Who finishes the work an AI agent starts

The productivity-software industry spent Monday talking about one idea from several directions: the AI agent, and how much of a business's work it can be trusted to run on its own. The answer that emerged across the day was consistent — an agent can start a task, but a person still has to supervise it and, often, finish it.

That is a useful lens for anyone choosing tools right now. A vendor demo shows an agent completing a job end to end. The announcements below, from the companies building these agents, are more careful. They describe where the agent helps, where it stops, and who is accountable for what it touches. If you are evaluating software this quarter, the gap between the demo and these caveats is the thing to study.

Microsoft treats AI agents as an insider risk

Security teams have a long-standing category called the insider threat: a person who already holds legitimate access and can misuse it, whether deliberately or by accident. The reason insiders are hard to defend against is that they are inside the trust boundary. Their actions look authorised because, technically, they are.

On Monday Microsoft applied that framing to AI agents. The company's position is that an agent acting with an employee's credentials should be watched the way you would watch a privileged insider, because the underlying models cannot be fully trusted. In TechCrunch's reporting, "Microsoft wants enterprises to closely watch the agents that employees use with systems not controlled by AI labs like OpenAI and Anthropic." [1]

The practical implication for a business choosing tools is about logging and scope, not about whether agents are good or bad. An agent that can read your files, send messages, or change records is exercising real access. Before you let one loose on a system, you want to know what it is permitted to touch, what it actually did, and whether you can reconstruct that afterwards. This is the same discipline that applies to any account with broad permissions — see our earlier note on the audit trail nobody thinks about for why that record matters more than the feature list.

When a customer's agent starts the job, the business still finishes it

Salesforce published a related argument from the opposite side of the conversation — not the agent you deploy, but the agent your customer brings. The idea of a "personal agent" is an assistant acting on behalf of a consumer: booking, querying, chasing a refund. The optimistic reading is that these agents reduce the load on a company's support desk by handling routine contact automatically.

Salesforce's view is that the net effect runs the other way. Its headline states that personal agents will drive more contact centre traffic than they offset, and the summary puts the reason plainly: "Personal agents can start a customer's request, but the business still has to finish it." [2] An agent can open a case, gather details and submit a request at machine speed and at any hour. The resolution — the judgement, the exception, the thing that actually satisfies the customer — still lands on the business.

For a company picking support and messaging tools, that changes the sizing question. You are not planning for fewer conversations. You are planning for more inbound, arriving faster, often pre-formatted by software, and still needing a human decision at the end. Tools that collapse scattered channels into one place to work become more valuable under that pressure, which is the theme of customer messages in one inbox.

What Anthropic's own workforce data says about agents

The day's most grounded data point came from SaaStr's read of a report Anthropic published about its own workforce. What makes it worth attention is the method rather than any single claim. SaaStr notes that "the authors used internal headcount data, agent session telemetry, calendar data, Slack patterns, and hiring scorecards, plus 17 employee interviews." [3] That is a company measuring the effect of AI on its own staff using its own operational records, not a forecast.

The headline figures are striking on their own terms: junior technical hiring down from 27% to 9%, agents working six times the hours of humans, and agents waiting on a human 80% of the time. The last of those is the one that ties back to the day's thread. Even inside the company building the agents, and even when the agents run far longer than people do, they spend most of their time blocked on a human. That is not a failure. It is what supervised automation looks like in practice: the machine does volume, the person provides the decision it is waiting for.

If a vendor tells you an agent will replace a function outright, this is a reasonable counterweight to hold up. The firm with the most incentive to show full autonomy reported that its agents mostly wait for people.

Stripe's read on how AI is changing SaaS pricing

Stripe approached the shift from the commercial side. It surveyed platform leaders about how the arrival of AI is reshaping the way software companies compete. The framing is direct: "AI is creating more SaaS competition, and platforms are feeling the pressure." The survey itself, Stripe says, draws on "a Stripe survey of 200 platform leaders" and "shows how they plan to evolve pricing, grow revenue, and retain customers." [4]

For a buyer, the relevant part is pricing. When a product's cost is partly driven by how much AI work it performs, the old flat per-seat number stops telling you what you will actually pay. Usage-based and hybrid models shift the risk of variable cost onto the customer. That is not inherently a bad deal, but it is a different deal, and it rewards reading the metering terms before you sign. The question to ask is simple: what exactly is counted, and can you see the count.

Natural-language analysis lands in the spreadsheet

Google's Workspace updates were quieter but point at the same direction of travel: putting an AI layer directly where people already work. One change lets you query a data warehouse in plain English from inside a spreadsheet. Per Google, "after connecting to BigQuery data with Connected Sheets, you can use Gemini to interact and analyze that data conversationally, right in Google Sheets." [5]

The appeal is obvious — a non-specialist can ask a question of a large dataset without writing a query. The trade-off to keep in view is governance. The moment a conversational tool can reach warehouse data, the question of who is allowed to ask what, and where the answer is stored, becomes a configuration decision rather than an afterthought. The feature is genuinely useful; it also widens the surface you have to govern.

The common thread

Read together, Monday's announcements describe a maturing position rather than a hype cycle. Microsoft says watch the agent like an insider [1]. Salesforce says the agent starts the work but you finish it [2]. Anthropic's own numbers say the agent waits on a person most of the time [3]. Stripe says the pricing of all this is still being worked out [4]. Even Google's in-spreadsheet assistant [5] lands in the same place: powerful, convenient, and in need of a boundary around it.

The lesson for choosing tools is not to avoid agents. It is to prefer the ones that make supervision easy — clear scope, a visible record of what was done, and a human kept in the loop for the decision that matters. That is the stance we take in what AI should and should not do in your business, and it is why 360REV keeps a human decision point and an audit record around every automated action rather than hiding the machine's work. The firms building these agents spent Monday saying much the same thing.