Gartner told HR chiefs on October 28 that 30 percent of employees replaced by AI will be rehired by 2029, often at a higher cost. The forecast lands while most boards still score agentic AI value by one number. That number is the headcount it removes. The companies with the strongest published returns measure something else.
Anthropic published a study of its own workforce on October 16. Agents now log six interactive hours for every employee hour. Code commits per engineer run at five times their pace of six months earlier, and the company is still hiring. So the agentic AI value in that study shows up as speed, security, quality and work that was never worth doing before. The ten benefits below run in the order they tend to appear, fastest first. Each carries one number and its source.
1. Speed: cycle time collapses first
A European health-tech platform told Anthropic it ships features 40 percent faster after rolling out coding agents. Code quality held, the company said. A global pharmaceutical company cut clinical study documentation from over ten weeks to ten minutes, according to the same report. Neither case lists a layoff. Both list a calendar.
2. Reach: work that was never worth doing gets done
Anthropic’s December survey of 132 engineers found that 27 percent of agent-assisted work would not have happened otherwise. Documentation, test coverage, dashboards and small fixes always lost the priority vote. Now they get picked up. Another 8.6 percent of coding tasks were papercut fixes nobody would have scheduled. That is new output, not cheaper output.
3. Security: vulnerability debt gets paid down
Partners in Anthropic’s Project Glasswing verified at least 129,000 software vulnerabilities between April and July 2026. More than 33,000 were rated critical or high. Engineers at a major open-source browser maker reported fixing 271 vulnerabilities in one release found with the model, ten times the prior release. No bank books a breach that did not happen as labor savings. It is still money.
4. Risk: losses stop before they post
A partner bank used the same model to stop a fraudulent $1.5 million wire transfer, Anthropic said in May. The attacker had compromised a customer’s email and placed spoof calls. The pharmaceutical company’s documentation agent halved review cycles because its output arrived compliant, the agents report says. In a regulated shop, fewer findings per audit is a line item. So is a wire that never leaves.
5. Quality: output holds when the harness holds
Anthropic compared agent-written code with human-written code from the same engineers. It found no difference in the rate of lines later reverted. A global beauty company reported 99.9 percent accuracy on conversational analytics, up from 90 percent with its earlier tools. But the counterweight is real. Google’s 2025 DORA report found AI adoption raises throughput and instability together. So quality depends on the eval suite around the agent, not the model.
6. Infrastructure: server bills fall
A hyperscale cloud provider’s best-known agent result is routinely misread as a labor story. Its Java upgrade agent moved tens of thousands of apps to Java 17. The company reported $260 million in annual savings, but most of it came from upgraded apps needing fewer servers. Teams kept the 4,500 developer-years the agent saved. Its chief executive said developers “hate this tedious though important work.”
7. Revenue: sales move when the calendar moves
The pharmaceutical company told Anthropic that each day of delay on a clinical study report could cost up to $15 million in lost revenue. Inside Anthropic, one account executive now talks to 30 percent more customers, the October study says. On the revenue line, agentic AI value comes from time to market. It also comes from salespeople who spend their hours selling.
8. Resilience: operations run around the clock
A large enterprise software vendor told shareholders in its 2025 annual report that AI agents now perform 90 percent of its internal support requests. A large online university cut one incident’s fix time from about two hours to 28 minutes, the cloud provider said in March. Both figures are vendor-reported. Both measure uptime, not payroll.
9. Decisions: data arrives faster
The beauty company has 44,000 monthly users sending 2.5 million messages to an analytics agent, the agents report says. Employees now query data directly instead of filing dashboard requests. Data analysis was the top non-coding impact for 60 percent of the 500 technical leaders Anthropic surveyed. A decision made on Tuesday instead of next quarter has a price.
10. Knowledge: expertise stops leaving with people
Anthropic’s standing instructions to agents grew from a few hundred to 8,500 in a year, the October study says. The health-tech platform’s agents update technical documentation on every code change. Leaders in the agents survey said employees now spend more time on skill development, 70 percent, and strategic work, 66 percent. At Anthropic, 44 percent of agent-assisted work was tasks people would not have enjoyed. What the company knows, kept in a governed harness outlives the next reorg.
The ten benefits at a glance
| No. | Benefit | One number | Reported by |
|---|---|---|---|
| 1 | Speed | Features ship 40 percent faster | Health-tech platform, via Anthropic |
| 2 | Reach | 27 percent of agent work is new work | Anthropic internal survey |
| 3 | Security | 129,000 vulnerabilities verified | Project Glasswing partners |
| 4 | Risk | $1.5 million fraudulent wire stopped | Partner bank, via Anthropic |
| 5 | Quality | No rise in reverted code | Anthropic employment study |
| 6 | Infrastructure | $260 million a year, mostly servers | Cloud provider, own blog |
| 7 | Revenue | $15 million per day of delay avoided | Pharma company, via Anthropic |
| 8 | Resilience | 90 percent of support requests handled | Software vendor, annual report |
| 9 | Decisions | 2.5 million data queries a month | Beauty company, via Anthropic |
| 10 | Knowledge | 8,500 standing agent instructions | Anthropic employment study |
The cost-only deployments are coming back
The table above is one side. The evidence against the headcount frame is just as public. Gartner expects more than 40 percent of agentic AI projects to be canceled by the end of 2027, citing unclear business value. Forrester found 55 percent of employers regret layoffs they blamed on AI. A Robert Half survey cited by Fast Company found 32 percent of hiring managers cut a role for AI, then rehired for it. METR’s controlled trial found experienced developers were 19 percent slower with early-2025 tools. None of this arrives by default.
What I take from the numbers
I spent twenty years selling IT services on rate cards, so I know the arbitrage reflex well. A cheaper hour was the whole pitch. Agentic AI breaks that pitch because the hour is no longer the unit. The completed task is, and the task now ships with evidence attached.
In my work in regulated loan origination, the agentic AI value never showed up on the headcount line. It showed up as files that moved in hours instead of days. It showed up as an audit trail the examiner could read, and fewer exceptions reaching a human desk. The underwriters stayed. They stopped re-keying.
So my position is flat. If the business case for an agent is headcount and nothing else, I would cancel it before Gartner counts it. The returns above come from a governed loop that measures speed, risk, quality and reach, then reinvests the capacity it frees. Instructions in, results out was IT. Intent in, outcomes out is agentic AI, and labor arbitrage was never the outcome. One hesitation, though. I am not sure the rehiring numbers hold for back-office roles the way they do for engineers. The Five-Stage Roadmap is how I decide which agents have earned the capacity to find out.
Which of these ten returns does your agentic AI value case measure today? And which one does your CFO actually believe?
Subscribe to the Blog: I publish a new post every weekday on one agentic AI governance question. Enter your email and each post arrives in your inbox the moment it goes live.
My Books: Three books on agentic AI for regulated enterprises: Intent In, Outcomes Out, Earned Autonomy and The Governed Loop. Reading order, formats and buy links are on one page.
© Dr. Harish Kotadia, Ph.D., All Rights Reserved, 2026
Dr. Harish Kotadia, Ph.D., is an Enterprise AI Architect with 20+ years of IT consulting experience serving Fortune 100 clients, specializing in agentic AI systems built on Anthropic Claude, AWS Bedrock and Google Vertex AI.
Disclaimer: This blog post is based on publicly available academic publications, vendor documentation, open standards, and news items from reputed media sources linked above. This post is intended for educational purposes, to help the enterprise agentic AI community build a shared vocabulary from public, authoritative sources.
Views and opinions expressed here are my own and do not represent those of any employer or client, past or present. The analysis presented is my independent interpretation of the published sources linked above and does not constitute legal, financial, or consulting advice of any kind.