Imagine you’ve had a busy few months at work and you’re ready for a vacation. Within a few hours, an AI agent can plan your complete itinerary: flights, hotel rooms, and an agenda full of local activities. You don’t know how it came up with this and you don’t need to know. All you care about is the result. If you asked it to plan another trip and it used completely different booking tools, you wouldn’t mind.
In everyday life, that kind of improvisation is a feature. The agent is free to choose its own tools, sources, and steps each time because only the outcome matters.
In professional settings, it’s a liability. When the process matters as much as the outcome, an agent that improvises can’t be explained, repeated, or audited.
Inconsistency is bad for business. That’s true whether you’re talking about commissions, pricing, or revenue reporting. When AI agents improvise solutions and make assumptions—even reasonable ones—they get inconsistent results.
AI agents need an explicit set of rules to guide them, but those rules are usually scattered across a dozen different siloed systems, from spreadsheets to hard drives to SaaS platforms. To break down these silos and learn the rules, agentic AI needs access to reliable data, meaningful context, and human expertise.
Agents need more context
To get insightful AI outputs, you need lots of high-quality data. Responsible companies know this and have spent the last few years feeding their AI systems the cleanest, most accurate data possible.
However, raw data doesn’t explain the rules behind the numbers. To tackle open-ended assignments, AI agents need to know how different systems relate to one another. They need context.
As an example, say you asked an agentic AI to determine the most cost-effective way to ship a product to a customer. The agent calculates shipping costs, transit times, and carrier options, and then recommends a route. What the agent doesn’t factor in is that the company’s supplier agreement includes a tariff exemption for goods routed through a specific port, or that the deal includes a volume incentive tied to a particular carrier. Without that context, the agent doesn’t flag a gap or ask a question. The agent assumes, and the assumption becomes the answer.
Now suppose that assumption touches something like revenue recognition or cost allocation. Applied inconsistently across enough transactions, the agent can create a material misstatement, or, in extreme cases, the kind of error that draws audit scrutiny or looks like fraud.
Part of the problem is the sheer number of systems with different rules. SaaS programs and spreadsheets are at digital resources with clear methods for AI ingestion. But there are also rules and context in legacy software, messaging apps, and employee handbooks that agents might not be able to access. Then there’s institutional knowledge, which is both the most difficult and most valuable kind of context to translate. Every normal workflow, backup procedure, and informal exception at your company lives in your employees’ heads.
Let’s say a business needs a revenue report for the board. Steven in RevOps knows how to find and format the raw data. Debbie in accounting knows how to analyze the information. Arjun in communications knows how to assemble it into something readable. That’s three established relationships and three different skillsets for a single workflow. An AI agent could theoretically do the same thing, but only if a company lays out every step in detail.
Don’t overlook the manual step
Developers and IT administrators know how to program an AI agent, but they don’t know how to price a product or record quarterly revenue. Your domain experts—accountants, analysts, operations teams, and other specialists—are the ones who have the rules and context that deterministic agents need. They can explain every step in a workflow to developers, including which systems and people are involved and what to do if something goes wrong.
This step is manual, and there’s no shortcut around it. Someone must map which systems talk to each other and where they should but don’t, like a CRM update that should trigger a finance system change but never does. Someone must audit legacy software and messaging apps the same way they’d audit a spreadsheet. And someone must sit with the person who’s been running the process for 10 years and capture the standards and exceptions they make without thinking twice.
At Alteryx, we built our AI workflows around the VURA framework, so the logic domain experts define stays Visible, Understandable, Repeatable, and Auditable by design. VURA sets the standard business logic must meet, so AI workflows stay consistent no matter who touches them next. That consistency is what lets developers and domain experts build from the same information instead of working from separate versions of the truth, and it’s what lets agents query the rules instead of guessing at them.
More agents won’t fix the problem
As AI agents become cheaper and more effective, leaders may be tempted to deploy more of them to fix their fragmented systems: 74% of executives plan to deploy agentic AI within the next two years. But no matter how sophisticated the AI models get, those agents won’t be able to accomplish much if companies continue to rely on scattered systems and improvised workflows.
I recently worked with a company that tried to use an AI agent to close its books for the month. Each time they prompted the agent, it came up with a completely different workflow, wasting valuable time and tokens on unpredictable outputs. Their solution was to build a second AI agent to manage the first. Naturally, this approach didn’t work.
If you do get the business rules right, the payoff compounds. Once agents understand how the business works, they can start making the judgment calls that used to require a person: the gut checks, the exceptions, the does-this-look-right moment. This will free your best people to focus on what only they can do.
The enterprise AI race will not be won by companies with the most agents, but by the companies whose agents best understand how the business actually works.
Andy MacMillan is CEO of Alteryx.