Rethinking standard operating procedures with AI agents
AI agents are reshaping CX Standard Operating Procedures, moving humans from step-by-step execution to oversight and optimization. Read: "Everything Starts Out Looking Like a Toy" #254
Hi, I’m Greg 👋! I write weekly product essays, including system “handshakes”, the expectations for workflow, and the jobs to be done for data. What is Data Operations? was the first post in the series.
This week’s toy: a mood board of automatically created images that are not real. What’s cool about this art project is imagining how to use this visual treatment to review, drill in, and zoom out of a pile of images and information. This mood board is fascinating and leaves me wanting more. I’d love to see a UI like this to explore search results and arrange multimodal information.
Edition 254 of this newsletter is here - it’s June 9, 2025.
Thanks for reading! Let me know if there’s a topic you’d like me to cover.
The Big Idea
A short long-form essay about data things
⚙️ Rethinking standard operating procedures with AI agents
What’s your first response when you see a service interruption or lapse as a customer? As a process nerd, one thing that often comes to mind is whether the service lapse happened because of a process break or whether there is an entirely new wrinkle for an existing situation that hasn’t been documented yet.
The standard operating procedure (SOP) sounds like it would fix most problems that ever happen. As the story goes, you walk along with the data through every step and state in a state machine, dutifully documenting the step of the “happy path” and “exceptions” and not knowing how you will validate (and observe) how often the SOP is followed.
The cynic concludes that in many organizations, SOPs go to Google Docs or Confluence to die a slow death where they will be occasionally dusted off and revised when it’s obvious that there’s a problem and the as-written SOP doesn’t match the as-experienced customer situation.
Because that’s the point, isn’t it? In most customer organizations, SOPs exist to drive accuracy, repeatability, consistency, and accountability. When the process they are modeling is relatively straightforward, they do a great job and can be used by anyone in the organization to predict the customer outcome most of the time. When people decide not to use them, we often find out … later than we should.
Yet something new is happening. The change is partly driven by the sheer complexity of software work and the variety of responses customers show when you give them. It’s also propelled by the twin need to lower cost and improve the service environment for both the customer and the human agent delivering the experience.
That thing is AI agents.
Yes, I know you’re probably tired of hearing about AI agents and you might never want to hear about them again. For you – if you don’t like the term AI agent – substitute some other word that you like. The point is that more and more of the customer experience is going to be automated soon, and you don’t have to fear it.
Let’s get in the wayback machine and go to … 2011 or so.
You are a time traveler to customer service software, and you see a stack of rules.
When you get an email, look for this keyword, then add a tag for “escalate”, then change the value of the priority of the ticket to “high”. Then send an email to an email distribution list letting people know when there are more than 10 tickets in queue that have breached the SLA of being answered within 3 hours.
The paragraph above will look very familiar to anyone who has built an automation with Zapier, Pipedream, Jira, or many other deterministic workflows that start with an event, evaluate information based on a process, and produce an output. The difference between what you could do then and what you could do now is that the “rule-based instruction” now looks like:
When answering a customer question, always:
Base your response on the most current company policies and procedures provided.
Reference the knowledge base, FAQs, and any internal documentation as needed.
If the question requires using a tool (such as [Tool A], [Tool B], or [Tool C]), use the tool to find or confirm the answer.
If you are unsure, or if a request is outside policy, politely escalate to a human agent or request clarification.
That’s … a lot different. And when prompted correctly, you get a response that is on par with a very good human operator.
SOPs were the only way to persist memory
Why did SOPs become popular? In a world with many different operators, SOPs let you someone new quickly and to ensure they would deliver a generally good answer. You couldn’t know whether the person would do a great job, and you sampled the calls or emails to assess after the fact if the service experience was aligned with the vision you wrote. SOPs also record the historical memory of the organization into a form you can track over time.
But people don’t always follow the steps you ask them to follow, for lots of reasons. Sometimes, the steps seem redundant and the operator has figured out a better way to deliver great service. Other agents may recognize that you don’t check for quality as often as you need to and that it’s worth it to avoid the procedure on days they don’t feel like using it. And still others are trying to navigate a rigid set of steps in an increasingly ambiguous environment.
SOPS preserve the shared understanding of how we complete tasks. But what if completing the tasks is itself an evolving process?
SOPs for robots AND humans sound a bit different
We’re starting to reach a point where many organizations are adopting automated processes for more than just “Tier 1” questions, like “How do I reset my password.” It turns out that customers have new expectations that can be met when the people are not around, but the agents are still following the same procedures.
In the old paradigm, you might send a customer a request to start an onboarding process, and if they had questions about the steps to complete the process, the timeline, or how to deliver information, you might have several back and forth conversations or zoom calls. An agent given a playbook and a procedure, along with a basic tool enabling them to access key data about a customer, can gather the information, answer questions, and summarize the result for the onboarding manager before an initial meeting.
The role of the human in the loop is shifting. Instead of doing every step by hand, we’re checking to see if some of the steps in the process could be completed by the robot. When it doesn’t work, we know a lot sooner and can iterate the instructions in the same loop as jumping in and handling exceptions.
Moving up the hierarchy of customer service needs
It helps to visualize this shift with a simple diagram. Picture a pyramid with three layers:
Manual SOP Execution:
This is where most teams still spend their time. Humans do all the steps, relying on training and memory.Validation & Testing:
As agents get smarter, they take the first pass at the process. Humans review outputs, tweak instructions, and step in for weird cases.Supervision & Strategy:
The top of the pyramid is where humans spend less time “doing” and more time coaching, refining, and asking, “How can this whole thing be even better?”
Most of the ops folks I know want to climb that pyramid—not because they want to do less work, but because the higher up you go, the more leverage you have. You move from plugging leaks to redesigning the whole plumbing system.
Building a partnership with technology
Let’s address the elephant in the room: does this mean we’re all being automated away?
Sure, some repetitive jobs will shrink or change, just as they have in every previous wave of tech. But the bigger story is partnership. The people who thrive will be the ones who figure out how to work with agents, not against them.
Here’s what that looks like in practice:
Translating ambiguous, tribal knowledge into agent-readable instructions.
Designing “SOP as code” that’s explicit, testable, and easy to tweak.
Setting up feedback loops so agents improve over time (and don’t make the same mistake twice).
Benchmarking agents’ performance and knowing when to intervene or escalate.
Dreaming up workflows that would have been impossible with just humans or just machines.
If you’ve spent any time in automation, you know: what matters is not the tool, but how you orchestrate it.
What science fiction gets right about the future
Any good story about agents needs to reference science fiction. It’s easy to think about “that computer that went rogue” whenever you consider an emergent system or the idea of giving up control to a process. Those are thoughts that happen when you don’t design good checks and balances in your system.
Like any technology system, there will be bugs. The lesson from sci-fi is that our job is not to let technology do all of the thinking and work for us. It’s that there are certain tasks machines can do much more readily than humans and it makes sense to use machines for those tasks. Our job? Design robust systems of supervision, transparency, and feedback.
How are teams using agents within workflows?
Task automation: chaining multiple tasks together with one execution. Instead of “if this, then that”, change data in multiple systems based on the required action
Persistent context: help both the agent and the customer know what happened for this customer in the past. Whether that’s simple summarization or contextual synthesis, agents are great at combing through data and providing a templated, styled result
“Swarming” on a problem: multiple agents might be involved in a process, including some who are automated and some who are human. Following a process, automatically sending out status, and driving resolution based on pre-defined rule is a scenario that can totally happen.
Getting started with agents
You might not be ready (yet) to start using agents in your customer service workflow. But how you do the work that’s necessary to be ready? (Spoiler: this is work you should be doing already if you are running a CS team)
1. Start smart and repeatable
Look for SOPs that are boring but critical: password resets, onboarding docs, and invoice approvals. Next, think about how you would do these tasks without a person. If you can imagine it, you might be ready to test this with an AI Agent.
2. Write like an engineer
Structure your SOPs for agents: clear, step-by-step, no hand-waving or fuzzy logic. If you can write it as a deterministic decision tree where you handle all of the states, either a human or a robot can solve the problem.
3. Trust, but verify
Run your agents in parallel with humans, then compare results. Where do they stumble? What did they miss? Iterate.
4. Move humans up the pyramid
As agents get better, focus your human team on what only they can do: exception handling, customer empathy, process design, and—maybe most importantly—supervision.
5. Document everything
Build your own “agent runbooks” so you know what your systems are doing, how they’re performing, and where to look when things go sideways.
More opportunity, (eventually) less worry
It’s easy to get anxious about all this. Will there be enough meaningful work? Are we automating away judgment? But the reality is, SOPs aren’t going anywhere—they’re just changing shape. The work of the future isn’t about mindless execution, but about testing, guiding, and improving the systems we build.
Start thinking of your team as a hybrid—part human, part agent. Use language like “agent ecosystem” or “digital partnership” to shift the conversation away from automation fear and toward capability building.
The organizations that will win are the ones that master this partnership—not those that rush to automate at any cost.
If you’re not experimenting with agent-driven SOPs yet, start now. Pick a process, formalize it, and see how far you can get. Don’t expect magic; expect learning. The winners in this new era won’t be the fastest automators, but the best learners and orchestrators.
In the end, the best SOPs—like the best teams—are the ones that keep getting better.
What’s the takeaway? Don’t just automate. Seek to elevate the conversation (and the role of the human agents). As AI agents take over routine SOPs, your value shifts to oversight, improvement, and strategic thinking. The future belongs to those who master human–agent collaboration.
Links for Reading and Sharing
These are links that caught my 👀
1/ Quality is rare - When it’s possible to have anything you want (practically) at the click of a button or the speed of thought, quality and craft don’t land at the top of the list by default. What creates a quality output? Karri Saarinen of Linear has a few ideas.
2/ Scene: you open Cursor… - Convention, invented by smart people who do a thing repeatedly, is one way to codify a quality output. I love this practical playbook from Addy Osmani on creating a prompt playbook for programmers. In a world where we are all going to be piloting LLMs and agents, knowing how to get started (and where you are likely going to run into walls) is a key skill.
3/ Rebuilding Mario - Imagine you are the designer of one of the most iconic video games in recent history. Now, reinvent that racing game by doubling the number of players and moving it from a track-based to open world metaphor. How do you do it?
What to do next
Hit reply if you’ve got links to share, data stories, or want to say hello.
The next big thing always starts out being dismissed as a “toy.” - Chris Dixon