AI Agents Take On White-Collar Tasks, but New Jobs Will Follow
Accounting, compliance, and CRM work are being automated by enterprise AI agents—yet panelists expect human roles to evolve alongside them

Accounting reports that write themselves. Compliance checks are completed in minutes. CRM systems that log meetings and suggest follow-ups automatically.
These are no longer prototypes—they're already working in enterprise settings. From automating financial admin to enhancing customer relationship workflows, companies including Schroders, Eunice, Melow AI, and Sage, often powered by Amazon Web Services (AWS), are rolling out AI agents capable of handling large parts of white-collar work.
Schroders, the UK’s largest asset manager, uses AI agents integrated with Salesforce to capture meeting notes, analyze tone and sentiment, and generate follow-up tasks—all without requiring human input.
At Eunice, agents compile due diligence reports by scanning public and private data, while Melow AI uses agentic automation to support data engineering teams with ontology mapping and pipeline creation.
Sage, a UK-based software company, is embedding its AI assistant Sage Copilot into accounting platforms across five countries. The tool can automate workflows, generate insights, and handle domain-specific compliance tasks using AWS infrastructure like Trainium chips and Amazon Bedrock.
Their impact is unmistakable: AI agents absorb repetitive tasks traditionally handled by analysts, accountants, and support staff. But panelists at a recent London event argued this shift doesn’t spell mass unemployment—if anything, it may unlock higher-value work and create new roles.
Replacing some jobs
At the AI Rush event on May 16, 2025, panelists debated the rise of AI agents in business and what that means for the workforce.
“Yes, we can see it. I really believe that AIs will replace some jobs. This is the reality,” said Konstantina Kapetanidi, who leads global CRM, Data, and AI Sales Enablement at Schroders. “Is it news to anyone? Can anyone say they don't see that right now?”
But her outlook wasn’t entirely pessimistic. “There will also be new types of jobs. That's why I am optimistic,” she said. “We need to upskill constantly. It’s a wave. Without tractors, the farmers couldn’t do what they do.”
Pepe Del Castaño, co-founder and co-CEO of Melow AI, agreed with the mixed outlook.
“I think they—AI agents—will help augment all the things that you can do as an employee, as a person,” he said. “It will help us do the things that we like, as opposed to being stuck doing all that menial, grunt work.”
Yiming Yan, Chief AI Officer at Eunice, emphasized the same theme: “They are here to help us reduce the amount of tedious work. Meanwhile, you regain the time and focus on more high-value-added work.”
Yan described a significant success: “With the help of our average AI agents, they actually managed to do six months of work within one month, with 90% cost reduction. The company gained time and moved them to the next project. The team got split into two. They managed to work on two projects and bring in five times the revenue.”
Embedded, not experimental
Tom Ellis, solution architect manager for software and technology at AWS, explained how these capabilities are becoming standard across enterprise software.
“The biggest area for agents right now is productivity assistance,” Ellis said. “You have generative AI assistants that can help you write code. Now you’ve got agents that can automate tests or fix bugs for you. It doesn’t replace the developer. The developer is still there, driving it.”
He added: “With an agentic approach, and wrapping a common protocol around your service or software experience, you can adapt and connect to other services. Perhaps they would have done that in the past with an API, but wrapping it with an agent gives you a lot more flexibility and power.”
AWS’s partnership with Sage is a leading example. The companies developed custom AI models that help small and medium-sized businesses automate financial tasks while remaining compliant with regulations. In one case, the Semantic Search API, powered by Amazon Bedrock, helps retrieve and rank relevant financial data more precisely.
Purpose-driven adoption
While the tech is powerful, panelists repeatedly stressed the importance of purpose-driven adoption.
“A lot of companies have this FOMO (fear of missing out)—they're saying, ‘Yes, let’s adopt it.’ But what does that actually mean?” said Kapetanidi. “They need to know why they’re adopting it and what use cases they’re solving for.”
She described a fundamental shift in mindset: “We need to sit down and say, if we had a blank piece of paper, how would we think of the process from the beginning? It’s all about reimagining the process.”
Del Castaño made a similar point: “How do you actually translate adoption into a good business outcome? That’s going to be the biggest challenge to unlock mass adoption.”
Even with powerful agents, use cases need to be constrained. “They’re not there in terms of sophistication yet,” Del Castaño said. “So you wouldn’t want to give an agent complete free rein over something that requires very deep regulatory or industry-specific knowledge.”
Yan added that current agents “work for certain areas” like information retrieval and summarization, but fail in judgment-heavy tasks.
“They won’t be able to handle all scenarios that a human can handle,” he said. “That requires experience, trust, and industry knowledge.”
Ability to reason
Panelists agreed on one thing: AI agents today are just the beginning.
“We’re building on the worst model we’re ever going to be building on,” Del Castaño said. “LLMs are getting better exponentially.”
Yan drew a comparison with child development. “We are expecting something that’s in its infancy stage to act as if it’s in a mature stage,” he said. “But it still has the ability to reason. Even if trained on imperfect data, it can identify what’s roughly right or wrong by comparing sources.”
Ellis emphasized that transparency will help the tech mature responsibly. “The more advanced models are showing us their chain of thought,” he said. “You can see how they got to a decision, and then correct or adjust the input as needed. You still need people. You still need judgment.”
Kapetanidi wrapped the discussion with a dual outlook: “So yes, I think we’ll see job replacement for now. But no, I think new roles will be created for the future. And we have to be ready for that.”