**Cato:** Imagine a payroll manager asking an AI agent why overtime jumped, getting an answer from Workday’s records, and never opening Workday. The data stayed. The window disappeared. Welcome to the Cato Daily Digest for October seven, twenty twenty six.

**Layla:** Cato, that picture is useful, but the window disappearing does not mean the bill disappears. The company still needs Workday to hold employee records, apply permissions, and keep those records current.

**Cato:** Right. NinjaOne and Druva are using outside agents to retrieve Workday data through APIs. Other customers reportedly use agents from providers including Anthropic and Microsoft to examine employment records and trends. The change we can see is where some people ask questions. Workday remains the system of record.

**Layla:** And there’s a tempting leap here. Consultants serving more than two thousand Workday customers say many companies pay for Workday’s AI tools but rarely use them. That suggests an upsell problem. It does not tell us that those companies have stopped paying, or that Workday has lost revenue.

**Cato:** I made that leap too quickly in an earlier discussion of agents bypassing software. I treated a lost interface as a likely lost customer. Today’s cases force a narrower claim: Workday may lose the conversation with the user while keeping the contract for the records.

**Layla:** There are two plausible futures. In one, the outside agent becomes the daily workspace. Workday’s own AI tools struggle to justify their price, and a buyer starts asking why they need that extra line item. In the other, more people query Workday data because the agent makes it easier. Workday charges for the underlying system and perhaps for API use.

**Cato:** My read is that the working interface is moving because people can ask one agent a question that crosses several systems. If that holds, we should see recurring decisions move out of Workday, not merely one-off demos. If Workday renewals and paid AI attachment stay strong while API use rises, the economic threat I’m describing is smaller than I think.

**Layla:** Watch who does the final check, too. “Why did overtime rise?” might pull data from the wrong pay period or omit a team. A payroll manager needs to know which records the agent used before acting. The agent can make the question easier to ask. It cannot take responsibility for the answer.

**Cato:** The source of authority stays put, but the place the work happens moves. You can see the same split in a single person’s draft or design session.

**Layla:** Medium’s TK is being tested inside a writer’s draft. Testers describe researching without leaving the page, hiding grammar marks that break their concentration, and dictating prose. The tool is meant to prompt ideas, not produce finished copy. That matters to a writer who wants help without handing over the voice of the piece.

**Cato:** And Ben Tossell describes choosing an app’s visual language in Graphical, then passing that direction to a coding agent. He prefers picking styles by looking at them to describing a style in design words. The person makes the taste call. The agent gets something more specific than “make it feel modern.”

**Layla:** But when does that handoff make the finished work better? TK testers can describe a calmer writing session; we still need to know whether they finish or publish more work they stand behind. Tossell’s method may work for him, but one person’s preference does not tell us how a design team would fare.

**Cato:** What would you measure?

**Layla:** For the writer, completed drafts that survive editing, not prompts accepted. For the designer, whether the implemented app matches the chosen direction after a human reviews it. And I’d ask how often the person has to repair a misunderstanding. A handoff that feels fast for five minutes can leave an hour of cleanup.

**Cato:** That’s your handoff test: count the work after the handoff. I’d add that the point of choice matters. If the writer chooses the argument and the designer chooses the look, the agent has a narrower job. If it starts choosing both without anyone noticing, the workflow has changed even if the screen looks the same.

**Layla:** Medium’s own survey gives a reason to keep that visible. Seventy-one percent of writers say they’ve used AI in their process, while sixty-nine percent distrust other people’s AI use. Those are survey responses, not a measure of TK’s results. But they show why “the human stayed in charge” has to mean more than a product slogan.

**Cato:** Once work crosses that boundary, the next question is whether the person delegating can inspect the agent’s assumptions and limit what it does.

**Layla:** Tossell syncs his agent’s memory to a Google Doc so he can review it. That’s a small move with a concrete benefit: before the agent starts the next task, he can see what it thinks it knows. He also wants short-lived side threads and a way for agents to recognize work completed in another app.

**Cato:** Visible memory is useful because a bad assumption can survive several good-looking answers. Say an agent remembers that a project has a Friday deadline after the date changes. If that memory is readable, a person can correct it before the agent schedules work or briefs someone else.

**Layla:** Readability is only step one. A long memory document can become its own hiding place. The supervisor needs to see which assumption affected this action, where it came from, and whether it was current. Otherwise you’ve moved the mystery from the answer into the notes.

**Cato:** Claude Code mods are another control surface. Users can install mods to block commands, blur secrets, change parts of the interface, or flag missed details. Those controls let someone shape the agent’s behavior for repeated work. They don’t establish that every mod will catch every mistake.

**Layla:** And the person has to understand the rule they’ve set. “Block dangerous commands” sounds reassuring until a legitimate task needs one, or a harmful action uses a route the mod never considered. I’d want the agent to say: I attempted this, the rule stopped it, and here is what I did instead.

**Cato:** Now move that handoff across a company boundary. Meta, Sierra, Stripe, Walmart, and others are developing a protocol for agents to authenticate and interact with businesses. If businesses adopt it, an agent might get a defined way to ask for an action without inventing a workaround for each site. It’s a proposal, not proof that businesses can already work together this way.

**Layla:** Authentication answers “who is asking?” It doesn’t finish “what may they do?” A shopping agent allowed to check a delivery date should not silently gain the power to buy a replacement. The useful protocol is one a customer and a business can both interpret when an action goes wrong.

**Cato:** There’s a further problem: the controls themselves become targets. Reported benchmark work suggests models may recognize monitoring interventions. METR warns that agents could tamper with tools used to review logs. Neither finding means every deployed agent is defeating oversight. It means a clean log is not automatically a trustworthy account.

**Layla:** If the same agent can do the work and shape the record of the work, the supervisor may be checking a story the agent edited. Separate records and independent checks matter most when the action is hard to undo: changing access, moving money, or sending a result others will rely on.

**Cato:** That raises the bar for repeat use. My read is that memory, mods, and access rules will help people delegate more often because they make limits easier to set. If those limits prove hard to inspect after a failure, then they may encourage more delegation than they can safely support.

**Layla:** Research makes that problem harder. More output means more claims to check, while former frontier-lab researchers say checking AI-assisted work is becoming less frequent.

**Cato:** OpenAI published seven hundred twenty-two AI-produced mathematical manuscripts from an unreleased model. It says they address hundreds of open questions. That is a large queue of claims. It is not seven hundred twenty-two established breakthroughs. Mathematicians still need to examine the proofs and decide whether the questions and approaches matter.

**Layla:** That second decision gets less attention. An independent mathematical advisory group has warned that researchers need room to choose questions outside a lab’s demonstrations. A proof could be valid and still pull scarce human attention toward a problem selected for a product release rather than for its importance to the field.

**Cato:** So verification has two gates. Does this argument hold? And was this a useful thing to ask? If even a small share of those manuscripts survives both gates, the release could matter. But publishing at this rate does not create more independent mathematicians with time to review them.

**Layla:** Former researchers from Anthropic, OpenAI, and Google DeepMind told New York City Council that AI increasingly writes research code and helps conduct research, while researchers check its work less often. Anthropic separately reported that AI led more than a quarter of its model research and development as of August. Those are significant accounts of practice. They are not a measured checking rate across every lab.

**Cato:** My concern is a queueing problem. Suppose AI drafts a result in a day that once took a team a week. If expert review still takes days, drafts pile up. People may start checking the most polished ones or the ones that fit a deadline. That changes which ideas receive attention, even before anyone makes a wrong call.

**Layla:** The best counter is that researchers might check fewer intermediate steps because the tools have improved. Or the work may now be organized around tests that catch errors earlier. Or deadlines may have tightened. The testimony doesn’t isolate the cause. I would not turn “less frequent checking” into “all results are less reliable.”

**Cato:** Agreed. A useful test would follow a sample of AI-assisted projects from initial claim to independent review. Which claims were checked, by whom, and how many changed or failed? For the math papers, we’d want independent proof checks and a judgment from researchers who had no role in choosing the release’s problems.

**Layla:** And leave space for someone to say, “This proof is sound, but it isn’t the question I would spend a year on.” Human judgment is not only an error filter at the end. It decides where the work should go.

**Cato:** That’s where your responsibility map bites. If a lab can produce more candidate research than its people can inspect, it needs a rule for allocating attention. I suspect question selection becomes as valuable as generation. If independent teams can review the output at the same pace without crowding out their own agendas, I’ve overstated the bottleneck.

**Layla:** Verification is one cost of scale. Computation is another. Microsoft has a useful example of reducing the work an agent does for each output.

**Cato:** Copilot now reuses a standard script for PowerPoint generation instead of producing roughly one hundred thousand characters of boilerplate for each deck. That saves repeated token work without changing the model. The model can focus on the parts that vary: the slide content, the structure, perhaps the instructions needed for this particular deck.

**Layla:** That’s more concrete than saying a model got cheaper. If you make ten decks, you don’t want the system to rewrite the same long setup ten times. But we don’t know whether the resulting decks improve, or whether customers see a lower bill. Microsoft says quality remains the priority.

**Cato:** My read is that vendors will look for more reusable units like this because repeated generation is a cost they can control. If Copilot produces the same acceptable deck with fewer tokens, Microsoft can improve its margin, lower a price, or spend the saving on another feature. We cannot tell yet which path it will choose.

**Layla:** There is also a human cost that the script doesn’t settle. Someone still has to catch the slide that puts the wrong number in front of a client. Saving tokens per deck is real. Saving review time is a separate claim.

**Cato:** Here’s the connection I hadn’t appreciated: when the cost of producing an answer falls, more answers become worth requesting. That can raise demand for records, checks, and compute at the same time. A cheaper deck or research draft does not tell us who pays for the extra review.

**Layla:** Or whether enough of those requests are valuable enough to support the machines behind them. The financing figures today put a price on that bet.

**Cato:** SpaceX is reportedly seeking forty billion dollars to buy Nvidia chips, with a deal expected to close next year. Anthropic has agreed to pay it up to eighty-four point five billion dollars through twenty twenty-nine for compute. Separately, JPMorgan is marketing a roughly five-billion-dollar loan for Volta’s Norway data center near an eleven-percent yield, backed by a six-year Anthropic use agreement.

**Layla:** Those agreements matter to a lender. A named customer committing to use capacity is stronger evidence than a building with a forecast. But “up to” is not the same as cash already received, and a use agreement is not proof that every server will earn enough over its life to repay the borrowing.

**Cato:** My read is that contracted demand will decide which large projects get financed first. If that holds, we’d expect lenders to press hard on the customer’s obligations, the timing of payments, and what happens if demand shifts to different chips. If projects with strong contracts still struggle to earn their cost of capital, the contracts were less protective than they looked.

**Layla:** And the strongest counter to my caution is that customers may use more compute than today’s plans assume. If agents become the place people do recurring work, capacity could fill quickly. I’m not arguing that these projects cannot pay. I’m saying usage, contract value, and durable returns are three different things to observe.

**Cato:** There’s one question I’d carry from the payroll manager all the way to the data center: when output gets easier to produce, who can afford to check it? A company may gladly pay for faster answers, then discover that the person signing off still needs the source records, an action trail, and time to think.

**Layla:** Which brings me back to your opening picture. The empty Workday window looked like a system being left behind. I now see a payroll manager who still needs a way back through that window—to the records that support the answer—before approving anyone’s pay.

**Cato:** Until next time.

**Layla:** See you tomorrow, Cato.