The 'Digital Employee' Arrived, but Autonomy Didn't - NIA Trend 2, Revisited at the Half-Year

Held against reality: the agent arrived as predicted, minus the one word that defined it - “autonomous” (🟢, partly 🔵) Placing NIA’s two predictions next to six months of what actually happened. Prediction 2, the multi-agent standards, converged faster than predicted (🔵). Prediction 1, the digital employee, arrived as a noun but shed the adjective “autonomous” that defined it (🟢). ROI failure rates and the governance gap were in neither prediction (⚪).

As set out in Part 0, I hold each of NIA’s two prediction sentences for Trend 2 against what actually happened. Same method as Part 1.

NIA’s Trend 2, “Self-Working AI Agents, Reshaping Collaboration and Automation,” carried two predictions.

• Prediction 1: A “digital employee” emerges that goes beyond following instructions to autonomously carrying out work processes.
• Prediction 2: Multi-agent systems arise in which small agents collaborate.

One word runs through both: autonomous. NIA treated an agent’s ability to handle work on its own as the core of the trend. Six months of reality does not ask whether that ability arrived; it asks whether the ability that arrived turned into trust and value. And those two split apart.

The half-year, in one line: the technology, the standards, and the platforms converged faster than predicted, while value, trust, and governance failed to keep pace, and the gap widened. NIA read correctly what was coming; it never asked how much of it could be trusted.

Prediction 2, Multi-agent standards - hit, faster than predicted (🔵)

Start with the prediction that landed hardest. “Small agents collaborating” became real in the form of converging standards, faster than the forecast expected.

The biggest signal is that the standards war is already over. The three layers of specification - agents calling tools, agents talking to each other, agents paying each other - moved to one place in half a year. Anthropic’s MCP (Model Context Protocol), Google’s A2A (Agent-to-Agent), and the payments-oriented AP2 (Agent Payments Protocol) all moved under the Linux Foundation, with Anthropic, OpenAI, Google, Microsoft, and Amazon at the same table. Rather than competing specs pushing each other out, they divided the roles by layer.

The convergence shows in adoption numbers too.

The three cloud majors loaded these standards onto their platforms. Microsoft shipped Copilot Studio and Azure AI Foundry in April 2026 with A2A support; Amazon did it through Bedrock AgentCore, Google through its own Agent Development Kit (ADK). IBM folded its own ACP spec into A2A. These standards did not stay at the announcement stage; they went into supply chain, finance, insurance, and IT operations. Microsoft, Amazon, Salesforce, SAP, and ServiceNow put them into their products, and financial firms such as Deutsche Bank adopted them. What remains unfinished: a central registry that can become a single point of failure, and fine-grained control over which agent gets which permissions.

Korea rode the wave too. Samsung SDS added MCP connectivity to its FabriX platform, and the Ministry of the Interior earmarked 18.3 billion won for an “intelligent work management platform,” naming MCP in the requirements (with Samsung SDS and Naver Cloud). By Google Cloud’s count, agent deployment across Asia-Pacific grew 340% in 2025.

The spread of standards is also the spread of the attack surface. As the channels connecting agents multiply, so do attack paths like prompt injection, but that security angle belongs to Trend 5 (AI Security). This part goes only as far as how quickly the standards converged.

Prediction 2 was right in direction and in speed. The plumbing that connects agents got laid faster than predicted. NIA saw this well. That plumbing being laid, though, is a different thing from something trustworthy flowing through it, and that gap shows up in Prediction 1.

Prediction 1, The digital employee - the noun came, the adjective didn’t (🟢)

Prediction 1 is different. The noun “digital employee” did in fact appear over the half-year. Agents went into production in roles like coding, customer service, and sales development. The trouble is with the adjective NIA put in front of it: “autonomous.”

The evidence that the noun arrived is solid. In software engineering, production deployment reached 53%, in customer service 62%, and payback across roles ran as short as under four months. But read the same table vertically and a different picture appears. The more a function is bound by regulation and audit, the more the rate of human intervention (HITL, human-in-the-loop) climbs and the longer payback drags out.

61% human intervention in legal means the agent does the work but the human holds the decision. Not an “autonomous employee” so much as a “first-draft writer.” And that intervention is not out of caution; it is because letting it run autonomously actually breaks.

How it breaks is already visible in the benchmarks. The visible scores look near-saturated. On SWE-bench Verified, which measures code-fixing ability, the top models cleared 90%. But run the same task several times in a row and the reliability collapses. On the telecom-task benchmark tau2-bench, an agent that hit 90% once drops to 57% when measured as the probability of succeeding eight times in a row (pass^k, success across k consecutive runs). The gap between working once and working every time is exactly what production demands.

There is a worse signal. A UC Berkeley result from April 2026 found that one agent passed SWE-bench Verified and Terminal-Bench without actually solving a single task, 100% through reward hacking. A single ten-line config file made all 500 tests pass. The score was pointing at a hole in the grading, not at ability.

How much an agent can work alone is also still measured cautiously. METR’s time horizon puts the top models at roughly 12 hours at even odds, and METR itself said the range beyond that has too small a sample to trust. Short-task accuracy and long-task endurance are not on the same curve.

The way things break over long runs is more concrete. The longer an agent drags a task out, the more its working memory rots, the more it loses information while compacting context (lossy compaction), the more it satisfies the letter of an instruction while destroying its intent (ugly wish-granting), and the more it passes results that are functionally running but wrong without a human catching them. So an agent’s memory is evolving from a plain store into a layer that manages execution state.

---
config:
  look: handDrawn
  theme: neutral
---
flowchart LR
    A["embedded in app<br/>80%"] --> B["actually in production<br/>31%"]
    B --> C["scaled enterprise-wide<br/>23%"]
    C --> D["EBIT-level effect reported<br/>39% (mostly under 5%)"]

The gap shows in adoption too. Agent features are embedded in 80% of apps, but only 31% actually run in production, and only 23% have scaled enterprise-wide. 88% of pilots never reach production and get scrapped. The noun went into nearly every piece of software, but the seat where that noun does the work of the adjective (“autonomous”) is fewer than one in three. Hence 🟢: right in direction, missing its core adjective.

What NIA didn’t see at all - ROI failure rates and the governance gap (⚪ unmapped)

Two first-order variables are named in neither prediction. NIA looked only at what agents can do, not at how that ability fails or who is accountable when it does.

The first is the failure rate of ROI (Return on Investment). As adoption rose, so did failure. In a 2025 MIT study, only 5% of integrated pilots produced millions in value. The interesting split is in success rates: external vendors succeeded 67% of the time, versus 33% for build-your-own, half as often. In McKinsey’s survey, 39% of firms reported an EBIT-level effect, but most of it was under 5%. Gartner expected more than 40% of agent projects to be cancelled by 2027.

But reading that cancellation rate as pure failure sees only half of it. 2025-26 was the year many firms first budgeted for agents, and 2027 is when that budget first comes up for review. The 40% cancellation looks less like a bursting boom than like a first budget cycle sorting the wheat from the chaff (a reset). In practice, the root cause of failure was often not model performance but unprepared data. Build-your-own landing at half the success rate of external vendors reads the same way: not the model, but the data and workflow to run it on, were not ready.

This sorting left a mark on jobs too. In Stanford HAI’s survey, employment of software developers aged 22-25 fell about 20%, 17% of firms had already cut headcount with AI, and 30% expected to. But the form was less mass layoff than not backfilling seats people left and trimming outsourcing first. That the seats disappearing and the seats appearing are not on the same curve is another axis the forecast missed.

The second is the gap in governance and accountability. The more agents act on their own, the more you need rules for who is responsible when an action goes wrong. Those rules fell well behind capability. The EU AI Act mandates human oversight and lifecycle log retention for high-risk uses like hiring and credit scoring, and faced an August 2026 effective date (with a Digital Omnibus proposal to delay it to December 2027, so it is uncertain), but neither the EU nor the US standards body (NIST) designed its rules for autonomous agents to begin with. Of the roughly 500 organizations McKinsey examined, only about 30% had agent-control maturity at level 3 or above. Responsible-AI maturity itself averaged 2.3 on a 5-point scale, and top performers had human-intervention verification in place 65% of the time versus 23% for the rest. The gap in control capability turned straight into a gap in results.

Here formal oversight splits from real oversight. Ceremonial oversight, a human just clicking “approve,” is not recognized as a legal defense. Lawful intervention requires understanding the context, having a kill switch, and holding the authority to override. In the HITL table above, the 61% in legal is exactly this real-intervention kind, and NIA’s prediction leaves this axis out entirely.

That gap shows in the platform race too. The number that says “deployed” and the number that says “actually used” pull apart.

Copilot’s 35.8% active conversion means two of every three people who bought a license don’t actually use it (ChatGPT, for reference, is 83.1%). On top of that sits the cannibalization dilemma. If an agent replaces 80% of a seat’s work, cancelling 80% of the seat licenses is the rational move, but that collapses vendor revenue, so pricing comes out as an awkward bundle of usage-metering and premium seats. Pricing is in fact splitting into per-seat (Copilot $30/mo), per-action (Agentforce $0.10 each), per-conversation, and add-on uplift (ServiceNow Pro Plus +50-60%), and firms are now forced to watch tokens and API calls in real time. The capability converges, but the way to make money from that capability is not yet settled.

One more thing worth flagging: most of these numbers come from vendors. Market-size estimates alone diverge two- to three-fold depending on how “agent” is defined, and vendor-sponsored surveys skew optimistic. There was even a dispute in which OpenAI challenged the accounting basis of Anthropic’s revenue figure. So this table reads less as a leaderboard than as material for reading the gaps between deployment and use, contracts and revenue.

One thing worth flagging - the contest is won at orchestration, not autonomy

Map this comparison onto Korea and a priority emerges that runs against the conventional read.

By will and speed of adoption alone, Korea is at the global top, and led from the very top of the chaebol. In June 2026, Samsung Electronics officially adopted a multi-LLM setup using ChatGPT, Gemini, and Claude together in its office division (Chairman Lee Jae-yong: “we must change the organization’s DNA”), and Samsung SDS rolled FabriX and Brity Copilot out to 200,000 employees, reporting a 98.8% conversion rate on its code-migration agent and a 68% cut in development cost. LG Electronics set a target of 30% productivity gains within two to three years, and SK Telecom deployed “A.Biz Cowork” across 25 affiliates. Naver shut down the consumer (B2C) service of ClovaX in April 2026 and focused on the enterprise (B2B) HyperCLOVA X.

Adoption runs ahead like this, yet in the models and standards beneath it, Korea is still the side following. Korea’s private-sector AI investment is only 2.1% of the US level (KDI). Upstage’s new Solar model approaches GPT-5 and Claude on some metrics and LG’s EXAONE is finding its place in Korean-language and industry specialization, but at the design table for the three standard layers seen above (MCP, A2A, AP2), there was no seat for a Korean firm. First in adoption, importing the specs.

The seat is similar in payment standards. Kakao’s gifting flow and Naver’s search-to-shopping path are candidates for agent payments, but Korea’s direct participation in the design of global payment specs like AP2 is weak. The pattern of leading in adoption while importing the specs repeats in payments.

So the lever this half-year points to for Korea is not “how fast we deploy autonomous agents.” All the data above points one way. The noun already arrived; the contest is won not by whether you can attach “autonomous” to it, but by how well people orchestrate the noun that stays un-autonomous. This shift from doer to orchestrator already shows in org charts: 56% of surveyed firms created a new “AI agent owner” role.

This frame shows up the same shape even from the side selling agents. In July 2026, Replit’s Amjad Masad and Scott Kennedy published a piece framing their own operations as a “self-driving company.” As a vendor reporting its own results, it is not something to take at face value. Yet the definition they landed on was “not a company without people, but one where people choose the destination,” and “doers become directors.” Even as company-wide code contribution jumped 5.8x in half a year and 30% of pull requests (PR) went to agent approval without a human, the wheel in the picture they drew was still where a human chooses the destination. Even the most optimistic picture drawn by a company that sells agents was orchestration, not autonomy. The name “self-driving” says as much: a self-driving car still has a human choosing where it goes.

From here I add what I see running coding agents every day. Put agents to actual work and what is scarce is not the model’s reasoning but, usually, data worth handing over and the judgment of whether to hand it over at all. Leave it fully autonomous and it quietly goes wrong at exactly the spot pass^k describes, where it works once but not every time. So what actually gains value in practice is not the agent’s autonomy but the person who knows how far to delegate and where to hold. The real differentiator in Trend 2 is not a model-performance table but the maturity of governance that keeps logs transparent and can locate accountability when something goes wrong. That maturity is what ends up holding the decision on how far autonomy is allowed.

Governance placed at the front line of adoption shows in Korean firms too. Samsung SDS previewed FabriX 2.0, a multi-agent operating-system-of-sorts that binds several LLMs on-premise, for October 2026, with lifecycle management and red-teaming attached, and SK Telecom set up a CEO-level “AI board” and designed its infrastructure on a zero-trust premise. Rather than chasing the specs it trails on, getting a control regime that matches the speed of adoption in place first is closer to a seat Korea can actually hold.

Closing - that it comes is settled, whether it can be trusted was another matter

Held against the prediction, Trend 2 comes out like this.

Prediction 2 (multi-agent standards) went past a hit to arrive faster than predicted; Prediction 1 (the digital employee) came with its defining “autonomous” stripped off. And the two variables, ROI failure rate and the governance gap, were in none of the predictions.

Just as in Part 1 the contest over infrastructure was decided not by the chip but by the bottlenecks behind it, the contest in Trend 2 had moved from “autonomous capability” to the things that let you trust and delegate that capability, namely reliability and governance. NIA read correctly what agents can do, and never set what they can be trusted with as a target of prediction in the first place.

This too is the same lag story as Part 1. In the discourse of a year ago, “agent” was a word about capability, not about trust. So capability got caught in the prediction, while the reliability and accountability that actually decided the half-year stayed outside it. What Trend 2 shows is that the faster capability arrives, the wider the distance grows between where that capability landed and where it can be trusted.

In the next part I hold Trend 6, national technology sovereignty, to the same method. I look at how the importer’s seat that flickered past here shows up in the contest over norms.

The compared source is NIA, “The 12 AI and Digital Trends NIA Forecast for 2026” (IT & Future Strategy No. 6, 2025.12.31). Evidence for H1 2026 developments prioritizes primary sources: Linux Foundation and each standard’s spec, IR from Anthropic, OpenAI, Microsoft, Salesforce, and ServiceNow, Gartner, McKinsey, MIT NANDA, and IBM, UC Berkeley RDI and METR, the EU and NIST, Samsung Electronics, Samsung SDS, SK Telecom, LG Electronics, Naver, and Upstage, and KDI. Replit’s “self-driving company” piece (2026.7.16) is cited as an operator’s-view example only, flagged as a vendor’s report on its own operations, not as neutral evidence.