AI in operations

Not reliably, on TIN's record. Octopus Energy had a person review and send every AI-drafted reply, said there would be no job cuts and went on to announce thousands of new roles. Klarna's assistant was credited with the work of hundreds of agents, then its CEO said the cost-driven push had cost quality and the company began re-recruiting humans. A workload-equivalence figure is a model, not a count of layoffs.

Nobody outside Salesforce can answer that with a measured outcome yet: TIN has no independently confirmed Agentforce customer-service result. The TSA's Ace agent, built on Agentforce, is reported to resolve most routine traveler questions without a human, but the figures come from the joint TSA and Salesforce announcement, and trade press restated them without checking the logs. What the record does show is that the autonomy setting matters more than the product.

Not of a bank running fully autonomous agents in production. The two most-cited figures are company claims: JPMorgan's COIN hours figure traces to a single Bloomberg article and was never independently measured, and Klarna's agent-equivalence figure was reversed. Newer flashes come from the banks themselves. DBS states a time-saving target for its credit-memo agents, not a result, and HSBC gives no baseline or sample size for its resolution-time claim.

Sometimes, and the measured range crosses zero. Pooled field trials at Microsoft, Accenture and a large manufacturer found GitHub Copilot raised completed tasks, with junior developers gaining most. A separate randomized trial by METR found experienced open-source developers were slower on their own repositories, though they believed the tools had sped them up. Most of this evidence measures code-completion assistants, not autonomous agents that open pull requests on their own.

It depends on who is measuring. The striking containment figures mostly come from vendor case studies that TIN could not corroborate with a second source. The one voice deployment a regulator examined, Presto's drive-thru order taker, was found by the SEC to need a human agent entering the order in every instance on its original version, while the company told investors it eliminated human order taking.

Yes, for selling fake agents rather than turning real ones against users. The DOJ and SEC charged the founder of the AI shopping app Nate over an automation rate the indictment calls effectively zero, with contractors completing purchases by hand; the charges are allegations in a pending case. The SEC also settled AI-washing cases against Presto, Rimar Capital and Delphia. Each punished a claim of autonomy the product could not back.

Ask for a documented outcome on work like yours: a named client, a figure with a date, and a source you can follow. In TIN's study of agencies listed in partner directories, only a small minority of first-listed case studies named the client, gave a number and showed a date. Being cited by AI engines lined up with search authority far more than with anything on an agency's case-study pages.

TIN's post on ai automation companies reads named deployments: Decagon at Notion, Kraken at Octopus Energy, UiPath at Omega Healthcare and Tractable at PZU, plus in-house builds at Klarna, JPMorgan and Manulife. Nearly every headline figure originates with the deploying company or its vendor. Klarna is the exception that turned: independent press reported the company re-recruiting humans after its CEO said the cost-driven push had cost quality.

Pricing on outcomes has started. Salesforce launched its Agentforce Help Agent with pay-per-resolution pricing, so customers are charged only when the agent resolves an issue. Its proof point is its own help portal, where the deployer and the vendor are the same company, so the resolution figure is a self-deployment claim rather than an audited result.

Coinbase's CEO said the company cut its internal AI spend while token use sat near a company high, through an internal LLM gateway that defaults engineers to open-weight models, task-based model routing and aggressive caching. Independent analysis argues the caching gain, not the cheaper models, did most of the work. Coinbase has published no per-lever breakdown, and the Chinese-origin models carry a live regulatory caveat.

Rarely, on TIN's record. Cashfree's figure for its Relay agents comes from its own blog. Blend said loans routed through its Autopilot agent moved faster and converted better but disclosed no percentage or time figure. Shopify's Sidekick usage comes from its investor materials. Each flash says so. Read vendor figures as accurately quoted claims, and ask who counted.

In a peer-reviewed Nature Medicine study, frontier general-purpose models outperformed two deployed specialized clinical AI tools, OpenEvidence and UpToDate Expert AI, in every one of its evaluations. It is a single benchmark in a single field, and the study notes these tools entered practice despite scarce independent evaluation.