Why Agentic AI Projects Fail, and How to Build One That Doesn't

How to Build AI Projects That Succeed

Agentic AI projects fail for three reasons that have nothing to do with whether the technology works: escalating cost, unclear business value, and inadequate risk controls. The fix is unglamorous. Track token and compute cost in real time, tie every agent to a measurable business KPI defined upfront, and stand up a governance framework before the first project starts. Robots are not obsolete. The winning model uses agents where judgment and ambiguity live, deterministic automation everywhere accuracy and cost discipline matter, and orchestration to coordinate both alongside people.

Agentic automation is the topic of the day, every day, in every feed. The useful conversation is not whether agents are powerful. It is why so many agentic projects stall before production, and what the ones that scale do differently.

Don Sweeney, Co-founder and Co-CEO of Ashling, sat down with Brad Hairston, from the strategy team at SS&C Blue Prism, to work through three things most teams treat separately and shouldn't: the operating model, the economics, and change management. Ashling and SS&C Blue Prism have collaborated for years. Here is the substance.

What Makes Agentic Automation Different

Most automation already running in enterprises is deterministic. It is predictable, rule-based, and structured, it runs in a data center or the cloud, and it needs little human supervision. That kind of automation is not going anywhere, and pretending it should is the first mistake.

Agents are a different animal. They perceive their environment and hold business context through a workflow, they reason and plan, they act, and they learn and adapt based on those actions. Crucially, you point them at a goal or an outcome rather than a fixed set of rules. Think junior analyst, not script or macro.

Agentic automation earns its place in four situations:

• Adaptability. The work requires interpreting context, handling ambiguity, and adapting to scenarios you did not script.
• Decision-making. The process needs an agent to reason through a problem and choose.
• Unstructured data. Emails, documents, images, and conversations carry the meaning, which is exactly where traditional automation falls short.
• End-to-end ownership. The agent orchestrates a whole process, calling multiple systems, handling exceptions, and adapting as it goes.

The catch that comes with all of that: agents bring far heavier governance and monitoring requirements, and that is the part that has slowed adoption. Companies want a solid foundation before they take on the new risks agents introduce, which is the right instinct.

One framing worth keeping: an agent is not set-and-forget. It needs onboarding, mentoring, and ongoing care the way a junior analyst does. Treating it like a switch you flip is how projects go wrong early.

The Convergence Everyone Is Now Naming

For years, enterprises ran separate centers of excellence by tool: one for RPA, one for document processing and OCR/IDP, one for low-code, one for BPM and long-running workflow. Those four were always overlapping and converging into the same space.

Gartner's Business Orchestration and Automation Technologies category, the "BOAT" slide, is the market finally naming that convergence. The important part sits in the middle: business process orchestration, the connective tissue between tools rather than the tools themselves. It coordinates humans, digital workers, AI agents, and integrations into one workflow.

Some people see that and think it is just business process management. It is bigger than that. When agents join the mix as autonomous workers executing real tasks, something has to bring it all together, or the result is fragmented and chaotic. The governance question stops being "what am I giving this one bot access to" and becomes "what does every component have access to, how do they share data, and when are they allowed to make decisions on their own."

The Progress Report Is Sobering

Two numbers frame where enterprises actually are.

MIT's The GenAI Divide: State of AI in Business 2025 found that 95% of generative AI pilots delivered no measurable P&L impact, with only about 5% of integrated systems creating significant value. Provocative when it landed, and largely right.

Right behind it, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value, or inadequate risk controls. Those same three causes show up across nearly every survey on the topic. They are worth taking one at a time, because each has a mitigation.

Failure Cause 1: Escalating Cost

Every prompt costs something. Anyone on a free ChatGPT license has hit the wall where it tells you to come back tomorrow, and that wall exists because inference is expensive. Now picture an enterprise automation prompting a large language model a hundred times a day, at scale, across many processes. The token cost is real, the GPU and cloud infrastructure cost is real, and agents that decide to change direction on their own can drive compute cost up further.

Too many companies ran pilots without modeling any of that, pushed as many exciting projects forward as they could, and hit a cost wall that killed the ROI.

How to mitigate it:

• Start small and deliberately. Prove the economics before you scale them.
• Track cost in real time. Monitor API calls, cloud cost, and compute usage as you go, not in a post-mortem two months later.
• Choose the right tool for the job. Many processes are still best solved by deterministic automation. If you do not need agentic capability, RPA or BPM will be far more cost-effective. Unless your company name is Google or Meta, the choice of the right tool is not going away. It becomes more important in an agentic world, not less.

This is augmentation, not replacement. Agents are a very shiny, very capable addition to the tool belt, not a substitute for everything already in it.

Failure Cause 2: Unclear Business Value

Organizations implement agentic AI because it is trendy rather than because it solves a defined problem. When the value is not clear, stakeholders do not see measurable benefit, support and funding wane, and KPIs drift out of alignment. An agent that shaves 5% off processing time rarely justifies its complexity and cost. Teams end up building impressive solutions that do not touch a real bottleneck.

How to mitigate it:

• Define measurable KPIs upfront and tie agent outcomes directly to business metrics: time saved, revenue uplift, error reduction. The framing shifts with the technology. The old goal was "automate invoice processing." The new goal is "all invoices paid and reconciled by month end," a broader outcome.
• Favor continuous improvement metrics over one-time ones. Headcount reduction is both the wrong metric and a metric you can only claim once. Track cost per invoice, cost per PO, cost per payment, quarter over quarter.
• Involve business owners early and co-design the workflows with the people the agent will actually affect. Handing a finished agent to a team that never helped design it is how you miss the mark.
• Communicate the benefits in the stakeholders' language, documenting success stories and ROI even when the wins are incremental.

Failure Cause 3: Inadequate Risk Control

Governance is front and center because executives read the same headlines everyone else does about AI going rogue, and because the regulatory list keeps growing. As agents become more autonomous, they can act on incomplete, biased, or misinterpreted information, and without robust controls a single erroneous action can propagate across many systems.

How to mitigate it:

• Stand up a governance framework before you begin. This is basic, and it is exactly what the successful companies do first.
• Assign clear accountability for each agent. A named person or team supervises it. Human oversight is not optional.
• Fold AI oversight into existing risk and compliance frameworks. A governance board built for the previous era needs to be extended, not assumed to cover agents.
• Log every agent action and decision. A full audit trail is critical, so you are never reverse-engineering what an agent did after something breaks.
• Set thresholds for when a human in the loop is required, and define which decisions can be probabilistic, which need human validation, and which go straight to a customer. Those are very different risk levels.
• Build an incident response plan. If an agent causes harm, how do you roll it back, shut it down, and prevent cascading failures?

One practical control: constrain where agents look for answers. Tell them which data sources to use, in what priority, and in what sequence, so you are containing where the results come from rather than letting them roam.

A Governance Framework in Six Parts

Ashling's position is that siloed, tool-by-tool centers of excellence are the wrong model, because the technologies have already converged. Six categories cover the governance and CoE structure that works across the board:

The talent and pipeline pieces connect directly to change management. People need to see a path for themselves as this rolls out, and the pipeline is how a program grows naturally, by demonstrating value and demand rather than mandating adoption.

The Operating Model Is an Evolution, Not a Reset

SS&C Blue Prism extended its two-decade-old Robotic Operating Model into an Enterprise Operating Model to account for agentic work. The ROM that helped companies automate at scale now adds data and model governance for AI and ML, AI model lifecycle management, human-AI collaboration patterns rather than just human-bot handoffs, responsible AI and decision transparency, and explainability requirements.

The point worth holding onto: roughly 90% of this has existed for 20 years. Nobody locked a room and emerged five hours later with a new methodology to chase agentic momentum. A tried-and-true model got enhanced to stay relevant. That is the honest version of how good frameworks adapt, and it maps cleanly onto Ashling's own governance framework, which many Ashling engagements have paired with the Blue Prism operating model.

Change Management Is the Multiplier

Change management gets named in every project and rarely gets the emphasis it deserves. In an agentic world it matters more, because trust has to extend past "does this work" to "will it make the right call in an ambiguous situation."

The sequence that works:

• Prepare the organization. Before anything changes, make sure people understand what is happening, who is involved, and who is affected. Be as transparent as possible, because people fill information gaps with their own assumptions, and those assumptions are rarely generous.
• Define the why. Most teams over-invest in the who and the what and skip the why. People need to get behind the reason, tied to corporate objectives. Everyone rowing in the same direction, knowing they are a valued rower.
• Reinforce relentlessly. People need to hear a message five to eight times before it sticks. Telling them once in January does not carry to a November go-live. Communication has to be constant and iterative.
• Enable with support. Train people on the new process, what their new role is, and how they benefit. Job aids, super users embedded in departments, and an FAQ or help desk all signal that there is ramp time and that nobody is expected to see the benefit on day one.

Two additions from the SS&C Blue Prism side. First, keep repeating that this is augmentation, not replacement, because someone who sees an AI agent will assume it replaces them and ten others unless you say otherwise, often. Second, process redesign becomes unavoidable. You cannot simply automate existing processes, because agents work differently and enable new operating models. More companies are doing genuine business process reengineering now than in the last five to ten years.

What SS&C Blue Prism Learned Running It Internally

SS&C Blue Prism put its own platform to work internally after the SS&C acquisition, in one of the most regulated environments there is, investment management operations. The results shared with the market: over 3,000 digital workers, 20 AI agents built, over $200 million in annual cost takeout, and a positive influence on $600 million in revenue.

The agents came at the tail end, aimed at work that genuinely called for a different level of automation: more autonomy, more ambiguity, unstructured documents like credit agreements and capital account statements, and workflows that vary each time they run. Governance was the throughline. Operating in a heavily regulated industry, the team built a gateway for real-time monitoring of every AI interaction, putting guardrails on both the prompts going into models and the outputs coming out, to mitigate hallucinations and prompt injection.

The lesson for anyone who has not started: it is a process that takes patience. Get the foundation right, deal with governance upfront, and it can become a real accelerator on both cost and revenue.

A Property Claims Example

Property claims illustrate the whole argument. A claim requires multiple people to finalize, approve, and close, and it arrives as a pile of documents across systems: ID uploads, photos of the damage, records, and receipts.

The four stages exist regardless of how the work gets done: intake, review, settlement, and closure.

Manual. People handle emails, attachments, incident reports, and coverage checks, searching for duplicates and reading through logs. Slow, difficult, and hard to repeat consistently. Skilled people spend their day moving information around instead of making the decisions only they can make.

With robots. Documents get collected and processed and systems get populated, which recovers real speed. The decision-making still queues behind people, so delays persist and repeatability stays uneven.

With agentic automation. Agents read the packet, classify it, check for duplicates and fraud indicators, and recommend next steps. Robots continue handling the mechanical work. People focus on exceptions, approving and validating what the agents surface.

The same pattern travels across industries. In finance, agents analyze transactions, robots reconcile them, and people approve exceptions. In insurance, agents interpret documents and flag anomalies while adjusters make the final call. In public sector work, agents classify cases and humans make sure outcomes stay fair and equitable.

Robots Are Not Going Anywhere

Duplicate claim checks and fraud checks generate many sub-processes per case, each with its own need for coordinated human review. That expands the addressable surface for agents, and it raises the bar on how tightly the handoffs between agents, robots, and people have to be orchestrated.

At the same time, plenty of steps have to be 100% accurate every single time and will be examined by auditors and internal controls. Payment processing is the obvious one. Deterministic, rules-driven automation is the right tool there, and it will stay the right tool. Nobody should want an agent independently approving a million dollar claim.

The takeaway from the webinar's 2026 outlook: this stops being framed as a choice between traditional automation and agents. Combining rules-based automation with probabilistic automation inside one governed process is the path forward, and the two do not have to live in separate lanes. Enterprises will also decide they cannot compromise on visibility and control once they understand both the power and the limits of purely agentic approaches.

AI Explainability Is the Real Adoption Blocker

Ashling's view on why enterprise agentic adoption has lagged the noise around it: explainability.

Moving from deterministic to probabilistic technology is a cultural shift for a large company, not a technical migration. That shift is manageable inside a digital product engineering team. Doing it enterprise-wide, where auditors, internal controls, and regulators are all in the room, is a different order of problem.

This is why the executable log matters as much as the execution. Full traceability across every action, whether taken by a human, a robot, or an AI agent, including the specific LLM calls and tool calls an agent made, is what converts a black box into something an enterprise can approve. UiPath applies the same standard to agents built outside its platform, which is what makes the traceability argument hold across a mixed estate.

If you are building the measurement side of this, Ashling's guide to measuring agentic AI success covers the frameworks that go past cost savings.