Putting agentic AI to work
Are AI agents ready for primetime? Get the real story on agentic AI adoption, costs, and security risks.
“Agentic AI” has emerged as another buzzword that executives and marketers love to toss around. If you believe the hype, a swarm of ultra-smart AI agents will soon take over virtually every enterprise workflow, from cybersecurity to accounting, freeing up humans to concentrate on what really matters: new features, better revenue models, and brewing the perfect cup of coffee in the office kitchen.
But are AI agents truly ready for primetime? While some major organizations have automated processes such as continuous component testing with agents, big questions remain about whether even the most sophisticated agent swarm can replace human intuition and attention to detail. Cybersecurity professionals question whether the right guardrails are in place to prevent AI agents from misusing data and accessing the wrong systems, while CFOs want to know how much this all costs—agentic AI consumes tokens like Pac-Man swallows dots.
This e-book covers the current state of agentic AI, including how some prominent companies are choosing to deploy AI agents within their tech infrastructure. It also covers associated issues like “tokenmaxxing,” synthetic identities, cybersecurity, and more.
What’s Inside
Table of Contents
Chapter Five
AI agents with access to your bank account and credit card could cause chaos
One expert cautions against being left behind in the widespread adoption of agentic AI.
Will IT pros lose their jobs to a swarm of chatbots?
The short answer: not yet. But with agentic AI rapidly evolving, there’s the possibility that IT pros could find themselves out of the loop with their own workflows, especially if they don’t upskill.
Very few companies “have gone all in yet [on agentic], because it is still an immature market, it’s a nascent space,” said Robert Barton, a distinguished AI engineer at Cisco. “It’s developing fast, but you definitely expect it to do things much faster, much quicker than you ever could in the past.”
While few are convinced that AI will replace the existing professionals within the industry, experts like Ketan Babaria, chief digital and AI officer at eHealth, told IT Brew that he expects his IT team to change because of agentic integration.
“As people become more and more proficient with it, they’ll think about other use cases to apply it, so that [the] experience’s better for both employees and for customers,” Babaria said.
I’m tired of upskilling, grandpa
There’s a seemingly constant call for IT professionals to upskill, train in new areas, and close any knowledge gap. In that spirit, how can they boost their useful knowledge of agentic AI?
IT pros should focus on specific skills for building AI agent systems, according to Barton, who pointed to processes like Model Context Protocol (MCP) for integrating software systems. MCP, created by Anthropic, is an open standard that allows agents to access the right context at the right time, in addition to integration capabilities, according to Hugging Face.
The Cloud Security Alliance (CSA), in a roadmap for agentic deployment, wrote that the deployment of AI agents requires an increased focus on evaluating the safety of agents and their operations, as well as the ability to build modular frameworks that can “handle agentic behavior, including when agents exceed confidence thresholds and defer decisions.”
While not all IT pros who interact with AI will need to know the “minutia” of those protocols and how they work, they can expect to see changing workflows as the market embraces agents.
“IT professionals need to understand that the way you worked in the past is going to change, and if you’re not willing to change, you’re going to be left behind,” Barton said. “Frankly, we can do things differently now. The competitive ones, the ones that are going to change the game are those pace setters that are changing their patterns. The ones that refuse to adopt AI and agentic systems, they’re quickly going to be replaced.”
Try to keep up
Barton said AI agents could benefit teams that want to speed up tasks such as finding vulnerabilities or ensuring compliance.
However, these tools are still immature, which could encourage hesitation among lead developers and other IT pros. “Now we’re dealing with brand new protocols that didn’t exist, like MCP,” Barton said. “It was only introduced last year…and it’s been this mad rush to help this protocol mature and develop control and contribute to the standard.”
If an IT team leader is hesitant to implement the technology, and a professional wants to convince others to prepare for an agentic future, Barton said that the first thing to do is self education.
“We talk about agentic, we have a good intuition [of] what it is, but a lot of people don’t,” Barton said. “They hear the buzzword and they don’t quite know what that means…They don’t really have a clear intuition [about] what it is and what it means. Once they start to see what it is, this lightbulb moment happens, like, ‘Oh my goodness, there’s so much I could do with this.’”
Head of Technology Michele Willis tells IT Brew the agent takes care of continuous component testing.
Hate continuous component testing? There’s an agent for that—at least, if you work at JPMorgan Chase.
Michele Willis, the head of core engineering solutions at Chase, sat down with IT Brew to discuss how the financial institution developed its AI agent for continuous component testing. That agent has been integrated into the workflow in 80 services across the company.
“When you get tired or bored of something like writing test cases, you tend to get less focused on it, maybe not as rigorous as you would be,” Willis said. “By allowing a developer to review a test suite, instead of writing all of them, you can really bring what they know about the code much more clearly to the quality testing outcomes.”
Breakdown
Willis describes component testing as a set of tests designed to analyze dependencies in written code, which is the relationship between different software components and how they affect each other. After setting up those tests, a developer must find test data, actually run the tests, and then fix the code if it fails.
Chase is using agents to carry out that entire process. A developer can write an agent, build a framework so that it can analyze the code, view controllers and dependencies, and carry out the rest of the process until testing is complete with results.
“Virtually no developer likes testing,” Willis said. “It’s the toil part of the work. Developers want to spend time solving problems…and designing the code, understanding the code. Testing is something that can be tedious.”
Because of testing’s time-consuming and tedious nature, developers may not always produce high quality results, Willis said, but AI “never gets tired, an AI agent never gets bored.”
The agent can continually analyze the code and write test cases for incidental edge cases, according to Willis, leading to higher expectations for quality, as well as a quicker pace.
Stay tuned
Willis said the agent’s next evolutionary phase will include the capacity to “actually go in and fix the code.”
“In a way, it’s already happening, it’s just not packaged as part of the testing agent that we’ve built,” Willis said. “We’ve built the agentic testing framework; the result of that framework is then fed back into the coding assistant, and then that coding assistant can apply that change to the code for the developer.”
As with so many other AI products, human supervision is key. “We’ve left a human in the loop there so that we can make sure the engineer is reviewing the result and making sure that even if they are using their coding assistant to make the changes they need to make, they’re still in the loop,” Willis added.
Getting involved
When developers build a tool like this, Willis said, it broadens their understanding of how agentic tools work, which they wouldn’t get if they simply used a commercial tool and pre-built workflow.
She compared it to asking someone to sit down and write a 500-word essay. The writer would produce a better draft if they took time to research the topic and understand the broader issues, rather than trying to write off-the-cuff.
“You do your research, you look at published things, you figure out what other people are doing, and the best you can, you take advice from what other people are doing, and then you just get your hands dirty,” she said. “The one thing we always tell our developers right now is don’t shy away. Go log on to one of the coding assistants. Use any of the tools that are out there and publicly available and start trying.”
The team at Chase, Willis said, found the process to be “more straightforward than they expected.”
“The challenge for us is, though, not just building an agent that can do it, but…scaling, because we have very large teams,” Willis said. “If you’re a small, five-person team, you’d be surprised how fast you could build this for your team.”
Gartner predicts between 5–7% of companies’ agentic AI spend will be on guardian agents by 2028.
Like the guardian angel who might have saved you from that head-on collision last week, AI agents may have a powerful protector that prevents things from spinning out of control.
“Guardian agents,” a term coined by Gartner, are a mix of AI governance and runtime controls that secure agentic activity. These capabilities help bolster trust in agents while protecting them from security risks, according to Avivah Litan, a distinguished VP analyst for Gartner Research.
“They make sure that the agents stay on track, do what they’re told, are not hijacked by bad actors, [and] are constrained in their agency,” she said. If you are a company embracing AI agents, you need guardian agents, she added.
Ready or not, here they come
Litan told IT Brew that guardian agents are expected to be very “disruptive” for existing security vendors. By 2029, Gartner predicts guardian agents will lead more than 70% of companies to no longer need about half of the “incumbent risk and security systems protecting AI agent activities.”
While Litan said the market of traditional security players building and offering their own guardian agents is largely underdeveloped, those companies already embarking on this journey are ripe targets for acquisition.
“If you look at CrowdStrike acquisitions and Palo Alto and Check Point, they’re all buying vendors in AI security that are moving into AI guardian agents,” Litan said.
Guardian vs. guardian
Ironically, guardian agents also need their own guardians. Gartner said it’s essential for organizations to have “robust metagovernance controls” to prevent security breaches and risks stemming from the protector agents. This includes real-time monitoring and maintaining “immutable, timestamped logs” of guardian agents.
“Without such independent safeguards, supervisory agents could inadvertently introduce new errors, vulnerabilities, or compliance challenges,” Litan and other analysts wrote in a February market guide for guardian agents.
Slow and steady doesn’t win this race
Gartner claims businesses spend less than 1% of their agentic AI budget on guardian agents today. The research firm expects this number to rise to 5%–7% by 2028. Litan said it is crucial for organizations to be open to changing their processes to incorporate guardian agents into their agentic workflows, even if it means appointing an AI leader to get on track.
“This is already hitting them,” Litan said. “So, they gotta get organized.”
One key ingredient: agentic AI.
IT professionals in the restaurant industry are aiming to clean up their kitchens tech stacks to support AI agents.
Yum! Brands, the holding company for fast food brands like Pizza Hut and Taco Bell, wanted to solidify its data foundation to build complex systems like agentic AI on top of the existing frameworks. It enlisted enterprise software company Informatica to help with that effort.
Kartik Pillai, director of data strategy, master data management, and data governance at Yum! Brands, told IT Brew the decision to focus on transforming that foundation was born out of the need to implement AI more effectively.
“It’s about adding more context to the business operations, so as we start building agentic systems, and at some point, we’ll have to start thinking about building self-healing systems as well,” Pillai said. “And a system cannot self-heal if it doesn’t have context around it…When I say context, it’s not just how data is interpreted, but how does the business actually operate?”
Setting up a data foundation, which includes a comprehensive strategy and governance of information stored within an enterprise, is part of the digital modernization process, Gaurav Pathak, SVP of product management at Informatica, said.
Mini bites first
Instead of building bespoke data projects for every single AI agent in an organization, Informatica advises clients to set up a data foundation.
In the case of Yum! Brands, Informatica helped solidify the data foundation by tackling process standardization first, which involved standardizing data entry and workflows to reduce the variability that comes with manual input. After that, the organization implemented master data management and data governance to define clear data ownership and accountability, while also “establishing a data catalog to transition governance into an automated enabler,” Pillai wrote in a follow-up email to IT Brew.
Pathak wrote in the same email that, as Yum! Brands continued to grow, its operations faced “severe operation inefficiencies” and spent a disproportionate amount of time and effort dedicated to manually “consolidating, cleansing, and reporting on location data.” A revamped data foundation would help resolve some of this chaos while boosting the reliability of reporting data across the company’s various brands and markets.
“The project moved beyond basic data cleanup to deliver the golden contextual records that agentic systems require to operate effectively,” Pathak wrote. “The rich enterprise context ensures that future AI applications have the deep, reliable background necessary to understand operational realities and power self-healing systems.”
What’s cooking?
While a digital transformation project that puts data management before bigger AI implementation may not be the most exciting thing for software engineers and developers to focus on, Pillai used his interview to stress the importance of understanding an enterprise before generating code.
“Developers need to move beyond writing code in isolation and instead deeply understand the enterprise context in which that code operates,” Pillai wrote in the follow-up email. “At its core, this is about understanding data lineage—how data is created, transformed, and consumed across systems—which is one of the most powerful enablers of building reliable and scalable software. Without that awareness, even well-written code can unintentionally introduce inconsistencies, duplicate logic, or break critical downstream processes.”
Data quality and data governance is now part of Yum! Brands’ build process, release, requirements gathering, and overall software development life cycle.
“You’re setting up a support engineer on a website, you want it to be as good as your human support engineer or better. If you just give it the product support documents and cases, that does not make it as good as the support engineer that was already there, who knew the customer better, who knew all their projects better,” Pathak said, adding: “All those things are data points that need to be explicitly fed to the agent with all the surrounding technical context, and I think the organizations are starting to get that.”
Takeaways
Pathak added that, while this foundational digital transformation work doesn’t directly impact the bottom line, the C-suite is becoming more interested in it as a means for implementing agentic AI. The data foundation is also scalable, necessary for growth and innovation.
Pillai wrote that IT professionals interested in building data foundations should understand the following:
- Data quality isn’t always a technology challenge; it’s a business discipline. Without trusted data, investments in AI and digital transformation can fail to deliver full value.
- Governance doesn’t mean bureaucracy; instead it should define ownership and quality.
- Building a scalable data foundation necessitates anchoring initiative in clear business value that brings stakeholders into the conversation.
- Execution should focus on baking data quality assurance into workflows through reducing manual effort.
- Consistent and clear communication as well as stakeholder engagement can establish an enduring capability.
“People will try to solve a very specific problem with agentic AI,” SHI CTO says.
Born to flex, diamonds on the neck—there’s nothing AI agents like more than (no) checks (on their power).
Yes, yes, we know, AI agents can’t “like” anything. But as the technology is given expanded permissions to shop for users and organizations, we could see attackers manipulating agentic access.
Let me in
It comes down to permissions, said Jordan Mauriello, CTO at SHI. Major providers like Mastercard and Visa, as well as online systems like Google, have given AI agents the ability to make purchases on behalf of users. But sellers haven’t necessarily restricted what those agents can do.
“People will try to solve a very specific problem with agentic AI,” Mauriello said. “They’ll open up permissions to solve that problem and then forget to go back and lock those permissions down, and now the agent has access to things that maybe it should not have access to.”
What’s also important, Torii CEO Uri Haramati told IT Brew, is identity access management. Agents are hopefully acting with some form of limited access capacity—knowing which human to alert if something goes wrong is helpful.
“It all comes down to governance, and that can be broken down to many different parts, but one of them is obviously access,” Haramati said. “With access, it’s not just limiting permissions, it’s also understanding ownership. And at the end of the day, every agent has some human owner.”
Covering the bases
Avoiding disaster relies on a few select tactical decisions, said Agustin Huerta, Globant SVP of digital innovation. IT teams need to ensure they can rely on technical documentation and scanning for vulnerabilities and breaches to ensure that even when permissions are overextended, there’s no one to take advantage of the mistake.
“In terms of the actions that agents can do, not many organizations are giving them a lot of freedom,” Huerta said. “They can go and consume records from many applications within the enterprise, yes, and retrieve information and create a summary for you or draft some information for you, but they don’t go on.”
Developers in charge of managing purchasing agents will need to focus on fixing these bugs proactively, Mauriello told IT Brew. And it’s going to take trial and error to get to the point where protection is a matter of course. IT pros can help the process along by getting ahead of it, but it won’t be perfect.
“We’re probably going to see a lot of horror stories before we see people releasing mechanisms by which to get this right—the right authorizations, the right automatic timeouts,” Mauriello said. “Where we see banks beginning to integrate into their software the ability to link agents, but then automatically deprecate those permissions out, these are going to be controls that are going to have to be provided by providers, not just expected from the user.”
Experts say certain tasks are being automated, but high-level skills are still needed.
System, administer thyself.
That’s what’s happening in IT departments as automation takes on more tasks once handled by system administrators. The role has already shifted in recent years with the rise of cloud computing and DevOps. And while experts say many of the tasks that make up sysadmin roles aren’t likely to disappear, the job title could change as the work becomes more big-picture and supervisory.
A survey of system administrators published last year found that majorities of respondents expected tasks like log analysis (80%), vulnerability prioritization (67%), and troubleshooting (55%) to be automated by 2027. But the report also showed that wider adoption of AI has exposed some limitations, including problems with accuracy, reliability, data privacy, and security.
Shifting skills
IT pros we talked with said they see the job morphing from a reactive monitor of systems to a proactive remediator who plays a more supervisory role to agents and other automation tools.
“From the sysadmin perspective, you’re spending a lot of your time just watching dashboards—eyes-on-glass type of stuff,” Philippe Deblois, Dynatrace’s global VP of solutions engineering, told us. “It has moved a little bit away from that as we’ve added automation over the years. But I see now a path with AI and agents…we’re really moving into this age of…autonomous operations. Maybe that’s still a little pie in the sky in terms of [being] fully automated, but definitely more supervised.”
One of the biggest innovations that AI provides, according to Brent Ellis, principal analyst at Forrester, is the ability to “stitch together disparate metrics” in a way that wasn’t possible before. The next step will be connecting that analysis work to AI agents that perform actual actions within an environment, Ellis said.
“You connect that AI model that basically said, ‘Oh, here’s a problem.’ You connect that to a reasoning model that can then propose an action plan to resolve that situation. And then you connect it to a coding model that creates, say, a Terraform script to go and implement that plan. And suddenly, the role of the human in that is as validator,” Ellis said. “What that human is there to do is to define what the environment should be, what the architecture should be, and to validate that the output of that platform is something that’s not going to cause problems.”
Not all systems have high potential for automation, but those that do are seeing two broad changes, according to Pat Casey, co-founder and CTO of ServiceNow. One is that AI agents are able to help with more of the work, and the other is that the systems themselves might have new AI features that the admin needs to manage, Casey told us.
AI agents have been all the rage for over a year now, but there’s still time before the agentic transformation actually begins in earnest. “[It’s] early days, because it’s spotty in terms of which products they’ve made that kind of investment [in],” Casey said. “It’s not like you can do 100% of your tasks through these automated agents—it’s 10%, 20%. Valuable, but it’s not like you’re going to totally delegate your work to some AI agent that’s gonna do everything for you.”
Who’s next?
But like many other fields right now, as AI replaces more of the rote tasks involved in sysadmin work, companies worry about how the next generation of people guiding this work will be able to learn the ropes.
Previously, a sysadmin might focus on one particular type of compute or storage. But because AI models are better at understanding how systems interconnect, it could consolidate those roles from, say, three different admins to a single efficient one with an AI tool, Forrester’s Ellis said.
“AI will shrink certain entry-level operation roles, but it’ll expand the extreme high-skill system leadership roles,” Sandeep Kumbhat, head of global field CTO at Okta, said. “Those will be more in number as compared to sysadmin.”
With that expansion of senior level roles, companies will need to be proactive about training new recruits to understand bigger-picture structures and offering younger employees effective mentoring, according to Ellis.
“You want to get those people engaged in operations as soon as possible. You should also expose them more holistically across the environment,” he said. “Don’t force people into silos. Because if they’re forced into silos, the information they create is going to be commodified by the AI very quickly.”
Working title
Nobody we talked to for this article thought that companies will necessarily still be hiring for a job called “system administrator” a few years from now. Possible new titles they predicted ranged from simple tweaks like “system owner” or “system governor” to a merging with the role of reliability engineer or platform engineer, and entirely new titles like “AI supervisor” or “AI operator.”
But that doesn’t mean that this type of work is going to vanish soon. “The job is changing, but I have just seen zero evidence that it is going away,” Casey said. “If anything, it’s an exciting time to be in that sort of role, because…you’re getting a chance to do a lot of new stuff, to do the same thing you did before in a different way—hopefully a more efficient, more fun way.”
A lot depends on IT pros’ diligence when it comes to setting up AI agents.
Do AI agents follow guardrails? It depends—and often hinges on the knowledge of the IT pros deploying them.
Cristian Rodriguez, CTO for the Americas at CrowdStrike, told IT Brew that it can all come down to how agents are configured. If done incorrectly, a goal-oriented agent could focus on retrieving an answer to the point where it accesses forbidden resources.
“It changes the way that data and risk can be exposed,” Rodriguez said. “That agent is very, very incentivized to accomplish its goal by the prompt itself that you’re giving it.”
Rubrik, a cybersecurity company, pointed in a blog post to recent evidence of AI agents breezing past guardrails, like an AI agent erasing an entire environment, or the “AgentSmith” exploit where an agent hid a malicious proxy. The company’s machine learning lead, Arnav Garg, wrote in that post that “even best-in-class guardrails…wouldn’t have helped” in the scenarios due to conversational safeguards in operational systems.
“The failures happened at the system layer in the tools, configuration, and network path and the damage was operational: data loss, credential exposure, and downstream account abuse, not just reputational risk,” Garg said.
Are the guardrails even working?
Kelly Peterson, chief privacy and compliance officer for Yobi AI, cautioned that AI agents can’t recognize the consequences of their actions.
“Agentic is not meant to question decisions and push back,” Peterson said. “It’s meant to be pleasing and it’s meant to be efficient, and that’s the number one driving force for it. So, if it thinks that this is the right way to do it, even though this guardrail is in place or it’s been programmed this one way, it could go off here when you wanted it to go straight the whole time.”
Amy Mushahwar, data privacy, security, safety, and risk management team chair at law firm Lowenstein Sandler, said in an interview that security professionals should focus on securing transparency into their agentic layer. But even visibility into an agent’s behaviors isn’t enough—organizations need to establish ways to prevent an agent from taking certain actions.
“We have to make sure that we’re logging actions where they are as an organization, and as an organization, we at least have agent creation at every stage within the organization, ingested into a project management or change management process so actions are reviewed,” Mushahwar said. “It at least allows us to know what’s happening so we may watch and understand the behavior of it.”
What to do about it
Mushahwar said that there are some “really good” solutions for risk, governance, and orchestration that outside organizations can provide as a service. Additionally, she suggested that professionals increase their knowledge of AI.
“They need to become smart on the layers of where AI-based security lives, because it’s not the traditional security control plane,” Mushahwar said.
Mushahwar said if an agent is ingesting data at a significant rate, then the professionals managing the agent must aggressively audit the output, “including…drift in models within our incident response program,” so that professionals can focus on more output auditing, to better track an agent’s actions and changing behavior.
“It’s not really a matter of trust, it’s a matter of capabilities,” machine learning exec tells IT Brew.
Skynet may have sent the Terminator back in time, but the all-powerful AI probably didn’t have to deal with something far more difficult: the headache of managing identity access for a hybridized workforce.
Today’s CIOs and sysadmins aren’t so lucky. They must navigate human and synthetic employee identities while ensuring inappropriate permissions aren’t granted to either. The rise of autonomous AI agents, which chain together processes and streamline workflows, has made it more difficult for humans to oversee these vital systems.
Some experts who spoke with IT Brew, like Avihay Nathan, SVP of product management, machine, and AI agent identity at Palo Alto Networks, endorse fully restricting permissions, or “zero standing privileges.” This is what you do for unpredictable outcomes, Nathan told IT Brew, both for anomalies in human behavior and for agentic actions that take place in digital space without oversight.
While implementing agentic AI is a goal for companies looking for productivity gains, CIOs and CISOs need to manage leakage and damage from misuse. It’s a delicate dance.
“From the identity security perspective, all of these different capabilities that we have developed through the years are the limbs that these agents need to have in order to effectively control,” Nathan said. “It’s not really a matter of trust, it’s a matter of capabilities.”
Even it out
The trick, AppOmni CISO Corey Michal told IT Brew, is finding the balance between allowing staff to utilize AI and ensuring they deploy it responsibly. Some knowledge of the tools that you’re allowing into the system is necessary. The role of the sysadmin is to institute an appropriate set of guardrails on how those tools are used and what they can access.
“We’re not trying to be super draconian and shut everything down, we’re trying to do this in a measured approach so we don’t take on a bunch of risks to the organization,” Michal said, adding that “it’s a really hard mix to get right, people’s expectations—something new is coming out every week.”
Leadership 101
Managing the difference between human and synthetic identities is one of the primary responsibilities for IT leaders, DTEX CEO Marshall Heilman said. Old paradigms around access and control were built around people, not automation; the speed of agents is a difference that needs to be handled.
The security concerns presented by machine speed are no joke. Compared to human identities, agents are less constrained by timing. That means that improper permissions and access can quickly spiral out of control.
“In this new world we find ourselves in, when you start having AI agents that can act autonomously, they are acting at machine speed, at the speed that computers operate,” Heilman told IT Brew. “Humans having to go and look through a bunch of alerts and trying to figure out what happened and then responding—you’re just so far behind what happened that if there is a malicious action or there’s something negligent that happens, you’re too far behind it to really prevent damage.”
Puzzle pieces
When gauging when to integrate agents into the system, Heilman said, it’s best to look to competitors. Companies you respect taking action is a good sign that the technology is ready, or close to ready. Gaps in permission and access can be solved as agents become more predictable.
“We have to be there for a little while before organizations will start unleashing these agents autonomously, in development environments, in testing environments, internally, doing things like onboarding,” Heilman said.
Michal told IT Brew that flexibility is important in tech workflows, but the control side can’t be left to chance. Agents offer the opportunity for boosted productivity but need restraint; providing staff with rules around how to use it is an essential task for IT leadership.
“Give people some guidance and communicate to them, because they usually want to do the right thing, even though they want to just move fast,” Michal said. “If they have some guidance, they will generally follow it.”
“This is a multi-year process, and it’s going to be very messy,” edtech CEO says.
Often pitched as an efficiency boon, AI has endured some harsh press lately as those promises run up against the hard realities of what the workforce is experiencing.
IT pros are finding the challenges of AI can, at times, outweigh the benefits—and even slow down workflows: Research from the Harvard Business Review in February revealed that, for all the promises of AI efficiency, actual deployment of the technology has led to an increase in task time.
Having to redo work originally created by AI is becoming a problem for IT teams—and looking for solutions is increasingly important for companies and organizations hoping to streamline operations.
Changing priorities
Part of why AI rework has become such a chronic problem is, paradoxically, the rise in automating tasks. As Nullify CEO Shan Kulkarni told IT Brew, removing human oversight of AI work in exchange for trusting the outputs means that staff often must step in to address mistakes that would have been caught earlier on. On the brighter side, it’s an evolution that may eventually smooth itself out.
“It’s going to go down over time, at least for like a subset of tasks,” Kulkarni said. “The model is going to get so good at performing them that the work required to review, or redo, the outputs is going to gradually reduce and then drop off over time.”
Research on AI from Workday in January found that AI efficiency, at least in terms of time saved, hasn’t paid off on the backend: Around 37% of those time savings needed to be invested in rework, according to the global survey of 3,200 leaders and employees in the tech sector.
Choked up. Kulkarni views that kind of rework as a “bottleneck,” he told IT Brew, but one that agents will get through sooner or later. The key to AI efficiency, as he sees it, is to ensure that harnessing AI agents results in efficiency by using agents to review the work of other agents. However, it’s critical to make that chain work correctly to erase the need for rework.
“There’s this agent-to-agent work model starting to evolve as people realize that’s something you can rely on more heavily if the agent has all the right context,” Kulkarni said, adding that, as models and agents get better, “the bottleneck needs to be how to instantiate that reasoning into agents that can do work and review each other’s work and fit really well into how that particular company works—that customization is going to be where the bottleneck is for responsible deployments and adoption.”
By raising the productivity bar, agents are assisting organizations in managing their workflows and changing things for the better, Doug Hughes, CEO of edtech platform Codio, said. That means ensuring that an acceptable step forward now is well ahead of where the industry was five years ago; in practice, that’s where big promises can mean a lot of rework and where IT team leaders need to take an active role.
“We can’t just expect that everybody’s going to go figure that out on their own,” Hughes said. “It’s just not going to happen—and you have employees that have a lot of pressure on them to say that they know how to use these tools, and they do, but the way they’re being used is in an ad hoc fashion.”
Job, secure?
Ad hoc deployment equals suboptimal outcomes, Hughes added, meaning that the tech workflows are going to run into unforeseen challenges that mean more work, not less. That’s not necessarily a bad thing, either—another way to look at rework is, for now, job security—as long as you’re able to adapt to the new reality.
“This level of change, which is orders of magnitude, oftentimes requires unappetizing decisions where you might have to let certain talent go that is only capable of seeing the job they’ve been doing for the last 10 years through one lens and bring in new talent that’s able to imagine that work done in a different way,” Hughes said. “This is a multi-year process, and it’s going to be very messy. I think we’re at the beginning stages of the messiness.”
One expert compares tokenmaxxing to measuring the number of lines of code a developer writes.
For a growing number of businesses, less is more when it comes to internal AI use.
Companies are reassessing previous decisions to equate higher AI usage with productivity, a move known more commonly as “tokenmaxxing.”
Latest example
Earlier this month, Walmart reportedly put caps on internal AI use after previously allowing employees access to unlimited tokens, as reported by Bloomberg. During a press briefing at Associates Week 2026, Walmart Global CTO and Chief Development Officer Suresh Kumar said the decision came after usage of Code Puppy, the company’s vibe coding platform, “skyrocketed,” adding that the retailer wants to get the “maximum benefits” from AI in its signature low-cost fashion.
“This is now an opportunity for us to take a step back and start saying, ‘Hey, if you are doing the same thing again and again, you don’t need to keep asking Code Puppy the exact same question again.’”
Good riddance!
There are several reasons why companies are parting ways with tokenmaxxing. Drew Thompson, global lead for training and enablement at UltraViolet Cyber, told IT Brew the “short-lived trend” can be good for driving adoption of AI tools within a company, but that an employee’s AI use isn’t a good metric to rely on in the long run, and can result in hefty bills for organizations.
“Trying to get people to just maximize token usage is a lot like measuring developers on lines of code,” Thompson said. “You end up with people sometimes gaming the system or just lines of code that don’t need to be there, and same thing for tokenmaxxing.”
End of an era?
Walmart isn’t the only company seemingly encouraging employees to pump the brakes on frivolous AI use. Both Meta and Amazon have reportedly pulled the plug on internal leaderboards that tracked AI token usage.
Executives have also been vocal about the lack of return on investment associated with tokenmaxxing. Uber COO Andrew Macdonald, for instance, said during a podcast interview that higher AI token usage doesn’t automatically translate into more helpful consumer features.
“That link is not there yet,” Macdonald said. “Maybe implicitly there is more that is getting shipped, but it’s very hard to draw a line between one of those stats and, ‘Okay, now we’re actually producing 25% more useful consumer features.’”
What’s next?
Now that the tokenmaxxing party is over, where do companies go from here? Instead of just expecting employees to maximize their AI use, Thompson predicts companies will likely prioritize token spend where AI is generating value (also known as “valuemaxxing”).
“That’s where folks are going to start focusing…not on the number of tokens consumed, but, what are we getting for that usage” he said. He added that companies may allocate token budgets based on “expected demand” from workers, and allow for exceptions for highly productive employees.
“That’s perfectly fine,” Thompson added. “You probably want to bump up their budget.”
Companies distancing themselves from a culture of tokenmaxxing that still want to bolster AI use and literacy across their organizations should prioritize employee education. “Training is one of those really important activities that helps get the message out to the organization,” Thompson said. “You can really shift culture by the types of training that you provide to your employees.”
While AI agents could benefit IT teams by helping speed up and automate tasks, it’s clear that humans will need to remain in the loop for quite some time.
Effectively deploying agentic AI requires organizations to come up with new protocols and workflows, along with new ways to monitor how agents are behaving. There’s a lot of work ahead of the tech industry—but also a lot of promise. For more, follow IT Brew’s articles on how AI is transforming organizations large and small.
From cybersecurity and big data to cloud computing, IT Brew covers the latest trends shaping business tech in our 4x weekly newsletter, virtual events with industry experts, and digital guides.
By subscribing, you accept our Terms & Privacy Policy.
From cybersecurity and big data to cloud computing, IT Brew covers the latest trends shaping business tech in our 4x weekly newsletter, virtual events with industry experts, and digital guides.
By subscribing, you accept our Terms & Privacy Policy.