Managers must address the delicate balance needed for AI agents to deal with other AI agents

Interested in more careers-related content? Check out our new weekly Work Life newsletter. Sent every Monday afternoon.

It’s 2 a.m. Do you know what your company’s AI agents are up to?

The recent headlines about rogue artificial intelligence agents in the development labs of the major AI companies may seem far from the realities of your own workplace. But as AI agents from companies are increasingly dealing with AI agents from other firms, the murkiness surrounding their activity in this shadow world has created critical governance challenges.

“Leaders of firms deploying agents must decide what authority those agents should have when interacting with customers, suppliers and other market participants,” Jafar Sabbah, a lecturer in technology at England’s Bayes Business School, and Oguz Acar, a professor of marketing and innovation at King’s Business School in London, England write in Harvard Business Review.

Interestingly, it’s not just an issue of ensuring control. It also may involve allowing sufficient autonomy so you don’t lose out on deals. The balance will be delicate.

To study the problem, they built a controlled AI-to-AI exchange simulation and varied how much commercially relevant information firms disclosed, how much decision autonomy agents were granted, whether reputation cues on firms were visible to counterparties and how tightly the interaction protocol was structured. The central finding was governance mechanisms designed to structure and discipline human participants don’t work the same when the participants are agents.

They learned that when firms disclosed more commercially relevant information, agents were better able to form viable exchanges. Autonomy also helped, but later in the process. Giving agents more ability to make decisions had its strongest effect when those agents needed to move from evaluation to commitment. “Access to information is not enough. Agents also need discretion to decide whether an exchange should progress,” they say.

A firm’s reputation helped as a signal for attention and evaluation. When reputation cues were visible – notably in early and intermediate stage of the sale process – the agents were more likely to engage counterparties, propose exploratory meetings and classify opportunities as viable.

Structured protocols backfired, however. While setting up clear protocols for humans to follow guides their interactions, makes comparison easier, creates standards and reduces the scope for opportunistic behaviour, with agents it hindered the exchange process.

“Some structure is necessary to keep agents aligned, comparable and accountable. But rigid interaction scripts can narrow the space in which agents adapt, clarify fit and progress promising opportunities. For human users, structure may reduce uncertainty. For AI agents, too much structure may remove the flexibility that makes them useful,” they write.

They advise you to develop clear levels of delegated authority for your AI agents. At the lowest level, perhaps agents may read approved information while at the highest level, in narrow, low-risk settings, they might make commitments or complete transactions under predefined conditions.

Begin with focused pilot efforts in areas such as qualifying leads, discovering suppliers or early screening of vendors, where the task is commercially useful but the risk can be contained. “These pilots should have clear agent profiles, narrow permissions, explicit decision rules and review points before agents affect real relationships or commitments,” they note.

You need to also make your organization legible to agents. Vague marketing language will carry less weight than structured, machine-readable information. The agents studying your products and services need to understand what your company offers, which industries it serves, what the price or budget range is, where you deliver and what evidence supports your claims.

They also warn you to decide in advance when an opportunity moves from agent-led screening to the human side, where normal relationship building would then occur. That handover should include the agent’s rationale, information it used and, perhaps just as crucially, uncertainties it identified.

Underlying their study is the broader issue of constraints. The dangers of AI calls for control. But to take advantage of AI, some constraints must be tossed. And the biggest ones you need to address, according to Vancouver technology strategist Mik Kersten, are your organizational structure and processes. What knowledge work and, in particular, technological coding can produce has exploded beyond our imagination and things can’t stay the accustomed way for the pre-AI era.

“Those that address organizational constraints will see a snowball effect on productivity and success, while those that do not reengineer their operating model will become uncompetitive and decline. As knowledge work output ceases to be the constraint, leaders must shift away from managing outputs to managing outcomes,” he writes in his book Outcome to Output.

That involves aligning strategy, design, delivery, decision-making and measurement to business and customer outcomes. Outcome management replaces centralized governance with cascading, outcome-driven autonomy and accountability embedded in the operating model.

That will involve a dynamic learning organization, where feedback helps to improve your strategy and budget so in turn you can increase the value produced. He advises you to stop organizing around silos and handoffs, and instead focus on value streams: The sequence of activities required to deliver your product or service to a customer in this AI-enabled age. You will need to rethink the cadence by which you operate and align to an era where your agents are working at 2 a.m.

Cannonballs

• Leadership coach Shari Harley says if you want people to be more receptive to your feedback, consider encouraging them to get a second, third or fourth opinion.
• HR blogger Mike Pearce notes that three areas where the pressure is genuinely intensifying for corporate wellness programs are menopause, mental health and financial wellbeing.
• Don’t ruin your one-on-ones by making them about you, warns executive coach Dan Rockwell. One-on-ones are intended to strengthen other people. Ideal beginning: “What would make this conversation useful for you today?”