PRAPI Research · 2026-10-24
The AI Rules Non-Technical Teams Actually Follow, and the Ones They Route Around
Most AI policies for non-technical teams are written once and quietly ignored. We asked people who set the rules for support, operations, sales, marketing and HR teams to name one AI rule their people genuinely follow and one that got routed around. The answers split cleanly: rules hold when they are one concrete sentence, enforced inside the tool, and fastest to follow. Blanket sign-offs and "approved tools only" mandates got routed around until they were narrowed to the cases that carry real risk.
10 contributors cited
Almost every rule that held had one thing in common: following it was the fastest path, or the only path the tool allowed. The rules that failed were written for a document, applied to every case the same way, or depended on someone remembering them at a busy moment. When teams fixed a rule, they rarely made it stricter. They made it narrower, or they moved it into the tool.
Pattern 1: Police the input, not the tool
The data rules that held were one sentence long and enforced by the software. Harrison Tang of Spokeo described the version his support team follows:
No customer personal information goes into any AI tool that isn’t on our approved list, and the approved tool automatically masks names, emails, and account details before anything is processed. It holds because it sits inside the work. Agents never have to remember the rule or make a judgment call, since the approved path is also the fastest one.
He knew it had landed when "a new hire asked why she couldn’t paste a ticket into a public chatbot, and three teammates answered before I could." Ace Zhuo, CEO of TradingFXVPS, whose VPS support tickets are full of trading account logins and broker server IPs, saw the same peer enforcement:
Recently a new support agent attempted to paste the entire ticket into ChatGPT to have it automatically write out a response for him. A team member stepped in before he hit enter and told him that he couldn't input that information into public AI tools.
The rule that did not hold was the tool mandate. Will Mitchell of StartupBros tried to keep all internal drafting in one approved place:
What happened is that folks used whatever was fastest on their phone, pasted the output into our docs, and nobody mentioned it. I only found out because the voice in a few internal briefs didn't match anybody on my team.
His fix: "instead of policing the tool, I started policing the input. The rule now is about what you're allowed to paste in, not where you paste it." Vic Akosile, Head of AI Transformation & People Strategy at Cloudflare, saw "only use the approved tools" fail on offer letters, where every new country and role type changes the template. "Our applicant tracking system couldn't keep up, so people found their own ways to get letters out the door." So the team built a tool for that one job, called OfferGen. "Now the approved way is also the easy way."
Pattern 2: Blanket sign-off turns into a workaround
The most common rule to fail was a manager approving every AI draft. Harrison Tang watched it happen at Spokeo:
In practice it created a queue, and agents started rewriting AI drafts by hand just enough to call them their own and skip the sign off. I caught it when I compared response times and noticed the tickets flagged as AI assisted were slower than the ones that weren’t, which made no sense until I asked agents about it.
Ace Zhuo saw the same thing at TradingFXVPS: "Within two weeks all such replies got lightly edited and sent out as if they were written by that support or sales agent themselves." The near miss was a sales reply promising "sub-1ms latency to all brokers", which he calls physically false.
Both fixed it the same way: by narrowing the rule, not tightening it. Spokeo now requires human review "only for refunds, account closures, and anything involving a legal or privacy request," with a weekly spot check on everything else. "Agents stopped hiding their AI use, and review time went to the cases that actually carry risk." At TradingFXVPS, sign-off now applies to claims about latency, uptime, pricing and refunds, and Zhuo says compliance has gone up.
Pattern 3: If a rule depends on memory, move it into the tool
Teams running AI agents found that the agents route around written rules too. Toni Kiss, Co-Founder and Creative Director at Clan of Creators, uses Claude Code to build client websites. The rule that holds is that the AI only ever saves a draft:
A person reads the draft and presses publish. It holds because it costs nothing to follow, and because the person who would be embarrassed by a mistake is the one clicking the button.
The rule that failed was "don't fire bursts of requests at a client's hosting."
The rule was written in the assistant's instructions three times. In a long session it still checked live pages one by one, about thirty requests in a row, because page by page felt thorough. The firewall banned our office IP, and every site on that hosting account went dark for us for about an hour.
The fix was to stop writing the rule down and turn it into code: a hook that blocks any command that would send a burst of requests to one client host. In Kiss's words: "if your rule depends on someone remembering it at the busy moment, it will be routed around."
Tom R., founder of KeptRing, which is building an AI phone receptionist for home-service contractors, is pre-launch, so his rules govern his own small operation rather than client results. Nothing goes live, and nothing goes out under his name, without a human sign-off. It holds because "the approval sits inside the work. It's one test call or one tap on a draft, not a policy document someone has to remember." The rules he does not leave to anyone's judgment are built into the receptionist itself:
Before any conversation, it says it's an automated assistant and that the call may be recorded, and it never claims to be a person. For a gas smell or CO alarm, it tells the caller to leave the house and call 911 or the gas utility from outside, and it gives no repair instructions. A contractor can change the greeting, the questions, and what counts as urgent. They can't switch off the disclosure or the safety script.
His test for any rule: "a rule people follow is one where the approved path is also the fastest path."
Richard Meadows, Head of Content at Streamrise, where an AI agent handles much of operations and support alongside the owner, traced his firmest rule to a specific failure. The rule: "no change to a customer's balance without an explicit instruction for that specific action. Investigation is free; moving money is not." It was written after June, when 41 refunds across 30 customers were applied directly and a check against the delivery provider's records showed that 30 of those 41 orders had in fact been delivered in full. "Every one of those refunds was reversed four days later."
His routed-around rule was review before deploy. On 3 October about 60 changes went live before anyone outside the work had reviewed them, because the work arrived as many small fixes and "review" quietly became "review later." Review is now a gate before deploy, and in his words, "60 small changes is not a small change."
Pattern 4: Limit what the AI may claim
Will Begeny, CEO of Slipcover Shop, which makes custom furniture covers, drew the line at what the bot is allowed to assert. His service and ops team follow one rule: "a bot can not tell a customer what will work (only what won't work)."
So, our bots will never tell a customer how to measure or what will fit; only a human will.
He says the team learned it the hard way. Most people wanted bots to "affirm and give good news to everyone (because it was fast and seemingly simple)", and "we paid for this efficiency on the back end." Using the bot to check a request against known exclusions turned out to be more accurate and more useful to both the service team and customers.
Where it still leaks: checking the facts
The rule that quietly lapsed in several teams was verifying what the AI said. Dora Bloom of iotum, a white-label business, put it plainly:
The one that quietly stopped happening was checking the source. AI output comes back sounding finished, so people paste it and move on. Nobody clicks the citations. We watched an AI answer quote a price for our own product that we have never published, and that is roughly what the habit costs.
Her fix was a person, not a policy: anything that states a fact about the product or its compliance posture now goes to the person who owns that fact before it goes out. "A stuffy document would not have caught that. One named reader does."
Joost Boogmans, founder of SocialFind, hit the same problem at scale. ChatGPT described SocialFind's recruitment services as "embedded," implying recruiters physically work at clients' offices. "Except we don't work that way." With the site now over 500 pages across Dutch, English and German, correcting those errors by hand "isn't sustainable," so the team is building a central knowledge base of its actual services and approved terminology for everyone's AI to work from.
giving everyone access to AI is the easy part. Getting everyone to work from the same facts is considerably harder.
The through-line
Across every answer, the rule that held was short, concrete, and built into the path people already take. The rule that failed asked for the same effort on every task, or relied on someone remembering it. Harrison Tang had the line for it:
A rule survives when following it is easier than avoiding it. If your approved path is slower than the workaround, people will take the workaround, so fix the path before you tighten the policy.
Contributors
A rule people genuinely follow: No customer personal information goes into any AI tool that isn’t on our approved list, and the approved tool automatically masks names, emails, and account details before anything is processed. It holds because it sits inside the work. Agents never have to remember the rule or make a judgment call, since the approved path is also the fastest one. The moment I knew it was working was when a new hire asked why she couldn’t paste a ticket into a public chatbot, and three teammates answered before I could. A rule that got quietly routed around: We originally required a manager to sign off on every AI drafted reply before it went to a customer. In theory it was a safety net. In practice it created a queue, and agents started rewriting AI drafts by hand just enough to call them their own and skip the sign off. I caught it when I compared response times and noticed the tickets flagged as AI assisted were slower than the ones that weren’t, which made no sense until I asked agents about it. What we changed: We dropped blanket sign off and replaced it with a narrower rule: human review is required only for refunds, account closures, and anything involving a legal or privacy request. Everything else gets a spot check each week instead. Agents stopped hiding their AI use, and review time went to the cases that actually carry risk.
The rule that is actually followed by my team: No customer credentials, server IPs or account logins are ever input into public AI tools. VPS support tickets are frequently filled with MT4/MT5 account logins as well as broker server IPs. This rule is followed because the consequences of a failure are very real and very frequently witnessed by our support team. Recently a new support agent attempted to paste the entire ticket into ChatGPT to have it automatically write out a response for him. A team member stepped in before he hit enter and told him that he couldn't input that information into public AI tools. He understood. The rule that got routed around: all AI-drafted sales and support replies have to be signed off by a manager. Within two weeks all such replies got lightly edited and sent out as if they were written by that support or sales agent themselves. (The near-miss was a sales reply saying that TradingFXVPS promised "sub-1ms latency to all brokers". This is physically false. The latency depends on where the brokers' servers are located.) Compliance has gone up because I have altered the rule so that sign-off is required for only those few occasions when it matters most - namely when we are making claims regarding latency, uptime, pricing and refunds, et cetera. The rest can be dispatched without any further review.
The rule that holds on my team is simple. Nothing goes out to a customer or a supplier with a number in it unless a human has checked that number against the source document. Pricing, lead times, MOQs, refund amounts. AI can draft the whole message, but the figure gets eyeballed. That one sticks because the cost of being wrong is immediate and personal. Somebody has to write the apology email. When I was running sourcing trips with hundreds of entrepreneurs, a single wrong quantity or quoted price meant a real supplier conversation went sideways in front of real money. The rule that got quietly ignored was my attempt to restrict which tools people could use for internal drafting. I wanted everything in one approved place. What happened is that folks used whatever was fastest on their phone, pasted the output into our docs, and nobody mentioned it. I only found out because the voice in a few internal briefs didn't match anybody on my team. So instead of policing the tool, I started policing the input. The rule now is about what you're allowed to paste in, not where you paste it. Customer names, supplier contracts, anything with a bank detail, those stay out. That one people follow, because it's a short list they can remember without opening a policy doc.
I've found that the AI rules people actually follow are the ones that protect their judgment. The one that holds: if AI helped write something going to an employee, like an offer letter or a policy change, a real person reads it and signs off before it goes out. It works because nobody wants their name on something they didn't read. The one that got routed around was "only use the approved tools." Offer letters were the biggest example. We have tons of templates. Every new country needs its own letter, and every kind of role changes it again, from interns to sales pay to equity. Our applicant tracking system couldn't keep up, so people found their own ways to get letters out the door. So we built a tool for that exact job. We call it OfferGen. Now the approved way is also the easy way. When a rule only slows people down, they'll find a way around it. The fix is to remove the slowdown.
The rule people follow: nothing the AI builds goes live on its own. We use Claude Code to build and edit websites, ours and our clients'. It can write a whole page, but it only ever saves it as a draft. A person reads the draft and presses publish. It holds because it costs nothing to follow, and because the person who would be embarrassed by a mistake is the one clicking the button. This week a new service page was built that way: draft in the afternoon, live only after I had read it. The rule that got routed around: don't fire bursts of requests at a client's hosting. Our host's firewall bans any IP that hits it too often. The rule was written in the assistant's instructions three times. In a long session it still checked live pages one by one, about thirty requests in a row, because page by page felt thorough. The firewall banned our office IP, and every site on that hosting account went dark for us for about an hour. What we changed: we stopped writing the rule down and turned it into code. A small hook now reads every command before it runs and blocks it if it would send a burst of requests to one client host. Verification moved to a single call on the server that renders all the pages and counts what we need. Today the hook stopped a four-page check, and the same check ran in one call instead.
I'm building KeptRing, an AI phone receptionist for home-service contractors. We're pre-launch, so these are rules I've set and enforced in my own small operation, not client results. The rule that holds: nothing goes live, and nothing goes out under my name, without a human sign-off. For the receptionist, the contractor calls a test number and approves or edits the script before a single customer call reaches it. Inside my company, AI drafts our outbound emails and listings, and I approve each email before it's sent. It holds because the approval sits inside the work. It's one test call or one tap on a draft, not a policy document someone has to remember. It also catches real mistakes. An AI-drafted directory listing once quoted a price that didn't match our live pricing page. Our "check every claim against the source" step caught it before anything was submitted. The rules I don't leave to anyone's judgment are built into the receptionist itself. Before any conversation, it says it's an automated assistant and that the call may be recorded, and it never claims to be a person. For a gas smell or CO alarm, it tells the caller to leave the house and call 911 or the gas utility from outside, and it gives no repair instructions. A contractor can change the greeting, the questions, and what counts as urgent. They can't switch off the disclosure or the safety script.
Context: Streamrise is a small team, and much of our operations and support work, from investigating customer complaints to preparing refunds and answering tickets, is done with an AI coding agent (Claude Code) working alongside the owner. So our AI rules are less a policy document and more a short list the agent and the people around it work under. The rule that genuinely holds: no change to a customer's balance without an explicit instruction for that specific action. Investigation is free; moving money is not. Why it holds: it was written the day after a real failure. In June an audit concluded that some customers were owed refunds, and 41 refunds across 30 customers were applied directly. A check against the delivery provider's own status records then showed that 30 of those 41 orders had in fact been delivered in full; our internal status column had simply not updated while the provider's feed was down. Every one of those refunds was reversed four days later. Since then refunds go through as a named action with a stated amount, and when an instruction could mean two amounts, both are written down and confirmed first.
We make custom covers for furniture—a whole business of edge cases—so the rule people on our service and ops team actually follow: a bot can not tell a customer what will work (only what won't work). So the data it needs is all about exclusions, where humans can review and assess the positives. It's a very different approach than others have taken (i.e., you want the bot to focus on the "happy path") but what it's really good at is checking a rule and confirming/affirming whether something violates that rule. So, our bots will never tell a customer how to measure or what will fit; only a human will. We learned this the hard way. And while most people wanted to have bots affirm and give good news to everyone (because it was fast and seemingly simple) we paid for this efficiency on the back end. But using the bots to de-risk an engagement is both more productive, more accurate, and more helpful to both the service team and the customers.
The one that holds is simple. Nothing that identifies a customer, or a partner's customer, goes into a public AI assistant. We are a white-label business, so the people using those tools every day in marketing and in support already understand what is at stake. We have also said in print that iotum and its subsidiary brands will never use or sell customer information for any AI training purposes, and you cannot say that outward while being careless with it internally. The rule survives because it is not written in stuffy policy language. It is one sentence, and the reason behind it is obvious to anyone who reads it. The one that quietly stopped happening was checking the source. AI output comes back sounding finished, so people paste it and move on. Nobody clicks the citations. We watched an AI answer quote a price for our own product that we have never published, and that is roughly what the habit costs. Now anything that states a fact about the product or its compliance posture goes to the person who owns that fact before it goes out. A stuffy document would not have caught that. One named reader does.
We use AI pretty heavily at SocialFind, particularly in marketing. One rule our marketing team genuinely sticks to is that the people in our recruitment videos must be real. We film actual people at actual workplaces. AI can help with editing, but we're not interested in inventing employees or fake testimonials. Recently, we experimented with AI-assisted video editing. Our marketing team wasn't convinced. They worried the editing made perfectly real footage look like "AI slop." So we started posting examples on Reddit to get feedback from people outside our own bubble. The rule we've struggled with more is keeping AI-generated content factually aligned with how our business actually operates. I remember ChatGPT describing our recruitment services as "embedded," suggesting our recruiters physically work at clients' offices. Sounds professional, right? Except we don't work that way. I had to correct it. That's the problem with everyone working in their own AI conversations. Something can sound completely convincing while being fundamentally wrong. We've now grown our website to over 500 pages across Dutch, English and German. Manually correcting these misunderstandings every time isn't sustainable. We're currently working on a central company knowledge base containing our actual services, approved terminology and verified case studies, so everyone's AI works from the same information.