What Agents Actually Do When You Watch Them at Work
Senior operators report real speed gains from AI agents, but every account ends the same way: automation tops out short of full coverage, and a human stays on the exceptions. The gap between that reality and the market story is where the work now sits.
The velocity is real
Start with the numbers operators gave us first-hand, not vendors. A senior fintech engineer reports coding agents lifted delivery velocity by 60 to 70 percent, turning feature work that took two to three months into one or two weeks. A logistics IT director describes a subset of his team co-developing with coding agents and shipping in one month what used to take nine. An AI engineering director moved a production application from zero to deployed in one month with three developers, against an estimated six months with eight to ten the old way.
The gains reach past engineering. A compensation leader runs pay administration for 6,000 employees with a single compensation person, crediting AI-built Excel automation. A finance leader generates and emails thousands of personalized bonus letters in minutes, replacing a one to two week mail-merge. A service desk deflected roughly 45 percent of routine internal tickets like VPN and login problems. A custom root cause tool cut incident reports for customers from three days to a few hours.
These are concrete and measured. They are also narrow. Almost every win sits inside a task with a clear start, a clear end, and a human nearby.
And it tops out short
A revenue cycle director at a healthcare provider reports that so-called autonomous medical coding tops out near 90 percent. A human team must remain to work the exceptions. Across healthcare, fintech, insurance, energy, and recruiting, operators told us the same thing in different words: autonomous automation keeps humans on the last tranche of cases and on final verification.
The reasons are specific, not vague caution. Insurance pre-authorization still needs someone to challenge a denial. Denial follow-up and appeals require detective work that operators say only humans do well. One operator frames current AI as training a dog that performs tricks but cannot teach itself, insisting a human still directs every analysis.
So the headline productivity figures and the hard ceiling are both true at once. AI clears the routine middle of a process quickly. The edges, where judgment and accountability live, do not move.
Where it breaks
The failures are as instructive as the wins. Agent-generated pull requests now bundle thousands of lines across dozens of files in a single review. The reviewer cannot meaningfully inspect that volume, so the core quality gate breaks down and the chance of merged errors rises. Speed at one stage created risk at the next.
Other breakdowns came from real operational mess. An interview-scheduling AI looked strong on paper but failed because manager calendars change on the fly. An operator pulled AI out of most recruiting after pilots auto-rejected candidates they later wanted and produced sourcing lists far off the mark. Another reintroduced human support after removing it, learning that simple flows work but complex cases need a clear handoff.
A consultant's Copilot trial went wrong immediately: an employee asked what their boss made, and the tool returned executive salaries. The technology worked. The permissions and governance did not. That pattern repeats.
The real bottleneck is not the model
One operator draws the sharpest line. The next frontier, trusting agents to take autonomous decisions, is blocked by the lack of an accountability and measurement framework, not by the technology. Another rates governance as immature industry-wide because LLM decisions are not explainable the way classical models are.
So operators build containment instead. The deployments that work are tightly contained: fixed input, three defined steps, fixed output, then validated by a separate tool against pass-fail markers. One designed an autonomy ladder where humans keep a veto that cannot be switched off, even at the highest level. Another runs monitoring models that flag anomalies, alert a named owner, and trigger documented procedures.
Underneath all of it sits plumbing. Several operators named integration as their single hardest problem, from piecing data across accounting, design, and CRM tools by hand, to the unsolved security of emerging agent-to-agent and tool-connection protocols. The leap most small operators want is not a smarter model. It is systems that talk to each other.
The workforce question nobody states plainly
The cost story is more complicated than elimination. An HR operator automated half of student support calls and routine HR work, but reinvested the savings rather than cutting headcount. AI changed where the dollars went. Contract analysis and negotiation gains showed up as throughput, not layoffs. The near-term payoff is capacity, not a smaller payroll.
But the pressure is real and uneven. Operators warned that automation hits administrative and support roles hardest, and that fewer of those staff are being hired. A practitioner of fifteen years reversed her view and now believes AI is breaking the lower rungs of the career ladder, which threatens the pipeline for mid and senior talent later. Call center agents asked to teach the AI correct predictions understand they are training the system that may replace them.
One operator was candid that broad automation is being throttled on purpose because it would eliminate jobs, and framed that as an executive decision, not a project one. That is a governance gap stated out loud: workforce impact is quietly shaping how far automation goes.
What these reports give you
Read together, the interviews replace slogans with a working map. They show which tasks genuinely automate, freight invoice leakage, ticket deflection, claims under a set dollar amount, documentation drafts, and which resist it, appeals, sourcing for niche roles, strategy, and the decision of what to build at all.
They also give you the shape of deployments that hold up: contained scope, a separate validator, a named owner, a human veto that stays on. And they name the real blockers honestly, accountability frameworks, explainability, permissions, and integration between systems and agents.
If you are deciding where to spend next, the value here is not a forecast. It is a set of first-hand boundaries from people who have already shipped, failed, and adjusted. That is harder to find than enthusiasm, and more useful when the budget is yours.
The research behind this analysis. If you own this topic, each report opens in full. If not, the executive summary is free to read.