Gartner expects 40% of enterprise applications to ship with task-specific AI agents by the end of 2026, up from fewer than 5% in 2025. Gartner also projects that more than 40% of agentic AI projects will be cancelled by 2027.
Both forecasts describe the same set of organisations. Agents are arriving faster than the ability to see what they actually do.
The layer that closes that gap has a name, and in 2026 it stopped being a developer convenience and became a board-level control. It is called agent observability, and it is usually the difference between a pilot that graduates and a pilot that gets quietly shelved.
What is AI agent observability?
AI agent observability is the practice of recording and evaluating every step an agent takes, including its reasoning path, tool calls, data retrievals and final output, so a failure can be traced to one specific decision rather than guessed at. It converts an agent from a black box into an auditable business process.
The unit of measurement is the trace. A trace is the complete record of one agent run, broken into spans that each represent a single action.
A chatbot produced one answer you could read. An agent produces a chain of twelve or thirty actions, most of which the user never sees.
That chain is where enterprise risk now lives. If you cannot replay it, you cannot explain it to an auditor, a regulator, or a customer who was given the wrong answer.
Why did agent observability become an enterprise priority in 2026?
Agent observability became a priority in 2026 because agents moved into production and the tooling market consolidated around that shift. New Relic launched its Agentic Platform in February 2026, and Cisco announced its intent to acquire the evaluation and guardrails startup Galileo on 9 April 2026, folding it into Splunk Agent Observability.
When infrastructure vendors of that size buy into a category within eight weeks of each other, the category has left the research phase.
The commercial pressure behind it is equally specific. McKinsey estimates AI agents could add US$2.6 trillion to US$4.4 trillion in annual value across business use cases.
The delivery record is thinner. Industry analysis of 2026 adoption data shows that while roughly two thirds of enterprises have experimented with agents, fewer than 10% have scaled them to deliver measurable value.
New Relic's platform illustrates what enterprises are now buying: multi-agent system visualisation, waterfall views of the full request lifecycle across every agent stage, and more than 50 integrations spanning models, vector databases and orchestration frameworks.
How is agent observability different from traditional application monitoring?
Traditional monitoring answers whether a system is up, fast and error-free. Agent observability answers whether a system that is up, fast and error-free is also doing the right thing. An agent can return a clean HTTP 200 response in 900 milliseconds while quietly citing a superseded policy document.
Four differences matter to a technology budget holder.
--- Traditional monitoring measures availability and latency. Agent observability adds correctness, groundedness and policy compliance as first-class metrics.
--- Traditional monitoring treats a request as one event. Agent observability treats a request as a tree of reasoning steps, tool calls and retrievals.
--- Traditional monitoring has deterministic pass or fail states. Agent observability scores probabilistic output, which means you need evaluation criteria before you need dashboards.
--- Traditional monitoring rarely involves the business unit. Agent observability requires a subject-matter expert to define what a good answer looks like.
That last point is the one most commonly underestimated. Observability is not purely an engineering purchase, and treating it as one is how organisations end up with beautiful dashboards nobody in the business can interpret.
What does an agent observability stack actually capture?
A production-grade agent observability stack captures trace-level visibility into every reasoning step, tool call, knowledge retrieval and decision point within a conversation, then attaches quality scores, cost, latency and security findings to the same timeline. Splunk Agent Observability, built on the Galileo acquisition, is explicit about placing all of these on one view.
The practical inventory an enterprise should expect looks like this.
--- Full traces and spans for every agent run, retained long enough to satisfy your audit window.
--- Tool-call records showing which system the agent touched, with what parameters, and what came back.
--- Retrieval logs showing which documents were pulled and which were actually used in the answer.
--- Automated evaluations scoring accuracy, groundedness, tone and policy adherence on a sample or on every run.
--- Human feedback loops that turn a flagged trace into a test case for the next release.
--- Cost and token accounting per run, per workflow and per business unit.
One capability separates mature platforms from immature ones: multi-turn simulation before deployment. Comparative reviews published in 2026 note that only a small number of platforms, including Latitude and Maxim AI, support pre-deployment multi-turn simulation, which catches failures that single-turn evaluation suites cannot see.
How does agent observability support AI governance and PDPO compliance in Hong Kong?
In Hong Kong there is no dedicated AI statute, so the Personal Data (Privacy) Ordinance applies in full to any AI system touching personal data. Observability supplies the evidence a compliance check needs: what data the agent accessed, why, and whether that access was necessary for the stated purpose.
The Privacy Commissioner for Personal Data completed a third round of AI compliance checks in 2026, covering 60 organisations. According to the PCPD, 57 of the 60 organisations, roughly 95%, use AI in day-to-day operations, around 79% have used it for more than a year, and around 51% run three or more AI systems.
The round found no contravention of the PDPO, which is the good news. The direction of travel is the point that deserves board attention.
The PCPD widened the 2026 checks beyond banking, finance, insurance and education to include accounting, food and beverage, innovation and technology, logistics and property management.
The PCPD's recommendations map almost one to one onto observability capabilities: governance structures, privacy impact assessments, AI audits, staff training, incident-response plans, and prudent controls specifically for agentic AI. In March 2026 the PCPD issued a separate alert identifying agentic AI as a distinct and elevated privacy risk.
An AI audit without traces is an interview. An AI audit with traces is an inspection, and only one of those two survives a follow-up question.
What does this look like inside a Hong Kong enterprise?
Consider three scenarios that reflect how agents are actually deployed in Hong Kong organisations of 200 to 500 staff. In each, the agent works technically and the failure is invisible without trace-level evidence.
A logistics operator deploys an agent to answer shipment status enquiries across email and WhatsApp. It resolves 71% of enquiries. Traces reveal that 9% of resolved cases quoted an estimated arrival from a cached feed that had not refreshed in six hours.
A professional services firm deploys a research agent over its internal precedent library. Partners report it is useful. Retrieval logs show it draws 60% of its citations from one 2023 folder because that folder has the cleanest metadata, not the best content.
A regional bank deploys an internal policy agent for its operations team. Adoption reaches 12% in month six. Traces show the agent asks a clarifying question in 44% of conversations, which is correct behaviour for a compliance context and intolerable for a staff member between two meetings.
Each of those findings changes a decision. None of them appear on an uptime dashboard.
What goes wrong when organisations skip observability?
Organisations that skip observability do not usually suffer a dramatic incident. They suffer an evidence vacuum, which means nobody can prove the agent works, nobody can prove it does not, and the budget conversation defaults to the sceptic in the room. Gartner attributes the projected cancellation of more than 40% of agentic AI projects by 2027 to unclear ROI and weak governance.
The measurement gap is stark. Survey data reported across 2026 analyses shows companies anticipate an average 171% return on agentic AI, while only 39% attribute any EBIT impact to AI at all.
A related finding is more revealing: around 80% of AI users report improved individual productivity, but only 37% report EBIT impact. Productivity that never reaches the profit and loss statement is usually productivity that was never measured at the process level.
Governance maturity compounds the problem. Analysis of Gartner, Deloitte and IBM findings indicates only about one in five companies has a mature governance model for autonomous agents, meaning roughly 80% of organisations deploying agents lack the infrastructure to manage them safely at scale.
The four failure patterns that follow are predictable.
--- Silent degradation, where a model or prompt change lowers answer quality and nobody notices for weeks.
--- Unattributable cost, where the monthly token bill cannot be split across business units, so no unit defends it.
--- Unfalsifiable value, where the pilot cannot be proven to have worked, so renewal becomes a matter of opinion.
--- Audit exposure, where a regulator asks what data an agent accessed and the honest answer is that the logs were not retained.
How should a leader sequence the first 90 days?
A leader should sequence observability before scale, not after. The sequence that works starts with defining what a correct answer looks like, then instruments one workflow end to end, then expands. Buying a platform before defining evaluation criteria produces telemetry without judgement.
A workable 90-day sequence for a Hong Kong enterprise looks like this.
--- Days 1 to 15: pick one agent workflow with a named business owner and a measurable outcome. One workflow, not a portfolio.
--- Days 16 to 30: have the subject-matter expert write 30 to 50 reference questions with agreed correct answers. This is your evaluation set, and it is the most valuable artefact you will create.
--- Days 31 to 50: instrument traces, tool calls and retrievals. Confirm retention meets your audit and PDPO requirements before you generate volume.
--- Days 51 to 70: run automated evaluations against the reference set on every change, and route flagged traces to the business owner weekly.
--- Days 71 to 90: attach cost per run and cost per resolved case, then present one page to the finance function showing quality, cost and coverage together.
Two related decisions sit alongside this work. The first is how you supply the agent with the right information in the first place, which is the subject of context engineering. The second is how you release changes without exposing every user at once, which is the logic behind a staged rollout.
Organisations that pair observability with an embedded delivery model tend to move faster, which is part of why the forward-deployed engineer role has spread through enterprise AI programmes this year.
What is the strategic takeaway?
Agent observability is not a monitoring upgrade. It is the mechanism that converts an AI pilot into a defensible investment case, and it is the only practical way to answer the three questions every enterprise agent programme eventually faces: does it work, what does it cost, and can you prove it.
Gartner's two forecasts for 2026 and 2027 are not a contradiction. They describe a filter. Forty per cent of applications will ship with agents, and more than forty per cent of agentic projects will be cancelled, and the organisations on the right side of that filter will largely be the ones who could see what their agents were doing.
Hong Kong enterprises have a specific advantage here. The PCPD has already published what good governance looks like, which means the compliance requirement and the performance requirement point in the same direction for once.
Building that visibility takes judgement as much as tooling, and it is easier with a partner who has watched several technology cycles arrive and settle. We understand AI. We understand you. With UD by your side, AI never feels cold.
Reviewed by the UD enterprise AI team.
Ready to Put Your AI Agents on a Measurable Footing?
Now that you have the framework, the next step is choosing the one workflow worth instrumenting first. We'll walk you through every step, from AI readiness assessment and evaluation design to deployment, tracing and reporting the results upward. Twenty-eight years of serving Hong Kong enterprises, alongside you the whole way.
