They’ve made stuff up. They’ve bungled basic tasks. In some cases, they’ve taken someone’s credentials without permission to log into an app.
If they were employees, they might have gotten fired. But they’re AI agents, which means leaders are rethinking what reliable AI content operations actually require.
A Business Insider story interviewed everyday people about some of the most egregious mistakes AI agents have made, but the examples were largely consumer-related. Designing agent experiences that mitigate the risk of making (or repeating mistakes) is even more critical for AI content operations within a large enterprise. It’s not that the technology is unreliable. It usually just needs more setup before launching.
Part of the problem is that the mistakes may not happen right away or go unnoticed. At first, content team members might be spending their time getting used to AI-assisted workflows rather than closely monitoring every output.
As AI is deployed, an agent might work as diligently as a newly onboarded team member but fail to double-check with an employee before taking action. Left unchecked, they could wind up consistently introducing problems like:
- Mistagging content based on the proper taxonomy.
- Routing content to the wrong stakeholder for approval, or in the wrong order.
- Writing meta titles and descriptions that violate established brand guidelines.
Depending on the severity and frequency of the error, content teams might become more resistant to further AI adoption, while audience engagement and trust could both take a hit.
What AI content operations inherit vs. what they need
Good agent experience design starts with more than policies, publishing rules, taxonomies, and your existing content libraries, though AI agents need to be trained on all of that too. It still leaves out a lot that you would likely pass on to a new hire.
Imagine a new staffer starting on the content team, for example. You would likely walk them through not only your top content pillars and brand value props, but also how you developed them.
Imagine a new staffer starting on the content team, for example. You would likely walk them through not only your top content pillars and brand value props, but also how you developed them.
There are usually reasons for the promises behind every product, service, and campaign, as well as a connection to the organization’s mission, vision, and values. This institutional knowledge goes beyond anything in an employee handbook.
There are usually reasons for the promises behind every product, service, and campaign, as well as a connection to the organization’s mission, vision, and values. This institutional knowledge goes beyond anything in an employee handbook.
You convey this kind of context to recently-recruited employees because it helps them buy into the “why” behind your work as they learn the “how.” AI agents may not feel like people do, but they reason and need a similar level of grounding to make better decisions and perform to their potential.
Leaving this until after the technology is woven into content workflows can be costly. According to McKinsey research, 60% of agentic AI costs stem from efforts to refine what they do and orchestrate them properly. This could explain Gartner Inc.’s dire prediction that 40% of agentic projects will be canceled in 2027.
Even if AI agents receive institutional knowledge, it’s probably confined to the right way to perform tasks in ideal conditions. Like any other business function, marketing teams operate in constantly changing conditions, where briefing an employee on what to do “in case XYZ happens” can make the difference between a small gaffe and a costly, reputation-damaging failure.
Setting up AI agents for success isn’t about giving them rules. It’s about corralling the institutional knowledge from sticky notes, Slack threads, and the minds of your most experienced employees that show them the exceptions to the rules.
Turning institutional knowledge into agent context
Begin tapping into that institutional knowledge well before AI content operations get formally underway. Assuming you’ve identified the use cases that make sense for your team, begin mapping out the potential “before” and “after” story of what everyday workflows will look like, focusing on the key performance indicators (KPIs) that won the green light for a deployment.
Don’t keep this high level. Break down workflows from beginning to end, discussing with those on the frontlines how everyone gets from A to B.
A blog post might start as an idea in someone’s head, for instance, but it’s often inspired by the data content teams review in their analytics platform. Then it gets outlined, drafted, reviewed, designed, published, and distributed through multiple channels.
AI agents can play a role in every one of these areas, but as they do so, ask the following questions:
1. What did the AI agent decide?
Let’s say your analytics platform shows it makes sense to publish a “part 2” to a popular blog post. The agent doesn’t decide that, but it might be responsible for developing the outline. What background materials does it need to do that?
This might include external data sources, such as market research, of course, but the results could be better if the AI agent were exposed to internal subject-matter expert insights captured on an intranet or in a Slack thread. The product team might have some additional context based on a feature they’re about to put into the next release of your flagship product.
2. Why did the AI agent decide what it did?
If the AI agent’s blog post outline was restricted to whatever it could find on the public internet, it might accidentally reference competitors or make assumptions based on speculations in a social media post. Reverse-engineering these decisions lets you see where you can provide additional context and training, much as you would with a new employee who messes up within their first few days on the job.
3. Under what conditions does an exception apply?
You probably already give outlines a thorough once-over before developing drafts, but in this case, a review is especially important for AI agent governance. It’s the human-in-the-loop AI needs because it prevents the work from going any further before claims are validated and the right stakeholders are assured the angle and approach are brand-safe.
Basic scrutiny from within marketing might be okay for regular thought leadership posts, but posts about new product releases could be an exception, with an AI agent routing the outline to a member of the product team for further input and guidance.
Use this same framework for evaluating how AI agents could help generate and deploy online ads, populate an email newsletter, update your brand’s social feeds, or create a landing page for webinar registrations.
Where human-in-the-loop AI oversight matters most
Excellence in AI content operations doesn’t mean constantly looking over an AI agent’s shoulder. That’s human oversight of AI done right: concentrated at the edges, not applied as blanket surveillance. This is something you can think through at the start of a deployment and fine-tune during the pilot phase.
This is not unlike having junior personnel shadow their senior colleagues as they join the team. When it’s clear they understand the way your organization does things, you can give them more autonomy.
You can even train AI agents, similar to those newer staffers, to recognize when to turn to a more experienced colleague for help. Exception management is the process by which human-in-the-loop AI dictates when a potential problem is escalated to a person.
Institutional knowledge isn’t always visible because it accumulates slowly among individuals and teams. It can be more like an oral history than a set of documented “if this/then that” guides. Building better-performing enterprise AI operations will depend on making this knowledge more legible for both other people and the agentic helpers intended to transform enterprise marketing for the better.
