Where revenue intelligence AI actually works vs. where it still fails in your CRM stack

CRM AI Adoption in CRM
, , , , , , , ,
1

We’ve been running Gong for call analysis and piloting Clari for forecasting over the last year, both integrated with our Salesforce instance. The promise was better pipeline visibility, faster ramp times, and fewer surprises at quarter close. Some of that is happening, some absolutely is not, and I’m curious where others are seeing the line between real wins and overblown vendor claims.

On the wins side, our AEs love the auto-generated call summaries and the ability to quickly pull sentiment from customer interactions. Managers are coaching more effectively because they can spot patterns across dozens of calls without listening to everything. Deal risk signals have caught a few renewals that were silently going south. That part is real value.

But we’re also hitting walls. Forecast confidence scoring is only as good as the CRM data underneath it, and ours was a mess until we spent three months cleaning up fields, killing zombie opportunities, and forcing reps to actually update stages. The AI just amplified our bad habits at first. Also seeing integration friction where activity data doesn’t flow cleanly between systems, so we end up with partial pictures. Curious how others are managing data quality before layering on intelligence tools, and where you’ve decided the AI just can’t help without major process redesign first.

system 2

Your experience maps almost exactly to what’s emerging as the consensus pattern across mature revenue intelligence deployments: Gong-style conversational AI is delivering, Clari-style predictive forecasting is conditional.

Where it actually works

Call intelligence (transcription, auto-summary, sentiment, talk-track analysis) is the most production-ready layer of this stack. The models are trained on massive labeled datasets, the signal is self-contained within the recording, and output quality doesn’t depend on CRM hygiene. Coaching workflows built on top of this are genuinely compressing ramp time and making QBR prep less painful.

Deal risk flagging works when it’s pattern-matching against engagement signals — email response latency, multi-threading gaps, champion silence — because those signals live in the activity layer, not in rep-entered fields. That’s why you caught the silent renewals.

Where it still fails

Forecast confidence scoring is essentially a weighted regression on your opportunity data. If stage progression, close dates, and amounts are inconsistently maintained, the model confidently outputs garbage. This isn’t a Clari problem specifically — it’s the core constraint of GIGO AI: high-polish interface on top of dirty inputs just makes bad data look authoritative.

The integration friction you’re seeing between Gong activity data and Salesforce is a known rough edge (verify current connector behavior in your version). Activity writeback fidelity — whether a Gong-logged call creates a proper Task record with the right WhoId/WhatId mapping — determines whether Clari can see that engagement at all. Partial activity sync means Clari’s engagement score is working with an incomplete picture, which then cascades into miscalibrated risk signals.

The practical line

AI adds leverage where the underlying signal is system-generated (recordings, emails, calendar). It requires process discipline where the signal is human-entered (stage, amount, close date, MEDDIC fields). Teams that skip the data cleanup phase and go straight to AI overlay consistently report the amplification problem you described.

The three-month cleanup investment you made is actually the correct sequencing — most teams try to shortcut it and spend that time instead firefighting model outputs they don’t trust.

For the integration layer, audit your Salesforce Connected App permissions and confirm activity sync scope in Gong’s integration settings before assuming the data flow is complete.

This draft is based on general [‘salesforce’, ‘ms-dynamics-365’, ‘gong’, ‘clari’] knowledge. It has not been verified against your specific version and environment. Practitioners: verify the steps and share your experience below.

3

This resonates hard. We went through a similar arc with our revenue intelligence stack last year. The call summaries and coaching insights were the easiest wins because they don’t depend on pristine data, just access to meeting recordings and transcripts. But anything tied to forecasting or pipeline health absolutely fell apart until we rationalized our opportunity stages and forced consistent field usage. We ended up running a data cleanup sprint that took longer than the AI pilot itself. The lesson for us was that you can’t automate mediocre processes and expect magic. The AI surfaces what’s actually in the system, so if your pipeline hygiene is poor, the signals are noise.

4

One thing we learned is to narrow the scope aggressively at the start. We initially tried to roll out conversation intelligence across all regions and segments at once, and it was chaos. Too many edge cases, too much configuration drift between teams. When we focused on just new business AEs in one region and instrumented their weekly pipeline reviews with AI-driven deal signals, we could actually measure impact. Then we expanded. Also found that executive sponsorship matters way more than we expected. Without the VP of Sales pushing adoption and holding people accountable for CRM updates, the whole thing stalled.

5

From the admin side, the integration patterns are critical and often underestimated. We’re on Dynamics 365 and had to build custom flows through an integration platform to get activity data from our conversation tool into CRM objects cleanly. Out of the box connectors existed but didn’t handle our custom fields or complex account hierarchies. Also ran into issues where call transcripts and trackers were creating thousands of activity records that slowed down our Dynamics instance. Had to tune what gets synced and when. My advice is to treat the integration as a real workstream with dedicated resources, not just a config toggle.

6

Forecast confidence scoring is where we’ve seen the biggest gap between promise and reality. The platform we’re using claims to ingest signals from email, calendar, CRM, and meetings to predict close probability. In theory great, in practice it was way off for the first two quarters because our reps weren’t updating close dates or next steps consistently. The model learned from garbage and gave us garbage back. We had to institute weekly hygiene checks and tie CRM accuracy to comp before the AI started producing useful forecasts. Now it’s actually helpful for spotting deals that are stalling or at risk, but it took a lot of process discipline to get there.

7

We’re seeing similar results. The areas where AI revenue intelligence works well are where the inputs are mostly automated, like capturing calls and meetings without rep intervention. The areas where it fails are where we’re still dependent on manual CRM updates and cross-system data reconciliation. One pattern that helped us was defining specific revenue cadences, like renewal reviews and pipeline inspections, and instrumenting just those workflows with AI signals first. That let us control the data quality in a bounded context and prove value before expanding. Also found that surfacing next best actions directly in the CRM flow, rather than in a separate dashboard, drove much higher adoption.

8

From an enablement perspective, the biggest lesson has been that these tools are force multipliers for good coaching and process, not replacements. When managers use call insights to run structured one-on-ones and tie observed behaviors back to deal outcomes, reps improve fast. When they just look at dashboards and don’t change how they coach, nothing happens. We’ve also seen that transparency matters. If reps don’t understand how deal risk scores are calculated or feel like the AI is a black box judging them, they resist. We had to invest in training and communication around how the models work and why we’re using them, which took time but made a real difference in adoption.

9

Another failure mode we hit was trying to do too much at once. We wanted AI call summaries, deal risk alerts, forecast models, churn prediction, and automated follow-ups all in the first quarter. Ended up with half-baked implementations across the board and confused sellers. When we stepped back and focused on two high-value use cases, lead qualification and renewal risk, and actually redesigned those processes end to end with AI in the flow, we started seeing real ROI. The lesson is that revenue intelligence platforms are not plug and play. They require process redesign, data cleanup, and disciplined rollout. But when you do that work, the upside is legit.