The Drafter Principle
My unsolicited take on AI in customer support, offered with genuine love for the very good people who spoke about it at the SD Summit and who I am absolutely not subtweeting (also, tons of memes).
I’ve gotten a few new followers, so if we haven’t met yet: Hi! I’m Ash. I’ve been doing the whole Customer Experience gig for the better part of two decades, and I have what my friends would generously describe as capital “o” Opinions about things like tooling and especially AI.
As I’ve mentioned previously, last month I was at the Support Driven EU Summit in Amsterdam (which sounds like some sort of arduous trek, but was only a train ride for me. And I am still not over how delightful it is that Amsterdam is just like 40 minutes away), and it was genuinely the most fun I have had at an industry event in years. I usually attend these things for the hallway conversations and come home moaning about the presentations. Not this time. Well played, Support Driven.
Nearly every talk touched AI in some way. Implementations. Ethics. Vendor promises versus vendor deliveries, which at this point is a genre unto itself. And before I go any further: I love these people. They are doing right by their teams and their customers, and nothing below is shade at anyone or anything that was said on that stage AT. ALL. What the talks DID do was rattle something loose that I have been turning over ever since I got off the train. A blog is nothing if not a place to put the things rattling around one’s head, so here we are.
My lukewarm Hot Take: the single best thing you can do with AI in a support organisation, for your customers AND your people, is to make it the Drafter of Things.
The drafter. Not the responder. Not the decider. Not the autonomous agent bravely fielding your customers’ problems while your team updates their CVs in the next room because they Know What’s Coming.
Drafter. That one word is carrying a lot, so let me explain.
No, there is too much. Let me sum up.
A support team produces a frankly absurd volume of writing. Knowledge base articles. Help center content. Escalation write-ups. Post-mortems. QA reviews. Coaching notes. And then, you know, the answering-customers part that is the actual job that everyone thinks is all that they do. All of it competes for the same limited hours, which is why the same person who writes a brilliant article on Tuesday may come up with absolute rubbish on Friday (for which I give absolutely zero judgement. Friday me is not the best of contributors either. See ref: this potentially rubbish article that is only going to be posted today [Friday] by some literal miracle).
I am a firm believer that, despite my deep dislike of a blank white page, the bottleneck is not actually the ideas part. At least not in these environments. In the business world, the bottleneck is in the throughput. The bandwidth. There is simply exponentially more that needs writing than there are hours in which to write it, and I am willing to put cash money down that every leader reading this just nodded, possibly without noticing.
THAT is the specific and very unsexy problem where these tools genuinely shine. The filling of those blank pages. So why not let the plagiarism machines produce the first draft of nearly everything that doesn’t require creativity? Everything from articles prompted by sudden spikes in tickets to draft replies to customer questions. And then there’s the QA process, where the machine can do something no human team has ever had the hours for: look at EVERY interaction and flag the outliers, top and bottom, for a human to actually review. Ask any support leader how much of their queue gets QA eyes on it today. I’m not going to actually ask you to check, because I already know. It’s guaranteed to be less than 5%. Which, yes. That’s still statistically significant, but . . . damn.
And now we must arrive at the part where somebody may start yelling:
What we need . . . what we MUST have . . . is also a human who reads every single word before it goes anywhere. Every article. Every reply. Every. Single. Fucking. Word.
I know. I KNOW. “Ash, you have just described AI with a human-shaped bottleneck. Also, you swear too much.” And I grant that this is exactly how it looks. And damn straight I do. But I am going to hold my ground, for two reasons: one about your customers, and one about your people.
The customer reason first. Every one of us has now been on the receiving end of machine-written text that was technically correct and emotionally wrong (or technically wrong AND emotionally wrong. It can be two things!). We all know the voice, and our customers do too. When a customer discovers mid-conversation that they have been talking to a machine (and again they absolutely DO discover it), the damage goes well past “this answer was wrong” and lands somewhere around “this company lied to me”. And that, of course, is even assuming that the answers being provided are completely accurate and hallucination free. Which they won’t be. This is only one person’s experience both as an implementer and consumer of AI support bots, but there has yet to be a bot created that can actually recreate the human support experience as we’ve been promised.
And then there’s the people reason. The pitch that goes along with most of these tools is headcount. Specifically less of it. The same (or more) output with fewer people at a lower cost. Just look at the savings! Please sign here. For most organisations this is not only a false economy, it could really be called a bait-and-switch (and I’m sure I don’t need to tell you, the trends are starting to show). The throughput gains may very well be real, but between the errors, the customer frustrations, the lack of bi-directional customer feedback, and the absolutely insane cost increases . . . those throughput gains are completely eaten.
Instead of capitalising on the supposed gains from AI implementations by terminating staff that are now supposedly redundant, the smart way to use those gains is by realising that you now have a team that finally has TIME. Time in which to use their product expertise to fix the processes or help articles instead of having to muscle through “the long way” because nobody has ever had time to figure out another way. Time to coach or train up to be tier 2 / 3 / Jr. Dev employees . . . because there WILL be tickets that a chat bot cannot resolve. And as a call back to the above, it’s not like they won’t be busy, I still strongly believe that they should be reading EVERY WORD that the AI generates before it gets sent out. To check for accuracy and the human factor™.
In the interest of full transparency I have to disclose that I run my own working life this way. I have an AI assistant (”Zelda”, so named after the Librarian of the Neitherlands in “The Magicians”. I welcome your judgement) that I work with every single day. It has never once published anything of mine itself, and anything it HAS created has been to my specifications and I have rewritten just about every word of it (and nobody who has ever worked with me is surprised by a single thing I just said. I’m told that I can be . . . particular . . . with my writing. Judgers) But to be clear, that’s not a failing on Zelda’s part, it’s this exact arrangement working as I’ve described. The draft gets me off of a blank page, which is exactly what I need. I couldn’t even imagine sending its words out as mine unread . . . and my inbox matters considerably less than our customers’ do.
So . . . I call this the Drafter Principle. The machines can own the first draft. But a human HAS to own everything after it. It’s what ensures that we, as humans, can maintain more of a handle on . . . well . . . us. And our future. It allows for an enormous amount of improvement: it creates more output, more consistency, dramatically less time staring at empty documents, and so much more; but it also allows for the human beings that make up the fabric of these companies to continue to exist, and maybe even with a better quality of life.
Don’t make me Butlerian Jihad this thing. I WILL pull this car over.
The “should we use AI in support” question was settled a while ago, whether we like it or not. You are using it, or you are about to be. The only question with any juice left is where the human sits, and my answer is: at the end. Every. Single. Time.
AI drafts everything. A human sends everything.
That is the entire message. Everything else is implementation detail.