For many organizations, AI is proving easy to pilot but difficult to scale. Pilots often look inexpensive because they run on narrow datasets with a handful of users, explains Ben Schein, chief AI and analytics officer at cloud software company Domo. “But the cost lives in deployment, the moment you connect that capability to real workflows and the systems of record behind them,” he says. “That’s when the real bill appears.” So CIOs must always budget for the gap between when it works in a demo and when it produces governed and durable value.

Organizations can easily get caught out because they run pilots as a technology experiment instead of a business initiative, he adds. “The interesting question is never whether AI can do the thing in a demo,” he says. “It’s whether it should run in this process, and whether it survives contact with production.”

There’s also a lot of pressure on IT teams to be doing something with AI simply because everyone else is, says Naren Gangavarapu, chief transformation and AI officer at Australian Cruise Group.

He calls it AI theater because there’s a big show around AI even though there aren’t that many successful applications of the technology in production environments.

AI costs out of control

According to John D’Emic, CTO at AI observability platform Revenium, one of the big traps when running a pilot is failing to anticipate how quickly consumption can spiral as adoption grows. “As an example from our own engineering org, back in May, a developer opened an AI coding session on his laptop, and it stayed open for four days,” he says. “By the time it closed, it had run 4,819 calls and cost us $3,762. We didn’t budget for this, and no alert fired. But that one session cost more than a lot of teams spend on their entire monthly AI tooling.”

While this showcases how a developer can make a costly error, Dmitriy Anderson, CIO and digital and social commerce leader at home and gardening retailer Leroy Merlin South Africa, believes the pilot trap frequently happens when employees with little or no software development experience vibe code applications. “It doesn’t matter if you can create something in 15 or 20 minutes if the result is AI slop,” he says. “Think dirty code, no consideration for safety, security, and possible data exposure.” In most cases, these pilots are developed with one of the frontier apps, and someone probably used their personal AI subscription, so the costs are negligible, he adds. But if you have a company of several thousand people, and you now want to roll this tool out more broadly, that’s where costs can get out of control.

This scenario is only exacerbated by the introduction of agentic AI, D’Emic adds. “Agents don’t spend money at human speed,” he says. “In the old cloud days, an engineer could spin up infrastructure in minutes and finance might not see the bill for a month, which was painful but recoverable. Agents, though, call APIs around the clock without waiting on anyone’s approval.”

Mind the trap

While cost is a big factor in the AI pilot trap, it should be treated as a symptom of a bigger problem, says Schein. The underlying issue is governance and observability. “An autonomous workflow can fan out into more queries, API calls, and model invocations than anyone scoped,” he says. “So if you can’t see what it’s doing, and spend compounds quietly, you only find out once the invoice arrives.”

In a recent LinkedIn post, Anderson outlined how in just six weeks he built a platform for a fraction of the sticker cost using three AI models orchestrated together. The traditional estimate to build the same tool would have required 2,472 engineering hours from a team, and was expected to take around nine months. “I went through the proper engineering steps and planning, and made sure the application passed a series of cybersecurity frameworks,” he says. “The purpose of this exercise was to showcase that AI can still speed up the process even if you take the time to work through the necessary steps. You can build with AI rigorously and securely.”

So to turn AI experiments into enterprise value, every AI interaction must be attributable: who triggered it, against what data, on which model, and at what cost, Schein says. For each workload, be sure to ask how often it runs, which model tier the job actually needs, and what triggers it, human or automatic. “A frontier model on an automatic trigger and a small model called on demand are completely different cost curves for the same task,” Schein adds.

For Anderson, it’s helpful to use AI to highlight potential gaps, assumptions, or blind spots in your ideas early on. “When you start building an idea, ask the agent to interview you,” he says. “It will go through every phase and ask questions about the important facets of the process, from scalability and budget to deployment options. You can even make AI write a prompt for itself, because it knows its capabilities and quirks better than you ever will. It’s called meta prompting.”

Anil Inamdar, global head of data services for the Instaclustr BU at NetApp, suggests CIOs cost out the whole program, not just the demo. “Generally, the model itself is the cheapest part of the program,” he says. For him, it’s important to have security and governance people in the scoping meeting, not the launch meeting.

He believes the pilot trap is also, or perhaps mostly, a sequencing trap. “A lot of teams are wired to build first and ask permission later, only to discover months down the line they can’t pass a security review or data privacy audit without a painful and costly rebuild. It’s also valuable to define what failure looks like before you define success.

“Pilots tend to die because of no result, which isn’t the same as a bad result,” Inamdar says. “Emphasize to the deployment team on day one that if a target result by a certain month isn’t seen, we shut it down. Otherwise, you’re funding a zombie pilot because everyone’s invested and no one wants to be the one to call it out.”

Enterprises of all sizes in all industries are rapidly deploying generative and agentic AI to automate processes. But the efforts aren’t always panning out.

Some reasons are new and unique to this technology. But others are related to issues we should’ve been prepared for because we saw them during the age of RPA. And in the rush to adopt new tech, some of these lessons are being forgotten.

“This new era of agents puts the same challenges again in front of us, and we need to think about the things we faced back when that revolution happened years ago,” says Agustin Huerta, SVP of digital innovation and VP of technology at Globant, a digital transformation company.

Those challenges often include selecting the right processes for automation, setting up systems to manage those processes, making sure automated processes get the right inputs, and managing the wider impacts of automation, including cultural.

1. Automating the right processes

All the lessons of RPA are carrying over, says Stephanie Bova, digital transformation officer at Novo Nordisk, including the biggest one that just because you can automate something, does it mean you should.

“We think hard before we start creating something,” she says. “Who’s going to maintain it, and where is it documented?”

And of course, is the process itself a good process. “Nothing gets built on a process that hasn’t been optimized anymore,” she adds. “We haven’t done a technology deployment on an unoptimized process for two years.”

And the company is now a lot more selective about how much automation it rolls out, but that wasn’t always the case with RPA. “At one point, everyone who wanted a piece of automation could get something built for them,” she says. “That’s not the case on how we’re approaching agents.”

There has to be real business benefit to the project, she says. “If you can show me the business outcome, we’ll consider it,” she continues. “But we don’t want or need hundreds or thousands of agents deployed. We want them all standardized and monitored, controlled, and auditable.”

Something similar happened a decade ago with RPA, says Huerta, when easy-to-use automation tools became available to people.

“When they were deployed without proper governance, systems got exposed,” he says. “They started stressing the overall infrastructure of the company, and some robots weren’t created in a way for a return on investment. The process ran faster, but consumed more in the cloud, so you ended up putting all the money you saved in the process into your cloud infrastructure, and the total ROI was zero.”

2. It’s not “set and forget”

Legal services company Purpose Legal uses the same basic approach for gen AI-based automation as it did with the previous generation of automation, based on ML, human oversight, and careful validation of the automated processes.

Take for example legal discovery, where documents are produced and shared with the opposing party in a legal case.

“Inadvertent production of sensitive data is a nightmare,” says Jeff Johnson, Purpose Legal’s chief innovation officer. “We always have to evaluate the data. Especially in the legal services context, we need people in the guardrails to make sure the process is on track.”

Without that oversight, problems can escalate quickly.

“If you make a bad decision you may get chastised by the court, lose the case, or lose the client entirely,” he says. “That happened in the past if you trusted automation too much.”

The AI tools today may be more sophisticated, he says, but they’re not perfect. “Even in the world of gen AI, it’s still something we need to watch out for,” he adds.

If anything, the oversight is even more important because of the scale at which AI can work, and how authoritative it can seem.

“Attorneys are more inclined to trust automation now because it interacts with them much more like a person would,” Johnson says. “It’s actually giving attorneys summaries of documents that look like another attorney wrote it, but that doesn’t mean it’s right.”

3. Reaping what’s sown

The need for good inputs goes back to the beginning of the computing era, if not earlier. “If we aren’t proving good inputs and putting good guardrails in place about where the AI gets its input, we get bad decisions,” says Johnson.

After all, data quality is a concern for any company rolling out automation, whether RPA or gen AI.

“Agentic AI won’t solve the entire data quality issue,” says Sabrina Joos, director of program and lifecycle management for new systems for the Americas at Siemens. But there are some differences, she says, in how it plays out.

In the RPA world, data quality was mostly about structured data and stable inputs. So, for example, if the data was formatted in a way the RPA didn’t expect, it might not execute.

“With agentic AI, the data quality issue becomes much more complex,” she says. “It’s no longer just about whether the data is correct, but if it’s complete and meaningful in context.”

AI systems can accept unstructured inputs or ambiguous data and make sense of it, but it doesn’t always interpret that data correctly.

“We don’t care too much about the format or typos since that’s not as much of an issue anymore,” Joos says. “But if there are assumptions that aren’t right, the process or workflow will still be executed. And this is where you have a risk that it will scale.”

For example, an AI can mix up two projects because they sound similar, she says. “Or, working on manufacturing solutions, it might not recognize the physical constraints of a system and will try to optimize and do something that a machine can’t do.”

Or two people might have a different understanding of an issue, and there might not even be an objective truth.

“You need to know where the interpretations are going to be made because there’s not enough information,” she says. “If we can identify this, we can trigger clarification questions. If I get a description from a customer, I might have a different view of it than you.” Solving the problem could involve additional conversations with the sales team, or double-checking with the original sources.

These data quality issues need to be considered early, says Jon Knisley,

director of AI value management at ABBYY.

“It’s really easy to run a pilot when it’s not in production,” he says. “But when you try to move it there, you get data issues. Where is the data coming from, and what’s the risk component?”

That’s also when the governance problems arise, as well as other challenges. These are all fundamentals that companies needed to learn in the previous era of automation and RPA, he says, since we’ve seen this technology cycle before.

4. Respecting change management

The biggest thing being forgotten about is change management, says Knisley.

“An AI project isn’t going to fail because of the model,” he says. “It’ll fail based on people and process. And there’s not that balance yet between the technology, people, and the process. Especially in North America, we want to solve every problem with technology. And that’s just not how the world operates.”

Back in 2019, according to a Forrester survey conducted on behalf of UiPath, 82% of respondents said change management was a challenge for RPA deployments. The same is true today. In a recent Kyndryl survey of over 1,100 business leaders, the speed of AI has outpaced workforce, governance, and operating models for 79% of organizations, and only 9% of organizations have implemented change management, redesigned roles around AI, and built workforce readiness.

“The biggest challenge we had in any digital transformation — and still have — is change management,” says Rahul Chhabra, director of applied AI at Herbert Smith Freehills Kramer, a leading global law firm.

And it’s gotten harder. With AI in particular, the technology is evolving so fast that change management is a quickly moving target.

“With RPA, it was sort of simple,” Chhabra says. “We had frameworks we could use to train people. We still have those learnings, but we have to enhance those processes.”

Something that works today might no longer work tomorrow, either. “You have to constantly iterate,” he adds. “What has worked can fail fast and only work in modules. So don’t try to solve the entire problem in one go.”

And employees don’t just have to keep learning new skills and adapting their work processes. Knowledge workers in particular also have to face the constant fear that AI will make them irrelevant. It doesn’t help when AI leaders amplify these fears. For example, Dario Amodei, CEO of Anthropic, predicted that AI will be capable of doing most or all jobs, not just entry level, in less than five years.

“Today, if lawyers do 10 tasks, maybe four of them will become obsolete,” says Chhabra. But that doesn’t mean four out of every 10 lawyers will be laid off. Even if most of the work is automated, Chhabra adds, there’ll be more for lawyers to do, not less.

“Today, a litigation matter might be worth $1 million,” he says. “But if it’s just $150,000, then a lot more matters are brought forward. So there’s going to be an increase in litigation and, therefore, more work for lawyers.”

RPA isn’t dead

So is RPA over? RPA wasn’t smart, says Traci Gusher, data and analytics leader at EY Americas. “It was useful, but it wasn’t intelligent. You couldn’t rewrite the process with RPA because it wasn’t technologically advanced enough.”

So, about 15 years ago, during the big RPA wave, organizations looked for ways to use RPA inside their processes, but the benefits were extremely limited.

“It was never so demonstrative that it would catch investors’ eyes,” she says. “It never got to that level of impact.”

Today, many companies are making the same mistakes with AI, she says. Instead of adding AI to existing processes, they need to rebuild them from scratch. “If you’re chunking it, you’re not going to get the results you want because it’s too small and incremental. That’s why I think over a period of time, RPA died a slow death.”

But AI can actually bring RPA back to relevance, she adds.

“There’s still very much a place for RPA in the AI wave,” she says. “You can use RPA for tasks and transaction-level activities, and integrate with agents. That might be the most cost-effective way.”

Unlike agentic AI, RPA is deterministic and completely predictable, it can run on-prem without leaking any sensitive data, and it incurs no token costs.

“We’ve seen consultants say we need to do this with agentic AI,” says ABBYY’s Knisley. “And they don’t have any reliability or governance. What they’re ultimately trying to do they could’ve done with regex for a tenth of the price, and 10 times the efficiency. You’ve got to figure out when you need to use agentic, script, or regex.”

So instead of throwing out RPA and going all-in on agentic AI, many companies are taking a more nuanced approach, using traditional RPA for processes that don’t require intelligence. Meanwhile, they use AI to help set up, test, manage, and upgrade the RPA, getting the best of both worlds.

“I think RPA is a very powerful technology and it has a place in the world today,” says Chhabra. “Especially on things that need to be deterministic, or you’re automating high risk or compliance workloads. You can mask the PII, but it’s still a risk to the company, so I’d rather use a script or some form of RPA automation.”

And a lot of governance will be rules-based, he adds, or based on RPA.

“There’s a lot of marketing speak that RPA is dead,” he says. “I don’t think that. Even the AI vendors are using RPA in the back, but now they’re calling it workflow automation.”