only road a company has. No corporation can reach into your head and set what you expect of it. It can only act — ship, document, design, publish — and those actions leave artifacts, and the artifacts become the record, and the record is what quietly retrains what everyone expects next. A company’s product copy is not a claim about its philosophy. It is evidence of it. Which is why you should read it like evidence and not like a claim.

11Look for verbs that change the world

There is another level beyond this. Some verbs merely produce information — summarize, explain, analyze, recommend, draft. Those can still belong comfortably inside the tool paradigm. The more revealing verbs change the external state of the world: send, file, update, schedule, book, publish, deploy, purchase, reconcile, register, contact, assign, close.

These verbs leave evidence outside the conversation. The CRM changed. The email was sent. The appointment exists. The invoice was reconciled. The website was updated. The prospect received the follow-up. The code is running in production. Something happened.

That matters, because the ultimate distinction between a chatbot and a working entity is not how eloquently it can describe what should happen. It is whether the world is different after it finishes.

12The pronouns change too

Eventually you will notice that it is not only the verbs. The pronouns begin changing. Tool companies say use AI to… Entity companies increasingly say give it… tell it… ask it… let it… assign it… hand this to…

And eventually: Nico handles that.

At that point something psychologically important has occurred. You are no longer naming the technology. You are naming the entity occupying the responsibility. That is why names matter.

That sounds almost absurd until you use these entities every day. Then it becomes completely natural. The same thing happened with human organizations thousands of years ago. We do not constantly describe someone’s underlying biological architecture. We refer to the entity and its role. Ask Sarah. Sarah knows what to do.

13The advanced student’s test

So when you are considering joining an AI company, investing in one, partnering with one, or simply trying to understand whether its leadership grasps what is happening, read its verbs.

Ignore the adjectives. Powerful. Responsible. Enterprise-grade. Agentic. Intelligent. Secure. Those tell you very little. Look at the verbs, then determine who owns them.

If the customer is constantly the subject — you create, you analyze, you decide, you manage, you update — while the AI perpetually helps, assists, augments, suggests, accelerates and enables, you are looking at the old world with AI added to it.

If the AI monitors, researches, prepares, contacts, updates, coordinates, executes, reconciles, publishes, maintains and escalates, you may be looking at something fundamentally different.

And if those verbs can be assembled into something that looks remarkably like a human job description, pay very close attention. Because the company is no longer merely building intelligent software. It has begun to understand that intelligence itself can occupy a role.

And once intelligence can occupy a role, the question is no longer what can this software help me do?

The question becomes: what am I still doing that I should have already given to someone else?

14The ledger

Already true

Grok Bot ships persistent, named Bots with jobs, compounding memory, scheduled routines, parallel execution and Bot-to-Bot handoff, on a shared cloud computer where each Bot has its own screen. IBM’s first Principle for Trust and Transparency states that the purpose of AI is to augment human intelligence, and its ethics guidance keeps responsibility with the human. IBM’s watsonx agents are simultaneously sold as acting independently with no prompt at every step, and described as delegating subtasks to one another. All of that is on the record, in each company’s own published words.

What has to happen for the test to hold

The grammar has to stay expensive. The test works because a sentence naming a responsibility, an authority and an escalation path is checkable on Monday, while an adjective is not. If a market emerges in which vendors routinely publish entity grammar over tool products and customers do not check, the signal degrades and the advanced student needs a harder instrument — probably the support contract and the incident post-mortem rather than the homepage.

Where I am probably wrong

I may be over-reading the IBM tension. A company can hold an augmentation philosophy sincerely and still ship autonomous agents, on the entirely coherent view that a human remains accountable for a system regardless of how many steps it takes unattended. That is not hypocrisy; it is the standard account of delegation inside human firms too, and it is roughly how every organization has always worked. The honest version of my claim is narrower than the one I have made here: IBM’s product grammar has moved further than its principles page has, and I am reading a lag as a tell. If IBM restates the principle to explicitly cover multi-agent delegation, my tell disappears and the philosophy turns out to have been elastic enough all along.