I was on a call a couple of weeks ago with a potential client who was just about to add half a dozen required fields to the opportunity/deal close process. A sales rep would have to fill them out before closing a deal so that the post-sales teams got the right context about the customer, what product features they were interested in, etc. Pretty standard. And if you were asking me 5 years ago, I would have told you that’s absolutely the right approach. Now, I don’t approach it that way at all.
Why I changed my mind
As a RevOps person, the data I’ve always been on the lookout for is structured: drop-down menus, numbers, dates, the stuff I can use in a report. And what AI and all of the tooling we have now changed is that unstructured data (free form text, audio, and video) is usable. So the first conversation I get into with almost anybody at this point is, what are you doing with your transcripts? Because that directly applies to how your team is interacting with the CRM and the go-to-market process in general.
Transcript data from sales, onboarding, customer success and support, and even internal meetings is the richest source of data we’ve ever had by quite a lot. A record in your CRM has maybe 40 or 50 key fields that you need filled out throughout the process (and that’s in a mature process, earlier stages you probably have less). Hopefully, a lot of those aren’t filled in manually. But now, you can probably fill out hundreds of different fields just with the transcript data if you’re using it correctly, and that requires absolutely no interaction from the sales/support/CS/etc rep in the CRM. It’s more about creating the right system of extracting the data from this new, very rich source of information that you weren’t using before.
The transcript is also more accurate than the rep
Let’s say a sales rep gets off a demo call and they forget to fill out some of the key areas in the CRM after the call, super common. They’re in back-to-back meetings all day in a lot of cases, they get to the end of the day, they do their CRM tasks of updating all of the records they’re supposed to update, and it’s all from memory. They may have taken some notes, and some sales reps are better at taking notes than others, but they’re still doing it from memory.
And human memory in particular is extremely poor. You can dive into all of the studies about eyewitnesses in the legal system. Somebody is absolutely sure it was this person and they’re way off, and a completely fabricated memory that’s clear in their mind has no basis in reality. That applies to all of our businesses and how we’re interacting with the data we’re entering into our systems. So if you build the right process of extracting the data from the transcript and directly filling it into the CRM (further down for how), or any other source of data you’re trying to collect, you’re getting more accurate information, you’re getting more rich information, and it’s much more usable further downstream.
What I told them to do instead
Okay, so there are five or six things the onboarding team needs to move forward with that new customer. Have AI scrape the transcript and put that into whatever onboarding ticket gets created at the close of the sale. One important point is to still make the fields required. For two reasons:
- So they’re still presented to the sales rep when closing and they can verify the information. Don’t underestimate AI’s ability to massively screw things up. That can be true and it still be helpful 95% of the time.
- In case AI doesn’t fill the field out due to lack of information in the transcript. In fact, part of the extraction instructions should allow the AI NOT to fill the field out if the information explicitly isn’t in there. That way the rep can either go get that via follow-up or they’ll have it from some other source than the meeting.
The information filled out by the extraction may not even be the most valuable part! Historically, onboarding or customer success may have asked the sales rep about a customer at some point in the post-sale process. Something that may not even have been in the extracted answers that’s relevant to retaining the customer. Now they can get those follow-up questions answered immediately from the transcript instead of relying on the rep’s memory.
How I actually build it
- Get the transcript somewhere you can reach it. I use the local transcription tools that run on your machine, not the bots that join calls, and I wrote a whole post about why. I personally use Notion’s but literally any will work, I just prefer the local ones to the ones that join calls. (Good chance I switch to WisprFlow’s when the Windows version comes out.)
- Write down the information needed throughout the sales and post-sale process that might be useful in the CRM. Highlight WHY it might be useful: this is the part people want to skip (I wrote about that too), but it’s the part that decides whether the extraction is any good. Also note that I said anything that “might be useful.” Operationally, this is opposed to the traditional advice of identifying what’s absolutely necessary (because more data to collect meant more admin time for the GTM team). Now you can be a lot more ambitious.
- Build the extraction in a no-code workflow builder. n8n and Make are the ones most people know at this point. I tend to use Gumloop. They’re all AI Zapier, basically. I prefer Gumloop’s interface, and I’ve seen it’s a little bit easier for non-technical people to pick up (so it’s easier to hand back off to a client to maintain after I build). In Gumloop, 99% of the time I just find the API documentation, dump it in, tell Gummie (Gumloop’s AI helper) I want this as the input and output, and it creates a custom workflow step for me that works.
- Write the output to the relevant CRM records (contact, opportunity, ticket, etc), and don’t require the rep to do anything.
- Still make the fields required to provide a human-in-the-loop step to review the output at critical points (such as SDR-to-AE handoff, closed deal to onboarding handoff, or onboarding-to-customer success).
Bottom line
If you’re about to add a required field so somebody downstream gets context, don’t. Pull the last 20 recorded demo calls, write the list of things onboarding keeps asking sales for, and check whether the answers are in the transcripts. They probably will be. Build the extraction for those, take the rep out of the equation, and end up with better data. The short version of this argument, plus the two tools I’d buy this month, is in issue 002 of the newsletter.
If you want a second set of eyes on which fields to pull and where to write them, set some time with me and bring the transcripts. If the process underneath is the real problem, the AI Effectiveness Audit is where I’d start.