Is your documentation ready for an agentic AI rollout in CX?

Agentic AI rollouts are becoming a routine part of contact center operations, and so are the rollouts that stall soon after launch.

When an AI agent in one of these rollouts underperforms, say by consistently approving refunds that should have been escalated, or escalating ones that should have been resolved automatically, the AI model tends to take the blame. In my experience reviewing these rollouts, documentation that was never made explicit enough for the AI agent to follow is a common, overlooked cause sitting underneath that assumption.

A standard operating procedure written for a person relies on judgment an AI agent does not have. A trained employee can read an ambiguous instruction and fill it in with years of context. For example, an instruction to use discretion when a refund request seems reasonable evokes unwritten judgment calls about account history, tone and whether a customer is asking once or asking for the fifth time.

In contrast, an AI agent reading the same instruction has nothing to draw on, so it executes the instruction exactly as written, and if that instruction is ambiguous, it can execute the wrong outcome consistently, at scale, until someone notices.

A documentation readiness check catches this before it reaches a customer. In my experience, it comes down to four things.

1. Get the right people from across the organization into the same room before the rewrite starts.
2. Write instructions explicitly instead of implicitly.
3. Target a reading level an AI agent can parse reliably.
4. Run a self-audit test before anything goes live.

This article walks through what each of those looks like in practice, based on how I have approached documentation readiness across client contact centers, including a migration of a 12,000-person contact center across eight different platform instances.

Standard operating procedures shape whether an agentic AI rollout succeeds in CX

Documentation doesn't often make for a compelling news story, especially when new AI model developments continue to dominate public interest. That attention influences where teams look when something goes wrong. When an AI agent underperforms, the model tends to be the first thing examined, and the standard operating procedures behind it are an easily overlooked cause.

The topic of standard operating procedures and agentic AI came up directly in a recent webinar conversation that I participated in with our partners at Genesys, titled AI Has Changed Customer Expectations. Now What? Alex Ball, senior vice president of Genesys Cloud CX, described a service interaction where an AI agent checked a device's telemetry and, within a minute, sent notification that there appeared to be something wrong with the battery and that a replacement would be sent. He pointed out that the AI tool had to be connected to the brand's backend system to make that possible.

But in addition to being connected to backend systems, this type of execution relies on clear documentation that guides the actions and decisions of the AI agent.

An AI agent will follow vague or incomplete instructions just as readily, which is one of several ways a rollout can go wrong. According to a Gartner survey of more than 3,400 organizations investing in agentic AI, more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.

A risk control is a rule meant to stop an AI agent from taking an action that creates financial, compliance or reputational exposure, something like a cap on refund amounts or a mandatory identity check before an account change. That control works when the standard operating procedure behind it is specific enough to enforce. An instruction that leaves room for interpretation cannot reliably stop anything, whether or not anyone labels it that way when a project gets canceled.

An AI model works from what has been written down in your SOPs or explicitly given to it because it has none of the built-up, informal context that a trained employee brings to an ambiguous instruction on their own. A trained employee can read "use judgment" and know exactly what that means in a given situation. An AI agent reading the same line has nothing to work with unless that judgment has already been captured somewhere it can access.

A related problem is institutional knowledge that never gets written down at all, such as the judgment sitting in call transcripts and in the instincts of top-performing agents. Our article on institutional knowledge in the contact center looks at how to capture that knowledge and put it to work for AI.

A documentation readiness check for agentic AI comes down to four steps

Each of the four checks below closes a different part of the same problem, and none of them require new technology, just discipline before the rollout instead of after. In my view, these must all take place before progressing with additional agentic rollouts.

1. Get the right people in the room

Repeat after me: No single department holds all the knowledge a rollout needs.

During a migration we led, moving a 12,000-person contact center across eight different platform instances, the largest improvements came from getting operations, IT, customer service and business leadership into the same room before the rewrite started. Each group held a piece of institutional knowledge that had never made it into any document, and none of them had been asked for it directly before. IT could point to system behavior that only showed up under specific conditions the documentation never mentioned. Customer service could point to phrasing that read fine on paper but triggered complaints the moment a live agent used it. Neither of those details would have surfaced from a documentation rewrite alone.

Getting the right people in the room makes it so that you, and your eventual AI agents, have all the pertinent information on the table. Once that information is on the table, you can focus on writing it in a way your agentic solution can understand.

2. Write instructions explicitly instead of implicitly

Making a standard operating procedure explicit means removing every point where a human reader would have used judgment and replacing it with a defined rule.

That sounds abstract until you see it applied to a specific instruction. Let’s consider a hypothetical example.

Imagine you’re a service leader at a large retail company. Your standard operating procedures tell human agents to use judgment to determine whether a damaged item should be replaced. That instruction might be clear enough to your experienced team members, and hey, when they were new, they saw firsthand how experienced peers handled these circumstances. They picked up the informal thresholds, the unspoken rules, and know when a replacement makes sense and when it does not. An AI agent reading the same instructions has no threshold, no condition and no defined outcome to work from.

It is for that reason that instructions written for AI agents must be explicit, rather than implicit.

The difference with explicit SOPs is that they remove every place where a person would have used judgment. It has nothing to do with adding jargon or padding the document, and everything to do with closing the spaces a human reader would have filled in without noticing.

Here is a five-minute test any team can run against its own documentation. Read each instruction line by line and ask whether a brand-new hire, with none of the judgment built up over years on the floor, would know exactly what to do here. If the answer is no, an AI agent will not know either.

3. Target a reading level an AI agent can parse

Targeting a simpler reading level makes documentation easier for both an AI agent and a new hire to execute correctly.

In that same conversation with Genesys I mentioned earlier, I made a point I repeat in almost every documentation review I run. Organizations tend to write the standard operating procedures at close to a twelfth-grade reading level, and pass these on for their human and agentic agents to follow. Unfortunately, it’s also common to come across organizations that write SOPs with so much corporate jargon that they don't seem to be intended to be read by anyone or anything at all.

Whenever I encounter these situations, I tell teams to write standard operating procedures intended for an AI agent at a level closer to elementary school than middle school. A simple test for this is the same one used to explain a complicated idea to a five-year-old, often referred to as “Explain Like I’m 5” (or ELI5 for short).

If an instruction would need more detail before a five-year-old could follow it, the same goes for an AI agent. And while that might not be strictly true, it’s merely a helpful guide to keep you on track as you write your SOPs for AI. Simpler language is not a compromise. It is easier for an AI agent to parse correctly, and it happens to be easier for a new human hire to follow as well.

4. Run a self-audit test before anything goes live

A quick self-audit test, run before deployment, can catch whether an instruction produces the outcome you intended.

Before any standard operating procedure goes near a live interaction, I have teams run it through an AI tool and ask the same direct question every time. Given this instruction exactly as written, what would you do? And then we review to see if that matches the outcome we intended. When the answer differs from the intended outcome, the instruction needs another pass. That’s also the case when you repeat the test numerous times and get different answers. You want your agentic solution to do what you expect, in the way you expect, consistently.

Taken together, these four checks are what it takes to build a strong knowledge base, one of three requirements TELUS Digital's guide to agentic AI names for any effective AI agent, alongside a clear purpose and the right tooling. Getting a knowledge base right is a documentation exercise as much as it is a technical one.

The executive's guide to agentic AI: Enhancing workflows to unlock high-value cost savings

This playbook breaks down real-world use cases, cost-saving formulas, and implementation tips to help you assess AI agent impact in your organization.

Access the guide

Have someone who has done this before review your documentation

The lesson underneath all of this is simple. When an AI agent underperforms, checking the documentation is one of the fastest and cheapest steps you can take before assuming the model needs to change. Ask whether the documentation was ever explicit enough to execute without guessing.

As Alex Ball said in our conversation, nobody wins awards for deploying AI. People win customers by making AI feel seamless, connected and effortless.

Getting the documentation right is one of the less visible steps behind that kind of experience, and too often it is the one that gets skipped.

TELUS Digital experts have led documentation reviews across contact centers with thousands of agents and multiple platform migrations, and the pattern repeats across industries. The documentation is rarely the exciting part of an AI rollout, but it is often the part that determines whether the rollout works.

If you are planning an agentic AI rollout, talk to our team. We help brands identify where AI agents can add the most value, then design, deploy and operate them, including readiness checks like the ones covered in this article.