Eighteen months of AI deployment recovered capacity almost everywhere. The pipeline lift it was supposed to produce mostly didn't show up. Four revenue leaders explain why the gap is a workflow problem, not a tooling one.

For a year and a half the pitch hasn't changed. Deploy AI, recover capacity, watch pipeline climb. The first two happened almost everywhere but the third one mostly didn't.

That gap is what I wanted to get into with Lindsay Rothlisberger (Head of GTM Innovation, Zapier), Lucy Alexander (Director of Agentic Prospecting, HubSpot), Davide Grieco (Head of Growth, Clay), and Elad Uzan (VP of Product, Lusha) on a recent panel. Lusha sponsored the session.

None of the four reframes they put on the table are really about AI. They're about the workflow infrastructure underneath, the metrics that start mattering once the tooling is live, and the operational discipline that separates a 10% pipeline lift from a number that just looks good in a board deck.

Here's what stood out.

Time saved per rep is the metric every AI vendor leads with. It looks good in the deck. It tells you almost nothing about whether the tool is touching revenue.

Lindsay's team at Zapier built around the time-saved number first, then corrected when the math stopped adding up.

"Right now, we're obsessing over recovered selling capacity. Did the rep get their week back, and did it show up in pipeline generated?"

Recovered capacity is the input. What the rep does with it is the signal that matters.

Lindsay layers in win rate on coached deals as a second indicator, a check that the recovered time is going toward work that closes. That's sharper than what most revenue teams run. It treats time saved as a leading indicator, valuable only when it converts.

Lucy runs the same logic from the marketing side at HubSpot.

"Our North Star is pipeline dollars generated by each rep each month. We wanted something more proximal to what we could measure."

She also walked through why she didn't pick closed dollars, even though that's the obvious revenue tie. Win rates get pushed around by pricing, by deal mix, by a separate work stream her team doesn't own. Too many variables to attribute cleanly back to her prospecting work.

Pipeline dollars per rep gives her a clean signal. Move that number and the work she controls is what moved it.

The logic underneath both metrics is identical. If your dashboard stops at the input, you're measuring whether the tool deployed. That's a far smaller question than whether the strategy is paying off.

Davide's take cut harder than most.

"When you scroll on LinkedIn, 99% of people are talking about the tools. The tools are not important at all. What's really important is how we design the campaign. That's never done by AI."

Clay sits at the center of the stack for thousands of revenue teams. Davide's position, from inside that company, is that the tool is the secondary layer.

The campaign design is the asset.

"In our case, it's us designing end-to-end, and then we use Clay to automate the hell out of it. But that's human thinking first, AI second."

He took the same argument to AI slop. The reason so much AI outreach reads as obviously generated isn't that the tools are bad. It's that the people running them outsourced the thinking, not just the execution.

"AI slop is not because the tools are intrinsically bad. It's that people are deferring the actual thinking about what good looks like, what's gonna resonate with my ICP, to the machines. As opposed to thinking themselves, and then using Clay or other tools to automate it."

Teams that bought tools first and tried to retrofit a strategy around them are rebuilding now. Teams that designed the workflow in plain language and handed AI the execution are compounding.

Davide is describing a shift from operator to architect. Reps and marketers used to write the messages and run the campaigns. Now they design the system and the AI runs the messages.

This is the theme the audience kept pulling at in the Q&A, and the deepest single insight of the session.

Both Lindsay and Lucy described their companies investing in what they call a context layer. It sits between the raw data in the CRM and the AI tools reps actually touch.

Lindsay broke Zapier's into three sub-layers.

The foundational layer is the stuff that rarely changes. ICP, use cases, company strategy. Ground truth every workflow references.

The playbook layer covers how reps engage at each stage and what their process looks like. It changes more often than the foundation, less often than the data.

The data layer is the live signal. Account history, conversations, what's working in market. It updates constantly.

"Those three layers of context are actually treated quite differently from a maintenance perspective, from how often they're utilized."

Lucy described HubSpot's version of the same architecture with a structural twist. A single engineering team owns the context layer. They maintain it as a database of insights and work with marketing and sales ops to keep it accurate.

Reps never touch the context layer directly. They use tools that read from it.

That separation is what makes it work. The tools change. The context layer stays canonical.

The other thing Lucy was clear about: context-layer work comes after validated use cases, not before.

"Two years prior to really focusing on the context layer at HubSpot were spent experimenting to prove which use cases context actually helped for the most."

That sequence matters. Teams building a context layer right now are often doing it backwards, standing up the central infrastructure before they know which use cases need it. Lucy and Lindsay did it the other way. Build small validated workflows first. Notice when the same context keeps showing up across them. Then centralize.

Lindsay called this the single highest-impact AI investment her team has shipped.

The workflow takes everything about a customer leading up to closed-won (conversations, emails, free-form text fields) and feeds it forward into the customer success handoff. Not as notes. As generated materials: the first kickoff deck, the account plan, the opening CS narrative.

The result was measurable improvement in customer time-to-value, on top of other work on the transition.

What made it work, in Lindsay's framing, was two patterns she now looks for in any AI investment.

First, whether the work involves unstructured data. Meeting recordings, customer conversations, emails, free-form fields, support tickets, "how did you hear about us" answers, closed-lost reasons. AI can read those as a new structured data point and surface them into a workflow.

Second, whether humans are currently playing middleware. Collecting things from different places, applying expertise, pinging someone else internally to tie the pieces together. Anywhere a human is doing connective-tissue work that wouldn't need to exist if the systems talked to each other, there's a strong AI workflow waiting underneath.

The sales-to-CS handoff hit both at once. A mountain of unstructured deal context, and a CSM doing connective-tissue work to turn it into a kickoff plan.

That's the playbook.

Lucy's HubSpot prioritization framework is the kind of detail that should make every marketing ops team rethink how they score leads.

The core split is one a lot of teams already use. Fit on one axis, intent on the other. Four quadrants. Each gets its own SLA on rep touches and timeline.

What's new is the data feeding the fit score.

"AI allows us to upgrade from static lead scoring to dynamic prioritization by analyzing factors like lead-gen forms and digital marketing readiness."

Her team uses Clay to run what they internally call the rep sniff test. Instead of pulling static enrichment fields from a vendor's database, they have AI look at the account's website the way a person would. Does the company run lead-gen forms? Is there evidence of digital marketing activity? Does the homepage suggest the prospect is at a stage where HubSpot fits?

These are fuzzy questions static enrichment can't answer. AI can.

The results settled it. The top 30% of dynamically prioritized accounts generated 95% of revenue. Reps who worked the prioritized list in order generated 10% more pipeline dollars per month than reps who didn't.

HubSpot's data scientists validated the lift with causal analysis. They controlled for tenure, prior attainment, and manager to isolate whether the tool was moving the number or whether better reps just happened to use it more.

The tool was moving the number.

A monthly score from a static enrichment vendor can't beat a score that updates on every signal AI can read. The real cost of static lead scoring isn't a wrong number on a spreadsheet. It's the AE hours burned on accounts the system would already have deprioritized.

Elad shared an experiment at Lusha that didn't pay off the way they expected.

They assumed reps would want fully generated outbound lists. Machine builds the list, machine writes the messaging, rep picks up the phone. End to end.

Reps rejected it.

The fix was structural. Reps didn't want less AI. They wanted to be the ones deciding what it optimized for.

"These are the signals I am curious about. These are the data points that I know are relevant for my ICP. And based on that, I want the model to take them into consideration and bring something hyper-personalized for me."

Same operator-to-architect shift Davide pointed at, applied to outbound. The rep isn't the recipient of an AI-generated list. The rep tells the model which signals matter, what counts as a strong fit, what kind of context unlocks a real conversation.

The model executes. The rep architects.

Anything else fails the trust test, and a tool reps don't trust doesn't get used no matter how good it is.

Davide had one of the sharper warnings of the session about where this conversation heads if nobody course-corrects.

The current narrative says every rep should build their own tools. Vibe-code their own outbound flows. Spin up custom dashboards. Compose their own workflows.

His prediction: this breaks.

"We're gonna give credits to everyone to code their own things. And then you have a hundred reps that have a hundred different processes. And then the numbers look bad, and you don't know what's happening because there's no framework to fix it."

The end state, he thinks, is the opposite of decentralized vibe-coding. Centralized teams (call them ops, GTM engineering, growth, whatever) build the tools. Decentralized reps use them.

Sellers and SDRs trying to moonlight as engineers means productivity goes down, not up. They're talking to Claude and shipping skills instead of picking up the phone.

Lindsay agreed and pushed the same point from the integration angle.

"I like to think that vibe coding random standalone apps is going to fade."

Her preferred model is workflow-embedded. A signal fires on an account. An agent researches, decides who to reach out to, drafts the content, and tees it up in the tool the rep already works in. The rep applies judgment and sends.

Standalone apps that live outside the workflow get abandoned. Embedded agents that act inside it stick.

If the panel's playbook compresses into something a head of revenue could run on Monday:

Measure pipeline generated, not just time recovered. Capacity that never surfaces in pipeline is a productivity stat.

Design the campaign first, then automate. The workflow you design is the asset. The tool runs it.

Replace static lead scoring with dynamic prioritization. The data to do it already sits in your stack. The question is whether you're using it.

Build a context layer, but only after you've validated which use cases need one. Three of the most ambitious AI shops in the space did it that way and don't regret it.

Embed agents into existing workflows. Standalone apps get abandoned. Triggered agents inside the tools reps already use will stick.

Keep judgment in the loop. AI scales execution. It doesn't scale taste.

When I asked which trends from the past 18 months will compound and which will fade, the answers lined up in interesting ways.

Lucy named data quality and freshness as the foundational investment teams underestimate. Her line on it was blunt: messy data ruins everything. The agents and automations only work if the data they read from is current and trustworthy.

She also predicted a structural shift in how revenue teams staff up. Strong individual contributors who can think deeply through workflows and solutions will get paid more. Middle-management layers will flatten. What's left for managers is coaching and judgment, not coordination.

Lindsay's prediction was about integration. The vibe-coded standalone app fades. The workflow-embedded agent that triggers on a signal and runs end to end sticks.

Davide's was structural too. The amateur seller-builder fades. The professional centralized GTM engineering team building tools that decentralized reps use sticks.

Elad's take was about infrastructure. The CRMs, the orchestration layers, the underlying data systems aren't going anywhere. The flashy demos sitting on top might rotate every six months. The infrastructure compounds.

He added one more thing I think is the underrated insight of the whole session. AI SDRs are held back by the inability to feed real-time context from rep conversations back into the model. The rep learns something on a call. By the time that learning reaches the AI's context, three more calls have happened. The latency is what kills the model's accuracy in practice.

Solve that latency and AI SDRs become real. Don't, and they stay a demo.

Elad closed with the line that's stuck with me since.

"Humans still want two things: trust in the process and the ability to focus more on creativity and selling, the parts of the work where they shine."

The companies that win the next 18 months won't be the ones with the most AI in the stack. They'll be the ones whose reps spend their reclaimed hours on work AI can't do, and whose leaders are tracking whether that's happening.

The hype cycle is over. The workflow problem is what's left, and it's where the next year of revenue gets won.

Want to get your brand in front of 60,000+ revenue leaders?

Partner with us and put your name at the center of the conversation.