Most failed programmes do not fail on technology; they fail on scope, value and control.

Gartner names escalating costs, unclear business value and inadequate risk controls as the reasons projects get cancelled. In our delivery work at Cybiant, five patterns sit behind those three causes.

The roadmap that follows is built to remove each of these failure points in sequence.

Run 2027 as four phases, each with a gate that must be passed on evidence before money moves to the next.

The gates are how you stay out of Gartner’s 40% cancellation figure: weak use cases stop early and cheaply, strong ones get funded to scale.

Q1 is where to start, and it is mostly decisions rather than code. Governance, cost control and skills run underneath all four phases.

Start where volume is high, inputs and outputs can be bounded, and the outcome is measured in money or time.

Those are the conditions under which agentic projects survive the move from pilot to production.

Score every candidate process from 1 to 5 on four tests, then fund the top three or four.

Then match the tool to the job, following Gartner’s rule: an agent where a judgement or decision is needed, automation for routine rule-based steps, and an assistant for simple look-up. Using an agent where a bot will do adds cost and risk for no gain.

Put an operating model in place before your first agent goes live: who owns it, what it may decide, and who can stop it.

Gartner’s 2026 Hype Cycle shows governance, security and FinOps for agentic AI emerging early in the adoption cycle, not after scale.

If you can answer yes to at least seven of these, you are ready to start 2027 at pace.

Each “no” is a first-quarter action.

Cybiant is the world’s leading partner that has authored the global standard for agentic service automation, and can design, build, run and teach it, end to end.

Every engagement uses the Service Automation Framework we wrote, so what we design keeps working long after hand-over.

The frameworks differ in legal force, but they converge on the same four asks: know your agents, limit what they can do, record what they did, and keep a person able to intervene. Design once against that common core and you can map to each framework rather than rebuilding for each.

Most internal service agents will not be classed as high-risk under the EU AI Act. Articles 12 and 14 are still the most precise public description of what “governed” looks like, which makes them a sensible design target even where they are not mandatory.

The strongest control is the one that does not depend on the agent behaving well. Prompts and instructions shape behaviour; permissions, limits and stop mechanisms enforce it. Singapore’s framework puts it as bounding risk by design: restrict access, limit autonomy, keep every action traceable to a human, and threat-model before go-live (Bird & Bird).

The stop mechanism is the one most often missing. Article 14(4)(e) of the EU AI Act asks for an interrupt that brings the system to a halt in a safe state (AI Act Service Desk). For an agent halfway through a multi-system update, that means designing rollback or compensation steps, not just a pause button.

An audit trail for an agent must answer one question for any action: what did it read, what did it change, in which system, on whose behalf, and who could have stopped it. Model-level logs of prompts and completions are not enough on their own; regulators and auditors will look for a per-action record (StackOne). The EU AI Act requires logging to be built into high-risk systems for their whole lifetime, not added later (Article 12).

Use a standard format. The OpenTelemetry semantic conventions for GenAI agents define spans such as invoke_agent and execute_tool, so one agent run becomes one trace with each tool call as a child step. The conventions are still marked as in development (Microsoft Learn), but adopting them now keeps traces portable across ServiceNow, UiPath, Power Platform and custom agents.

Make logs trustworthy. Write to append-only storage with integrity checks, separate log access from agent access so an agent cannot edit its own history, and set retention by the strictest rule that applies. For high-risk systems under the EU AI Act that is at least six months for automatically generated logs (Prediction Guard).

Protect the people in the logs. Agent traces often contain personal data. Mask or tokenise sensitive fields, and align retention with PDPA, GDPR or sector rules.

A human approval on every step does not scale and quickly turns into rubber-stamping. The better question is where a person adds real protection. Two factors decide it: how much harm a wrong action causes, and how hard it is to undo. Singapore’s framework names the same factors, asking organisations to weigh the scope and reversibility of an agent’s actions before setting its autonomy (Bird & Bird).

Checkpoint placement matrix: impact by reversibility. Move an agent between quadrants as the logs earn it more autonomy.

Only actions that are both high impact and hard to undo need approval before the agent acts; everything else can run with review after the fact, a cancel window, or logging and sampling.

Autonomy should be earned. Start an agent in a stricter quadrant, then move it as logs show reliable behaviour, and move it back if incidents or drift appear.

Governance works when it is part of how automated services are designed, not a review bolted on before go-live. The Service Automation Framework, authored by Cybiant and maintained by the Service Automation Framework Alliance, already treats automation as a service with users, a designed interface and managed delivery (Cybiant). Agents fit that structure: the autonomy tier becomes a design decision in the service blueprint, and exception handling becomes part of service delivery rather than an afterthought.

The confusion between these two categories causes two very different, and equally costly, mistakes.

The organisations getting the most value from automation in 2026 aren’t choosing one approach over the other. They’re using both, deliberately, based on the actual shape of the process:

• Rules-based, high-volume, low-ambiguity tasks stay on RPA — faster to build, cheaper to run, easier to govern.
• Judgment-involving, cross-system, or exception-heavy processes move to agentic AI, with appropriate governance built around them.
• In many mature deployments, agents sit above a layer of existing RPA bots, orchestrating them and handling the exceptions the bots themselves can’t resolve.

The processes seeing genuine agentic deployment right now share a profile: high volume, a clear goal, a bounded set of systems to work in, and enough historical cases to define policy for the exceptions. Common examples include:

This is exactly the gap the Service Automation Framework (SAF) was written to close. As the framework’s author and first accredited partner, Cybiant approaches agentic AI as an extension of a disciplined automation practice, helping organisations work out which processes belong on RPA, which belong on agentic AI, and how to govern the transition between them.

Strip away the marketing and most agentic systems share the same five parts. Understanding them is the fastest way to evaluate any vendor’s “agent” claim.

• Perception. How the agent takes in information: reading a document, watching a queue, receiving an API event, or parsing a screen.
• Reasoning and planning. The layer, usually a large language model, that interprets the goal and decides what sequence of steps might achieve it.
• Memory and context. What the agent retains between steps and between runs: the conversation so far, retrieved records, past outcomes. Without this, an agent cannot handle a multi-step or long-running process.
• Action and tools. The concrete capability to do something: call an API, fill a form, update a record, send a message. This is where agentic automation overlaps directly with RPA, since bots are often the hands an agent uses.
• Orchestration and governance. The control plane that decides which agent handles what, enforces permissions, logs every decision, and routes exceptions to a human. This is the layer most narrow “AI agent” demos skip, and the layer that determines whether a pilot survives contact with production.

The three terms get used almost interchangeably in vendor marketing, and they should not be.

The practical implication: agentic automation does not replace RPA, it sits above it. Robots remain an efficient, auditable way to take action in legacy applications that do not have clean APIs. What changes is who decides which robot runs when, and that decision moves from a fixed script to a reasoning agent operating inside a governed set of boundaries.

Two things converged. Large language models became reliable enough at multi-step reasoning and tool use to plan a process rather than just describe one, and enterprises already had a decade of RPA, orchestration, and process mining investment providing the execution layer for agents to plug into.

Analyst data shows the pace of interest clearly. According to Gartner’s 2026 CIO and Technology Executive Survey, only around 17 percent of organizations have deployed AI agents so far, yet more than 60 percent expect to do so within two years, among the steepest adoption curves Gartner tracks for any emerging technology. Separately, Gartner has forecast that 40 percent of enterprise applications will carry task-specific AI agents by the end of 2026, up from under 5 percent in 2025.

Most analyst models describe agentic maturity as a progression rather than a switch. A commonly cited version, reflected in Gartner’s own research, moves through roughly five stages: embedded assistants that answer questions inside an application, task-specific agents that complete one bounded job end to end, collaborative agents that coordinate with each other across a process, and eventually autonomous agent ecosystems that manage entire functions with minimal human direction, expected toward 2028 and 2029 rather than today.

Almost every enterprise currently sits in the first two stages. The gap between “we have a working pilot” and “we have an autonomous ecosystem” is not a model upgrade, it is years of governance, integration, and organizational change, which is exactly why a maturity assessment matters more than a product demo when deciding what to build next.

The processes seeing genuine agentic deployment right now share a profile: high volume, a clear goal, a bounded set of systems to work in, and enough historical cases to define policy for the exceptions. Common examples include:

What is common across all four is that none of them are wide open. Each has a defined goal, a bounded action set, and a clear escalation path. That boundary is what makes agentic automation safe to deploy, and it is also the first thing many first attempts get wrong by scoping too broadly.

The same flexibility that makes agentic automation powerful is what makes it risky without the right controls. Analyst forecasts are specific about where failures come from.

In practice, the failure pattern looks less like “the AI made a mistake” and more like “nobody had defined what the agent was and was not allowed to do before it was given access to a production system.” Scope creep, missing audit trails, and unclear escalation ownership cause far more damage than any single wrong answer from a model.

Treating agentic automation as a governed capability rather than a feature means answering a short list of questions before any agent touches a production system.

This is precisely the layer that separates a working pilot from a program that survives its second year. It is also where most off-the-shelf platform features stop and where an explicit framework, like the Service Automation Framework Cybiant authored, has to take over.

Agent – A software component that interprets a goal, plans steps, and takes action through tools, adjusting based on results, rather than following a fixed script.

Orchestration layer – The control plane that assigns work to agents and bots, enforces permissions, and coordinates multi-agent or multi-step processes.

Tool calling – The mechanism by which an agent invokes an external system, API, or robot to carry out a step of its plan.

Human in the loop – A governance pattern where a person reviews or approves an agent’s decision before it takes effect, typically used for higher-risk actions.

Guardrails – Explicit rules, permissions, or policy checks that constrain what an agent is allowed to do, independent of what it decides is the best path.

Agent memory – The information an agent retains across steps or sessions, such as case history or prior decisions, needed to handle multi-step or long-running work.

• Gartner, Hype Cycle for Agentic AI, 2026

• Gartner, 2026 CIO and Technology Executive Survey, as referenced in Gartner’s 2026 Hype Cycle for Agentic AI commentary

• Gartner, Predicts 2026: The New Era of Agentic Automation Begins, published December 2025

• Gartner press release, Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, August 2025

Share of agentic AI projects at risk vs. share of organizations governance-mature

Source: Gartner Predicts 2026 research on agentic AI project risk and governance maturity

Because agentic systems act rather than only answer, they introduce a different category of risk than earlier generative AI tools. Any organization deploying agents needs strong permission-based access controls as agents connect into more enterprise systems, clear accountability for who is responsible when an autonomous agent errs, and ongoing monitoring treated as a permanent operating cost rather than a one-time project task.

Two frameworks anchor most formal governance programs today. The NIST AI Risk Management Framework is a voluntary US methodology built around four functions, Govern, Map, Measure, and Manage. ISO/IEC 42001, published in 2023, is the first certifiable international standard for an AI management system, structurally similar to ISO 27001 for information security, and it is increasingly used as a procurement signal between enterprises and their vendors. Neither framework, nor the EU AI Act, was originally built with autonomous, tool-using agents in mind, so organizations need to actively extend them to cover cascading failures across connected agents, scope creep beyond an agent’s intended task, and gaps in accountability.

Share of enterprise application software revenue driven by agentic AI

Source: Gartner, enterprise application software revenue forecast, best-case scenario to 2035

Most enterprise automation history has followed a predictable arc: organizations first ask whether a task can be automated, then whether it can be scaled, and more recently whether an intelligent, connected ecosystem can be built around it. At Cybiant, we believe agentic AI pushes that progression one step further, toward automating decision-making itself, not only the execution of tasks. We call the resulting model the Self-Driving Organization™: an enterprise where core operational, service, and decision-making workflows are orchestrated by interconnected AI agents rather than managed through manual intervention at every step.

This is deliberately a workforce elevation strategy, not a workforce replacement strategy. When agents absorb operational burden, human talent is freed to focus on judgment, creativity, relationships, and strategic thinking, the things machines cannot replicate. Our implementation experience across hundreds of organizations on every continent points to five dimensions that must be addressed together.

Most organizations we work with have strong coverage in one or two of these dimensions and significant gaps in the rest, which is exactly why an independently authored maturity standard, rather than a single vendor’s implementation guide, matters. That standard is the Service Automation Framework (SAF), which Cybiant authored, and Cybiant was the first organization globally to become accredited in it. The SAF defines maturity levels, operating model requirements, governance structures, and capability benchmarks for how service automation programs are designed, operated, governed, and scaled, so every deployment Cybiant builds is proven, measurable, and audit-ready rather than improvised.

Five actions consistently separate organizations that capture real value from agentic AI from those funding expensive experiments.

This article reflects publicly available research current as of July 2026. Figures should be verified against original sources before use in board or investment materials.

• Gartner, Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, gartner.com
• Gartner, Gartner Says $234 Billion in Enterprise Application Software Spend Is at Risk from Agentic AI, July 2026, gartner.com
• Gartner, 2026 Hype Cycle for Agentic AI, gartner.com
• Gartner, Predicts 2026: The New Era of Agentic Automation Begins, gartner.com
• Gartner, 2026 CIO and Technology Executive Survey, as reported in Gartner’s Hype Cycle for Agentic AI
• MIT Sloan, Agentic AI, Explained, mitsloan.mit.edu, February 2026
• Futuriom, Agentic Infrastructure Operations: What It Is and How to Do It Safely, futuriom.com, April 2026
• McKinsey & Company, The State of AI in 2025: Agents, Innovation, and Transformation, mckinsey.com
• McKinsey & Company, State of AI Trust in 2026: Shifting to the Agentic Era, mckinsey.com, March 2026
• S&P Global Market Intelligence, enterprise agent production-deployment tracking, 2026
• TrustCloud, ISO 42001 & NIST AI RMF: Mastering Responsible AI Governance in 2026, trustcloud.ai
• GAICC, Global AI Governance Comparison 2026: EU AI Act vs NIST AI RMF vs ISO/IEC 42001, gaicc.org
• ModelOp, NIST vs ISO: Compare AI Frameworks, modelop.com
• Cybiant, Introducing the Self-Driving Organization, cybiant.com
• Cybiant, Service Automation, cybiant.com

The evolution of enterprise automation has moved through distinct phases, starting from basic scripts and macros, moving to RPA mimicking human UI interactions, and then to Intelligent Automation layering in machine learning. Each wave brought efficiency gains. But each also revealed the same fundamental limitation: systems bound to the paths their designers could predict.

Modern enterprises operate in environments that defeat static logic. Supply chains shift overnight. Regulations evolve. Customer expectations change. In this context, the inability of traditional automation to handle variability becomes a strategic liability.

Agentic Automation addresses this gap by enabling systems to reason, plan, and adapt within defined boundaries. Rather than encoding every possible path, enterprises define goals, policies, and constraints, and agents determine how to achieve objectives dynamically. Autonomy moves from innovation to operational necessity.

A common misconception is equating agentic systems with conversational AI or chatbots. The whitepaper is clear: enterprise agents are not chat interfaces. They are embedded within operational environments, invoking APIs, executing automations, validating outcomes, and operating under strict governance.

Agentic systems operate across a spectrum of autonomy, from agents that only recommend actions to fully autonomous systems operating under strict policy constraints. The whitepaper defines three levels:

Most enterprises should start with advisory or semi-autonomous models. As governance matures and trust is established, the scope of autonomy can be responsibly extended.

Most enterprises should start with advisory or semi-autonomous models. As governance matures and trust is established, the scope of autonomy can be responsibly extended.

The whitepaper is emphatic: governance is not an enhancement; it is the condition that makes agentic automation viable at enterprise scale. In traditional automation, failures are localized and predictable. In agentic systems, failures can propagate across systems, processes, and organizational boundaries.

Effective governance in agentic automation rests on three interdependent elements:

Data security and privacy are equally critical. Agents must be treated as first-class identities within the enterprise security model, operating under least-privilege principles with strict data handling and regulatory compliance enforced throughout the agent lifecycle.

One of the most practical contributions of the whitepaper is a five-level maturity model that helps enterprises understand where they are and how to advance responsibly, from manual scripted operations through to fully managed autonomous operations:

Most successful enterprise implementations currently operate at Level 4: Semi-Autonomous Agentic Systems, balancing autonomy with human supervision and exception handling. The key insight: use the maturity model as a planning tool, not a scorecard. Advance incrementally, selecting use cases aligned with current readiness and investing in governance before scaling autonomy.

The whitepaper draws on real-world enterprise implementations to identify patterns that consistently determine success or failure. Common pitfalls to avoid:

What consistently works: start small, embed agents into existing processes, maintain clear human oversight, measure outcomes not just efficiency, and invest in governance from day one.

In the long term, competitive advantage will not come from who deploys agentic systems first. It will come from who governs them best. Organizations with mature governance frameworks move faster, not slower. Clear policies, controls, and accountability mechanisms reduce internal friction, shorten approval cycles, and accelerate deployment.

Human roles will evolve alongside these systems, shifting toward defining goals and intent, designing policies and constraints, supervising agent behavior, and making strategic and ethical judgments. Enterprises that invest now in architecture, governance, and organizational readiness will be better positioned to leverage increasing autonomy without sacrificing control or trust.

The complete 36-page guide includes a one-page Enterprise Readiness Assessment Checklist, detailed implementation roadmap, and all reference architecture diagrams.

The solution lies in separating those responsibilities. Rather than asking one agent to balance two competing objectives, a Double-Agent Framework assigns a dedicated agent to each task.

The Primary Agent is configured to be aggressive and high-recall — catching every suspicious transaction the benchmark model missed. The Secondary Agent then takes on the filtering role, reviewing every alert and evaluating whether each one genuinely warrants investigation.

This architecture mirrors how high-performing human investigation teams actually operate. A first analyst surfaces potential cases; a senior reviewer applies judgement before escalation. The difference is scale — an AI-powered double-agent system can operate at the speed and volume of digital transaction flows that no human team can match alone.

According to figures cited in Cybiant’s research, money laundering in Malaysia alone accounted for an estimated 5.04% of GDP in 2024 — over USD 20 billion in losses. At the same time, the annual cost of compliance has reached nearly USD 2 billion. AI-driven AML approaches have the potential to reduce those compliance costs by close to 50%.

But cost efficiency without detection accuracy is a false economy. The Double-Agent Framework is designed to deliver both: lower operational burden and higher detection performance, simultaneously. And when the false positive rate is reduced upstream by the Secondary Agent, the alerts that reach investigators are more likely to be real — changing the nature of the work from managing volume to applying expertise.

The Double-Agent Framework is the natural next step in Cybiant’s AML research programme. The foundations are already proven — the Primary Agent architecture works and the performance benchmarks are established. The next phase involves developing and testing the Secondary Agent: evaluating how effectively it reduces false positive rates without sacrificing the high recall the Primary Agent delivers.

AML is not a problem that gets solved and stays solved. The Double-Agent Framework isn’t just a research concept — it’s a practical architecture that financial institutions can begin designing for today.

The complete paper covers our full methodology, statistical analysis, model architectures, and extended findings — including a detailed breakdown of False Positive trade-offs across all four approaches.

Sustainable IT — sometimes called Green IT — is no longer a nice-to-have. Regulatory pressure, investor scrutiny, and stakeholder expectations have made it a strategic priority for organizations of all sizes and across all sectors.

The challenge is that most organizations approach sustainability in IT in a fragmented way: swapping hardware for energy-efficient models here, shifting workloads to cloud providers there, reporting carbon metrics when required. What is missing is a coherent, end-to-end strategy that embeds sustainability into every IT decision.

The Green IT Framework, developed by DASCIN (The Data Science Institute), is the only comprehensive, vendor-neutral methodology available today that guides organizations in implementing sustainable and energy-efficient IT operations. Accredited by APMG International, the Green IT Framework provides a structured learning and implementation path from foundational awareness through to advanced practice.

The framework focuses on eight key components:

Crucially, the Green IT Framework is technology-neutral. It does not prescribe specific vendors or tools, making it directly compatible with any AI or automation platform an organization is already using.

The most powerful insight is that AI Automation does not just need to be managed within the Green IT Framework; it can actively accelerate an organization’s Green IT objectives.

For organizations looking to bring AI Automation and Sustainable IT together, we recommend a phased approach aligned to the Green IT Framework maturity journey:

The organizations that will win in the data economy are not just those that automate the most — they are those that automate most responsibly. AI Automation, when deployed within the structured guidance of the DASCIN Green IT Framework, creates a compounding advantage: operational efficiency, cost reduction, ESG compliance, and a demonstrable commitment to a sustainable future.

The organizations that will win in the data economy are not just those that automate the most; they are those that automate most responsibly. AI Automation, when deployed within the structured guidance of the DASCIN Green IT Framework, creates a compounding advantage: operational efficiency, cost reduction, ESG compliance, and a demonstrable commitment to a sustainable future.

The research utilized a comprehensive dataset containing over 9.5 million transactions. Four distinct processes were developed and tested:

The AI agents were built using the UiPath Cloud interface and leveraged Large Language Models (GPT-4o) to perform contextual analysis and prioritize red flags.

The integration of AI agents delivered statistically significant improvements in the detection of money laundering activities compared to standalone models.

The complete paper covers our full methodology, statistical analysis, model architectures, and extended findings — including a detailed breakdown of False Positive trade-offs across all four approaches.