Most revenue teams have conversation intelligence. Far fewer have a conversation intelligence program. The difference shows up in what happens after the call: does the data change how deals are run, or does it sit in a dashboard nobody opens?
Magic is a useful place to look for the answer. Their go-to-market team is roughly 140 people across 20+ countries, all selling into North America, and they were one of Airspeed’s earliest customers. Today they run around 75 custom insights across sales, support, customer success, and key accounts.
Alex Wilson oversees sales support, customer success, RevOps, sales enablement, and several other teams at Magic. He joined Adam Blech, Customer Success Lead at Airspeed, for our Revenue Execution Series, a monthly session where customers share how they actually use the platform. This post pulls together the playbook he walked through, from why keyword tracking falls short to how Magic runs pricing experiments, measures reps without telling them, and keeps 75 prompts from turning into a mess. You can watch the full session on demand.
Keyword tracking tells you what was said. Prompts tell you what it meant.
Alex has been using conversation intelligence since 2016, and he remembers the early version well: did the rep say the competitor’s name, yes or no? That approach has two problems.
The first is transcription. Accents, fast talkers, and slow talkers all trip up speech to text, so a keyword tracker is only as good as the transcript underneath it. The second problem is bigger. A keyword match can’t tell you why something came up.
Take the most basic tracker there is, a competitor mention. In practice, that one mention could mean very different things:
- The prospect is moving away from that competitor
- They are in an active buying cycle with that competitor right now
- They used that competitor years ago and have opinions about it
A keyword tracker flattens all three into a single “yes”. A prompt based custom insight reads the surrounding conversation and returns the context, which is the part you can actually act on. As Alex put it, “the keyword tells you, was it said, yes or no. The prompt tells you what it actually meant.”
That’s what Airspeed custom insights are built for. Teams define what they want to extract in plain language, from concrete data points like budget or competitors to softer signals like sentiment, buying intent, or how well a rep handled a step in the process. For starting points, see Airspeed’s inspiration for custom insights.
Start with call tags, not insights
Alex’s first piece of advice surprised a few people: before you go deep on insights, get your call tags right. Call tags are how Airspeed classifies a conversation, and they can trigger insights automatically.
The logic is simple. If the platform recognizes a call as a discovery call, it fires every discovery insight and nothing else. Because tags are prompt based too, Magic can classify a discovery call correctly whether it happens on Google Meet, Zoom, or the phone.
This matters more as your library grows. You wouldn’t want all 75 insights running on every discovery call, demo, onboarding session, and product feedback chat. Tagging lets you get surgical about which questions get asked of which conversations, and everything downstream (insights, scorecards, CRM fields, alerts) inherits that precision.
Try this: when you create an insight, the trigger sits at the top of the setup screen. Pick the call types it should run on before you write a single word of the prompt.
Three ways Magic uses custom insights
Alex frames every insight with one question: which team needs this, and what will they do with it? That’s why Magic’s insights serve executives, product, marketing, and operations, not just sales managers. In practice, they fall into three buckets.
| Use case | What it answers | Magic example |
|---|---|---|
| Lifecycle data and team behaviors | What are we learning about the deal, and is the team doing what each stage requires? | Exit criteria per stage, from discovery through the 90 day check-in; consistent use of the referral program |
| Deliberate experiments | Is the change happening, and what impact is it having? | Pricing experiments with control and variable groups, scoped to specific reps |
| Silent baselines | What is really going on when nobody knows they are being measured? | A two month, behind the scenes run of Magic’s own qualification framework |
1. Lifecycle data and team behaviors
The first bucket tracks the customer journey itself. What are we learning about this lead, and how is the deal evolving from discovery to closed won to the first 90 day check-in? The customer’s goals shift along the way, and insights capture how.
Alongside that sits behavior. “It’s one thing to set a playbook,” Alex said. “It’s another thing if it’s being done or not. And then if it’s being done, to what degree.” Every stage has exit criteria, and insights show whether they are being met consistently.
Company goals live here too. Magic’s referral program is a win for the client, the rep, and the business, but only if reps bring it up. A set of insights checks that it is being embedded across the lifecycle rather than mentioned once in a blue moon.
2. Deliberate experiments
When Alex joined Magic three and a half years ago, the culture was closer to “write the playbook and forget it”. He and the CEO have spent the time since building a habit of constant experimentation.
Pricing is a good example. Magic doesn’t publish pricing on its website and has tiers based on need, which gives it room to try different approaches. Over the past two years they have run several pricing experiments, each with a control group and a variable group. Insights answer three questions: are reps actually doing it, how are clients receiving it, and what happens to the deal afterwards?
The detail worth stealing: insights don’t have to roll out to the whole team. You can assign them to individual users or teams. Some of Magic’s experiments involved just three reps, and the insight only tracked those three. That keeps the data clean and avoids a flood of false negatives from people who were never part of the experiment.
3. Silent baselines
Tell a team they are being measured on something and their behavior shifts, for better or worse. Sometimes you need to know what is happening before anyone adjusts.
Magic’s most recent example is its own qualification framework, a cousin of BANT, CHAMP, and MEDDIC. They ran it quietly in the background for two months. Reps weren’t told, so the output gave an unbiased read on opportunity quality. Marketing then used that data to evaluate programs, without reps showing up differently to leads they had been told were “good” or “bad”.
The baseline also improved the insight itself. The first prompt for M (money: do they have budget, are we speaking to the budget holder, and how likely are they to buy) worked fine. Over time it evolved into a 1 to 5 range, which made opportunities far easier to compare. That only happened because someone kept going back to it.
Governance: how to manage 75 insights without losing the plot
Once you pass ten insights, things get messy fast. Magic has around 75 (Alex was quick to add that he’s not saying you should, just that they do). Three habits keep the library useful.
Organize with groups. Airspeed has a groups feature inside insights that isn’t set up by default. Magic uses it two ways: by lifecycle stage (discovery, demo, onboarding, and so on) and by program. All referral prompts, for example, live in one place.
Run every new insight through QA. Magic has a dedicated QA group that works as a holding pattern. QA means two things there: are we actually doing the behavior, and is the prompt itself returning the right data? The team reviews outputs, checks false negatives against the real calls, and only “graduates” an insight into its permanent group once it holds up.
Revisit every 30 days. “I don’t think insights are set and forget,” Alex said. In past tools he’d install a tracker and not look at it again for a year. At Magic, the team aims to improve each insight roughly every 30 days.
A tip for when the library gets big: keep a copy in a Google Sheet. It’s easier to reorder, merge, and delete there, and nothing beats a cross functional meeting where everyone can highlight and comment in one shared sheet. Magic has swung between too many insights and too few, and the sheet is how they condense and repurpose.
Deciding what to build: start with the calls
Magic didn’t brainstorm its insights in a vacuum. It started by listening. Alex suggests picking a successful deal and listening to every call in full, from discovery through demo and beyond, with three lenses:
- Skills and behaviors: what was the rep supposed to do here, and did they?
- Data: what information do we wish we’d captured from this call?
- Gaps: where are the blind spots in skill, behavior, or data that we keep running into?
Insight or scorecard?
After the June session on coaching scorecards, this is the question customers ask most. Magic’s rule of thumb: an insight measures a milestone or gathers data you don’t want reps typing in by hand (or want them to start from and refine in the CRM). It also covers blind spots you keep wishing you understood.
Insights can graduate, too. Once Magic trusts the data, knows the behavior is happening, and finds it measurable, it sometimes moves the insight into a scorecard to track it over a longer period, then deletes the original.
Make it cross functional
The real power of an insight is that it’s one to many. A single budget insight means one thing to a rep and something else entirely to marketing, who also want to know why the budget sits where it does and what the lead spent before. Magic involves marketing, product, and operations when it designs insights, because those teams always have questions the calls can answer.
Adam sees the same pattern at Airspeed: reps and CSMs run the conversations, but product and marketing leaders need the aggregated picture across dozens of calls a week.
The “asked twice” rule
Alex also looks at how his teams use Airspeed’s AI chat. “If you ask the question more than twice, it probably deserves to be an agent, an insight, or a scorecard,” he said. Stop rewriting the question and get the answer proactively. Airspeed agents are built for exactly that hand-off.
Getting insights into the CRM (and keeping a human in the loop)
The bottom of the insight setup screen is where the data goes: a Slack alert, a CRM property, or both. Magic uses both.
Slack alerts that protect revenue. One insight flags whether a deal is, or could become, a multi person team deal. It sends a yellow alert to managers and up to Alex, so leadership can step in early whether or not the rep raised their hand. “That’s saved us not just time,” Alex said. “That saved us revenue.”
Sync modes that build a living record. When you connect an insight to HubSpot or Salesforce, you choose how it writes to the property:
| Sync mode | What it does | When Magic uses it |
|---|---|---|
| Append | Adds each new call’s output to what’s already there | SPICED fields: Situation from discovery, then the updated Situation from the demo, then onboarding |
| AI merge | Combines existing content with the new call into one summary | Complex insights that need to land in a single sentence property |
| Overwrite / don’t overwrite | Replaces the value, or leaves an existing one untouched | Simple fields where only the latest (or first) answer matters |
Append is the quiet hero here. Magic runs SPICED as its sales methodology, and appending turns a single CRM field into a record that grows with the deal.
Pre-fill, then verify. Magic hasn’t handed CRM hygiene entirely to AI. Insights pre-fill fields, and when a rep moves a deal to the next stage, the required fields pop up with Airspeed’s suggestions for them to confirm or edit. It’s faster than typing from scratch, and the data stays trustworthy.
Where data gathering is pure admin, though, Magic has removed the human entirely. During a lead source experiment, reps had to fill in three or four properties (interest level, a quality rating, and so on) after every attended call. They stopped doing it. So Magic used the answers reps had given to write insights that now fill those fields automatically. At three to four minutes a deal and 40 to 50 new deals a week, that’s roughly two to three hours of rep time back every week.
One more rule: Magic never connects a new insight straight to the CRM. It goes through QA, gets to V2, and only then syncs, so bad data never reaches the source of truth.
Insights as fuel for follow-up emails and agents
A well built insights program pays off in places you might not expect.
Personalized sequences at scale. At Magic, deals tend to stall after discovery, in the stretch before they reach the “not committed” pipeline stage. There’s a HubSpot sequence for exactly this moment, but reps were asked to personalize it and, unsurprisingly, didn’t.
So Magic built four insights that pull context from the discovery call and return it as a sentence or a short paragraph. Those go into HubSpot as properties, and the properties drop into email templates as variables. The result reads something like: “Hey Adam, it was great to connect last week. I remember you saying…” followed by the prospect’s own words. “There’s nothing better to win a deal than what the lead said versus what we said,” Alex pointed out. Single follow-up emails are easy to draft in Airspeed already; insights are what make a seven email sequence feel personal.
Faster, sharper agents. Airspeed agents can read raw transcripts, so they will answer “what are our most common objections?” either way. But if you have already been tracking objections through insights, the agent knows exactly where to look. The data is collected, organized, and still comes straight from the customer’s mouth, which makes it easy to aggregate across a quarter, a segment, or a vertical.
The same applies to stage drop-off. If you keep losing deals at one point in the process, an insight can surface future reasons to re-engage, or feed a win-back strategy, before anyone asks.
How to take your insights from V1 to V5
Most teams start with yes/no questions and picklists. That’s a fine place to begin. Alex’s advice is to improve what you have before adding more, because the jump from V1 to V3 is often where the real value sits.
Magic’s pricing insight shows how that evolution looks. V1 was a basic budget range. It grew into propensity to buy, then prior budget, which competitors the lead was evaluating, and, based on all of that, how likely they were to purchase.
The trick is to ask what action the insight should trigger. Once you think about the action, you start asking broader questions, and that’s usually when someone in the room says, “Hang on, we can get that?”
A checklist to run against your own library:
- Call tags are set up, and every insight triggers only on the call types it belongs to
- Each insight has a named owner team and a clear business action attached to it
- Insights are organized into groups by lifecycle stage and by program
- New insights sit in a QA group until the output is checked against real calls
- Experiments are scoped to the specific reps or teams involved
- Every insight gets reviewed at least every 30 days
- Insights only sync to the CRM after QA, with the right mode (append, AI merge, overwrite)
- Any question asked more than twice in AI chat becomes an insight, scorecard, or agent
- Insights that don’t drive an action get deleted
That last point is the one Alex returns to most. “If it’s just data for the sake of data, I don’t think it belongs.” The most powerful thing an insight can do is change behavior, skills, or pipeline. If it isn’t doing that, you can retire it.
The takeaway
Magic’s insights program works because every prompt earns its place. Each one is tied to a team, a stage, and an action; each one gets checked against real calls before it touches the CRM; and each one is revisited until it pulls its weight, or gets deleted. That’s the difference between collecting call data and running revenue execution on it.
The Revenue Execution Series runs monthly. Missed the June session on coaching scorecards? Watch the recording. Got a topic you’d like us to cover next, or want help reviewing your own insights? Reach out to your Airspeed customer success lead and we’ll find time.
Not an Airspeed customer yet? Start a 30 day trial and set up your first custom insights on your own calls, or book a demo to see them run first.
Frequently asked questions
What are custom insights in conversation intelligence?
Custom insights are questions you define once and run automatically across every relevant call. In Airspeed they are prompt based, so they extract context (pain points, competitors, budget, buying criteria, sentiment) rather than just flagging whether a keyword was said.
How is prompt based AI call analysis different from keyword tracking?
Keyword tracking tells you whether a word appeared in the transcript. Prompt based analysis reads the surrounding conversation and tells you what it meant, for example whether a competitor mention signals a switch, an active evaluation, or past use. It is also less dependent on perfect transcription.
How many custom insights should a sales team have?
There is no magic number. Magic runs around 75 across sales, support, CS, and key accounts, but only because each one is tied to an action and reviewed regularly. Start small, improve each insight before adding new ones, and delete anything that does not drive a decision.
What is a silent baseline?
A silent baseline is an insight that runs without the team knowing they are being measured. It shows how reps behave before awareness changes anything, which makes it useful for qualification frameworks and lead quality analysis.
Should insights write directly to the CRM?
Only once they have been checked. Magic runs every new insight through QA and a second version before syncing to HubSpot. For fields reps own, it pre-fills and asks reps to confirm at stage changes. For pure admin data, it fills the field automatically.
When should an insight become a scorecard?
When you trust the data, the behavior is happening, and you want to track it over a longer period, move it to a coaching scorecard.
