The rise of autonomous AI agents marks one of the most significant technological shifts since the rise of the internet. Unlike traditional AI tools that simply respond to commands, agentic AI systems like IBM's HR bots and Salesforce's Agentforce possess decision-making agency. They can independently analyze real-time data, execute contextual decisions, and take targeted actions - from resolving customer service inquiries to autonomously screening job candidates.
I wrote about AI agents in a previous post last year and want to build on that in this piece. Because agentic AI is no longer just theoretical. Real-world deployments are starting to emerge and the technology is evolving. So, what is the most effective and ethical way to promote agent-human collaboration in agentic AI deployments as they scale? And how can we harness agent power without compromising human values, safety or accountability?
There are plenty of examples of agentic AI now. For example, IBM has developed a variety of HR agents that tap into an extensive catalog of prebuilt conversational AI automations, also known as "skill-flows". They can handle complex HR tasks while meeting compliance and company policy requirements.
An HR agent could streamline employee onboarding by giving new employees personalized support, providing real-time answers to common employee queries, offering tailored guidance and supporting essential procedures and advanced tasks, such as pre-employment checks, learning recommendations, onboarding profile creation and IT request submission. The overall goal of an agent like this would be to help new hires settle in quickly and to improve retention rates.
Salesforce's Agentforce platform helps enterprises build and deploy agents using natural language. It has a library of pre-built agent skills that organizations can work with, including Sales Development and Sales Coaching, enabling organizations to create autonomous AI agents that enhance customer interactions and streamline workflow. Salesforce has already deployed its agentic AI tech internally in customer support; AI agents now handle 83% of customer support queries independently, with human escalations dropping by 50% since.
More broadly, OpenAI recently released a preview of its first AI agent, OpenAI Operator, which can perform computer-based tasks like browsing the web, without needing custom APIs. It can handle a wide variety of repetitive browser tasks, such as filling out forms, ordering groceries, and even creating memes.
Despite the increasing prevalence of AI agents, a full-scale displacement of human workers is unlikely anytime soon. Instead, we should envisage a future in which human workers are collaborating with AI agents. AI can assume responsibility for repetitive tasks, thereby boosting overall productivity, while humans focus on strategic, creative, and interpersonal aspects of work.
AI agents are starting to act not just as workplace assistants but also as emerging economic participants. In some scenarios, AI systems may even hire human consultants for tasks demanding specialized judgment or creative input. This shift highlights how agents can influence the very structure of work relationships, prompting new forms of collaboration and compensation models.
Enterprises are, of course, still moving cautiously, partly because of the potential for AI bias, errors and hallucinations when deploying large language models, as well as concerns about data privacy, security and compliance breaches. So, how can we build collaborative workflows involving human workers and agents while taking these issues into account?
Three pillars of ethical agent-human collaboration
When integrating AI agents into collaborative workflows, there are three important factors to consider:
1. Transparency by design
If you are going to use AI agents, you should have full transparency about how the agents are making decisions. In other words, If you can't audit it, don't deploy it.
For example, IBM watsonx.governance, can track AI models across their lifecycle, with bias detection mechanisms and governance tooling to ensure compliance with ethical standards. Its dashboards, reports, and alerting capabilities can be used by the enterprise staff to audit, and report that AI models are meeting requirements for fairness, transparency, and compliance. These capabilities are supported by features like AI Factsheets, which document model metadata across the lifecycle, and customizable metrics to track fairness, bias, and explainability, ensuring adherence to FPACCTS principles (Fairness, Privacy, Accountability, Compliance, Confidentiality, Transparency, and Safety).
Salesforce's Atlas Reasoning Engine enhances AI transparency and trust by providing inline citations, using Retrieval-Augmented Generation (RAG) to ground responses in verified data, and employing advanced "System 2" reasoning techniques to minimize errors and ensure accuracy in critical applications. It refines queries by expanding them with additional context, evaluates retrieved text for relevance, and synthesizes responses based on the most pertinent information. In this way, it improves decision-making with reliable, data-driven outputs.
The Perplexity Deep Research agent and the OpenAI Deep Search agents enhance collaborative research by providing AI-driven comprehensive analysis. They conduct numerous searches and analyze hundreds of sources to create detailed reports, ensuring transparency through cited sources and improving accuracy in various domains.
2. Human-in-the-loop safeguards
Another best practice when deploying AI agent technology is to always ensure that there is a human in the loop when agents are making important or sensitive decisions, such as those that involve medical diagnoses, legal contracts and hiring approvals.
An interesting example of this is demonstrated by IBM's use of AI agents during the hiring process, balancing efficiency and empathy at scale. While Watson Recruitment processes 4 million annual applications, human recruiters focus on nuanced candidate interactions by leading final interviews, cultural fit assessments, and salary negotiations. This hybrid approach achieves 90% operational efficiency gains in administrative workflows and there is evidence to suggest it can achieve a reduction in hiring bias through anonymized AI screening. It leads to 40% faster candidate-to-interview matching, while maintaining personalized engagement - 87% of chatbot-assisted applications advance to interviews versus 53% in traditional processes. By preserving human oversight for strategic decisions, IBM's model indicates that ethical AI adoption can enhance both productivity and candidate experience.
It is also important to acknowledge the broader ethical and practical implications when AI begins to be used in areas like performance reviews or even team management. Some organizations are experimenting with 'AI middle managers' that allocate resources and track progress, but these pilots maintain robust human oversight to handle exceptions, complex negotiations and final approvals. This underscores a growing philosophical discussion about setting boundaries around AI, especially in decisions that significantly impact human lives.
3. Accountability frameworks
Who is liable when an AI agent makes an error? It's likely to be the user organization or the company that has built and deployed the technology on its behalf. While the repercussions could negatively impact business performance, depending on the type of error, there could be legal implications and brand damage.
Unsurprisingly, AI regulations are starting to address concerns like this. For example, the EU AI Act mandates rigorous risk assessments for high-risk AI systems in sectors such as education, which are classified as high-risk due to their potential impact on individuals' fundamental rights and educational trajectories. The Act also delimits the liability of the provider and of the user of the AI system for any harm it may cause. AI systems must comply with criteria such as transparency, accuracy, cybersecurity, and quality of training data. Non-compliance can lead to substantial fines, ranging from 35 million euros or 7% of global turnover to 7.5 million euros or 1.5% of turnover, depending on the infringement and company size.
Various frameworks are being proposed to ensure that agents' decisions are in line with company policy and industry regulations. For example, IBM has suggested a technology-driven approach through its IBM Alignment Studio. This aligns large language models to rules and values described in natural language policy documents, such as government regulations or a company's own ethical guidelines.
Salesforce's Agentforce platform incorporates the Einstein Trust Layer that has guardrails and other security features including the ability to block agents from accessing sensitive data without authorization.
Overcoming implementation challenges
Next, I will describe three important implementation challenges enterprises need to overcome when deploying collaborative human-AI agentic systems.
Challenge 1: Scaling without losing control
A future vision of agentic technology is likely to involve the deployment of multi-agent systems with specialized AI workflows, requiring the ability to balance scalability with operational governance. However, as enterprises begin to integrate hundreds of interdependent agents - each handling tasks like document parsing API integrations, or decision-making - traditional architectures are likely to struggle with state synchronization (ensuring all agents have up-to-date data), fault isolation and resource allocation.
IBM has developed an open-source tool kit, its Bee Agent Framework, which aims to address the complexities of large-scale, agent-driven automation, orchestrating agents via hierarchical task delegation. Its supervisor-agent-worker architecture breaks workflows into atomic operations, serializing agent states for pause/resume functionality and migration across nodes. Sandboxed code execution prevents rogue processes from destabilizing systems, while OpenTelemetry integration enables granular tracing of cross-agent interaction. The Bee Agent Framework is part of a broader ecosystem that includes tools like Bee UI for visual interaction with agents and Bee Observe for telemetry collection and management.
Challenge 2: Bridging the trust gap
If we want human workers and AI agents to collaborate in integrated workflows, a key requirement is that humans trust the outputs of the agents, which still remains a challenge. In fact, a Salesforce study highlights that 54% of workers doubt the accuracy of AI outputs, while 59% question potential biases in algorithmic decisions.
Additionally, AI agents often require real-time human input for context-sensitive tasks that involve multiple data streams. Integrating and validating this data can become a bottleneck if organizations do not have well-defined processes. Ensuring accurate information flow is crucial for building trust in agent-driven decisions. By prioritizing transparency, reliability, security and user-centered design, organizations can foster greater trust in AI agents and promote more effective human-AI collaboration.
So, what strategies can help to address this lack of trust?
- Anthropic's Constitutional AI methodology, which it has used in its Claude AI assistant proposes training agents to refuse harmful requests by aligning with predefined ethical principles written into a constitution. For example, its current constitution for Claude AI draws on the likes of the UN Declaration of Human Rights and Apple's Terms of Service, among others. The system uses self-supervised learning to iteratively critique and revise responses against constitutional rules, achieving a 37% reduction in harmful outputs compared to human-supervised baselines. Crucially, this alignment occurs during model training, enabling autonomous harm reduction without requiring ongoing human monitoring.
- Salesforce's Agentforce platform incorporates Explainable AI agents which provide visibility into AI decision-making processes. The agents explain the rationale behind AI-driven decisions in simple terms, enabling human workers to confidently collaborate with AI agents. This strategy has been pivotal in enhancing trust and adoption in enterprise workflows.
Challenge 3: Workforce transitions
The integration of AI agents into the workforce is likely to drive a fundamental restructuring of organizational roles and operational patterns.
As AI automates routine tasks, new human-centric roles are already emerging to bridge technical and ethical gaps: Collaboration Designers optimize hybrid workflows by aligning AI's analytical capabilities with human creativity, while AI Trainers fine-tune systems to domain-specific contexts, such as healthcare diagnostics or financial compliance. Ethics Auditors monitor algorithmic decisions to mitigate biases and ensure regulatory adherence.
In their adoption of AI, organizations like IBM and Walmart, for example, are phased transitions - from initial AI augmentation pilots to full restructuring - in which humans shift from task execution to strategic oversight using AI-generated insights.
Adapting to AI in the workplace requires addressing cultural challenges to ensure a smooth transition. Companies need to clearly communicate how AI enhances human roles rather than replacing them. For example, Siemens Energy uses simulations to show how AI helps workers achieve better results, like faster problem-solving. Upskilling programs are also essential, with platforms that provide personalized training to help employees develop their own skills and expertise alongside AI tools.
Upskilling programs are critical. For example, IBM has trained over 7 million workers via its SkillsBuild platform, which uses AI to personalize learning paths for roles like AI ethics auditing and hybrid workflow design.
The road ahead: Collaboration, not replacement
As agentic AI becomes more sophisticated and takes on more tasks in the workplace, organizations that deploy them successfully will view agents as amplifiers - not replacements - for human talent.
For example, in healthcare, IBM's watsonx Assistant has been able to reduce clinician burnout by automating tasks like insurance verification and appointment scheduling, cutting administrative workloads by 20% in pilot programs. By streamlining processes the AI enables clinicians to focus more on patient care while maintaining human oversight of critical decisions.
Similarly, retailers are adopting AI tools like Salesforce's Agentforce to enhance efficiency and customer experience. Agentforce automates order management, appointment scheduling, and shopper personalization at scale, streamlining tasks like order modifications and product recommendations. With pre-built commerce skills for guided shopping and loyalty promotion creation, retailers can better serve their customers while reducing costs and increasing conversion rates. Companies like Saks and OpenTable have already seen productivity improvements, with AI handling routine enquiries and enabling employees to focus on more meaningful customer interactions while delivering seamless, personalized shopping experiences.
Another example is IBM's watsonx Code Assistant for Z which addresses technical debt by automating COBOL-to-Java translation, targeting the 70% of global banking transactions reliant on legacy COBOL systems. By giving users the ability to reduce development time by up to 80% for application analysis and 30% for documentation, it streamlines modernization, lowers costs, and mitigates risks while maintaining performance and security.
As AI advances toward Artificial General Intelligence, ethical collaboration frameworks will define which enterprises thrive. Tools from IBM like watsonx.ai as well as offerings from other providers, can provide the 'scaffolding' to deploy these systems, but ultimate success will hinge equally on establishing smooth and frictionless human-AI collaboration within businesses.
This article was originally published on the IBM Community.