How agentic AI is rearchitecting enterprise workflows

2025 was always going to be a pivotal year for AI Agents. The technology had matured, enterprise interest was high, and the use cases were becoming clearer. As with any major platform shift, the path from early adoption to broad deployment has involved real learning — about what works, what doesn’t, and what it actually takes to get Agents into production at scale.

Key speakers

• Praveen Akkiraju: Managing Director, Insight Partners (Moderator)
• Joao Moura: Founder and CEO, CrewAI
• Ritika Gunnar: VP of Data and AI, IBM
• Itai Asseo: VP of AI Research Incubation, Salesforce
• Chad McAfee: VP of AI and HPC, Oracle

Key takeaways

• Agent sprawl is real. Enterprises are deploying Agents at scale, but the shift from building to managing the full Agent lifecycle — optimizing, securing, and governing in production — is where the work now lives.
• Most agentic workflows sit on a spectrum between fully autonomous and fully deterministic. Enterprises that understand this build better systems.
• The biggest driver of success isn’t the technology. It is clarity about outcomes, willingness to rethink processes, and investment in reskilling people.
• Back-office use cases dominate current adoption. Most enterprises are operationalizing Agents internally before deploying them externally.
• Three things will unlock broader adoption: ease of use, trustworthy and repeatable outputs, and manageability at scale.

These insights came from our ScaleUp:AI event in October 2025, an industry-leading global conference that features topics across technologies and industries. Watch the full session below:

The state of Agents: Hype, reality, and what changed

Coming into 2025, the industry declared it the year of AI Agents. What actually happened was messier — and more interesting.

Moura, whose open-source framework CrewAI* has seen rapid adoption, described the year as a turning point. Early on, companies were still figuring out what Agents even meant for their operations. By mid-year, the conversation had shifted. Enterprises began treating Agents not as experiments but as infrastructure. One CrewAI customer scaled from 2,000 Agent groups to over 120,000 in just 15 days — a pattern Moura sees repeating once something “clicks” inside an organization.

Gunnar offered a complementary view from IBM’s point of view. Agent sprawl, she noted, is now a real phenomenon. Enterprises are building hundreds of thousands of Agents using a mix of open frameworks and proprietary tools. The challenge is no longer just building them. It is managing the full Agent development lifecycle — from design and deployment through optimization, security, and governance in production.

Asseo framed the moment with a useful analogy: The industry is somewhere between a 56K dial-up modem and the broadband era. Companies are still logging into AI deliberately, treating it as a separate tool. What’s coming is ambient intelligence: AI woven into every process, invisible in the same way a search engine is now invisible.

“A lot of these companies are already thinking about AI agents as infrastructure now, and there are definitely many companies that are getting to the other side of this with major success.”

— Joao Moura, CrewAI

Where Agents are actually being deployed

The panel was consistent on one point: Most of the real activity is happening in the back office.

Customer care, HR, finance, supply chain, and IT operations are the dominant use cases. Enterprises are moving carefully — deploying internally first, proving out reliability, and only then considering customer-facing applications.

Gunnar identified three primary clusters: customer experience assistants (increasingly personalized and action-capable, not just informational), internal productivity workflows across HR, finance, and sales, and IT and software development, where tools like coding assistants have seen some of the fastest uptake.

Moura added a revenue-generation angle that is gaining traction. One telecom customer built an entire new line of business — a credit and lending product — using Agents to handle analysis and pre-approval, rather than hiring a team to staff it. Another company, a major CPG with nearly 700 brands, has set a target of $1 billion in combined savings and incremental revenue over five years through Agents, running country-level pilots across Brazil, India, the UK, and Canada.

“We’re entering this zone now where just like software had a software development life cycle, agents have an Agent development life cycle.”

— Ritika Gunnar, IBM

Why pilots stall — and what separates success from failure

The panel was candid about what goes wrong. And the answer, consistently, is not the technology.

Gunnar pointed to IBM’s experience across more than 200 agentic engagements since the start of the year. The enterprises that see real ROI share a common thread: They started with a clear use case, a defined outcome, and a genuine willingness to rethink the processes and skills around it. Those that struggle tend to deploy Agents into existing workflows without questioning whether those workflows make sense in an agentic world.

Moura put it plainly: Building agentic systems is fundamentally different from building traditional software. The skills required are different. The mental models are different. Companies that invest in training their teams to think differently about how they work are the ones that see results compound. Those who treat Agents as a drop-in replacement for existing processes often stall.

Asseo identified two dimensions that determine whether an Agent actually delivers value: capability and consistency. Capability — how much the AI can do — gets the most attention. Consistency — how reliably it performs across different situations — is often the harder problem and the one that determines whether real-world adoption follows. Much of the knowledge that Agents need to be consistent lives inside individual experts within the organization. Getting that knowledge out, structured, and usable is as much a people challenge as a technical one.

“Always start with those outcomes, with the process changes, with the people changes that you need — because the technology, as we all see, is evolving at such a rapid pace.”

— Ritika Gunnar, IBM

The infrastructure layer: GPUs, latency, and the economics of scale

McAfee brought the infrastructure perspective and a note of pragmatism.

The GPU gold rush, he argued, is slowly maturing. The dynamic is shifting from acquiring as many GPUs as possible to building a coherent GPU strategy. Enterprises are beginning to think in terms of hybrid environments: smaller, efficient models at the edge for high-frequency tasks; frontier models accessed via API for heavier reasoning work. Older GPU generations, vacated by frontier labs chasing the latest hardware, are becoming viable and cost-effective options for enterprise workloads.

For large organizations, the economics only work when GPU investments are consolidated across departments rather than fragmented by team. McAfee’s point: convergence — across on-premises and cloud infrastructure — is where the unit economics start to make sense.

He also highlighted a factor that gets underweighted in Agent deployment decisions: latency. Agents that don’t respond at a pace that feels natural to users simply don’t get adopted, regardless of their underlying capability. Observability is the other underrated requirement — understanding when an Agent fails, why it failed, and how to surface that feedback to users is essential infrastructure for any deployment that expects to scale.

“When an Agent doesn’t give feedback at a pace at which people expect, the adoption is much more difficult.”

— Chad McAfee, Oracle

Digital labor and ambient intelligence

The panel’s closing exchange moved toward the larger question: what does software actually look like in a world where Agents are doing meaningful work?

Asseo summarized Salesforce’s framing: digital labor. The premise is that teams will increasingly include AI Agents as participants — not just tools — and that learning to manage that mix is the next competency enterprises need to build. Salesforce CEO Marc Benioff’s line, that we are the last generation to manage only humans, captures the direction.

Gunnar added a structural observation: Every piece of software is going to be accessible as an Agent endpoint. User interfaces matter less when any system can be reached programmatically. That shift changes how software is built, how it is sold, and how it is used — and it is already starting to reshape how IBM thinks about its consulting and technology offerings.

Moura closed with first principles. The genie is out. Enterprises will adopt Agents — the only variable is how fast and how well. Three things will determine which organizations get there first: ease of use (technical and non-technical users alike need to be able to deploy and manage Agents), trustworthiness (outputs need to be reliable and repeatable), and manageability (governance, data visibility, and security cannot be afterthoughts). Organizations that solve for all three will lead. Those waiting for the technology to mature before investing in the surrounding capabilities may find the gap harder to close than they expect.

“It needs to be extremely easy to use. It needs to be trustworthy in terms of being reliable and repeatable. It needs to be manageable — otherwise you don’t have governance, otherwise you don’t understand what data is going in or going out, you open yourself for a new vector of attack.”

— Joao Moura, CrewAI

These insights came from our ScaleUp:AI event, an industry-leading global conference spanning technology, investment, and enterprise leadership. Watch all sessions on the ScaleUp:AI YouTube channel.

Note: Insight Partners has invested in CrewAI.