.png?width=24&disable=upscale&auto=webp)
Much of my week is spent collaborating with product and data teams, focused on deploying revenue-generating real-time analytics and AI-driven functionality across revenue operations, where data quality is critical. Ironically, the underlying revenue data often presents a challenge to consistent revenue growth, working against these very initiatives. I usually get pulled in after someone has tried the straight-forward approach and discovered, again, that nothing about their GTM data is simple.
This conversation has occurred so often that it's now a clear and repetitive trend: A sales or product leader assembles the team and expresses the desire for AI to move beyond mere summarization. This sentiment is usually met with agreement, as everyone is weary of reports that simply reiterate what is already known. The crucial follow-up question then arises: If we grant AI the ability to take action, what specific actions is it permitted to take, and at what point in the workflow?
At that point, the focus shifts from models to systems. You’re no longer debating prompts or UI polish, but deciding whether you can trust an AI agent to observe what is actually happening, reason about what it means, and trigger the next best move across your revenue motion.
Over the last year, I’ve seen enough early adopters take a run at this that some clear lessons have emerged. The teams who get real value from agentic AI deployments tend to do a few things differently from everyone else.
A lot of strategy decks start with big visions: autonomous pipelines, self-healing forecasts, and AI-run territories. In practice, the early adopters who are actually winning pick something much smaller as their starting point.
One pattern I often see is that they start with inbound lead handling. Instead of completely overhauling the sales funnel, the immediate priority is to prevent the loss of high-intent leads due to slow routing or response times. They pick a handful of high-signal inputs such as intent level, account fit, recent activity, and market intelligence. Then, they give an agent permission to do a very specific job: decide where this lead goes and what is the best first step.
What this looks like in practice is the agent could only do three things at launch: route to the right SDR, trigger a first-touch sequence, and flag exceptions when the data does not fit their rules. No fancy branching. No multi-agent choreography. It is a simple win. The handoff requires no supervision, and marketing shifts its focus, no longer preoccupied with tracking the fate of individual leads every Monday.
IDC research sponsored by Outreach points in the same direction. In that work, 47% of organizations report stronger forecast accuracy after adopting revenue AI, and 45% say they can handle higher sales volume without adding headcount. These are manageable victories, the result of tightening specific workflows where automation can safely take over.
For SalesTech builders, I recommend the following: scope your first agentic feature around one workflow where the pain is obvious, the area of impact is limited, and the inputs are under your control. And most importantly, make sure to have a measurable ROI attached to it. Ambition and expansion can come later; reliability and trust has to come first.
Most SalesTech products have already done their first lap with generative AI. We’ve seen call summaries, deal risk narratives, coaching prompts, and highlight reels show up directly in the product experience. Sellers and managers generally like these features because they reduce documentation time and give them a quick read on recent developments.
Then the novelty often wears off, usage flattens, and nobody can point to a meaningful change in pipeline or revenue.
Here's the simple truth. A neat summary won't magically redirect a hot lead. A well-written risk analysis doesn't automatically book the next meeting, update the shared action plan, or pull in an executive when a deal is looking shaky. When your team is already stretched thin, giving them more insights without the means to act on them just adds to their mental load. It turns into a weight rather than an aid.
An example scenario involves a team running into exactly this problem. They launch polished meeting recaps and risk scores. Two quarters in, they realize the only folks that consistently read them are managers preparing for QBRs. Reps appreciate the summaries, but they still have to manually update fields, trigger sequences, and align stakeholders. Instead of being a teammate, the AI feels like another report.
Their second pass looks very different. Instead of asking what they could summarize, they ask what next action this insight should reliably trigger. For deal risk, that means an AI agent proposing (and, in some cases, auto-scheduling) a follow-up with the right stakeholder, or updating a mutual action plan template. The focus of the conversation shifts dramatically from appreciating a cool recap to recognizing genuine utility, with the realization that this actually saved a few steps.
You can see the same shift in how vendors talk about themselves. Gong, for example, positions its Revenue AI Operating System as a unified layer for visibility, intelligence, and automation that lets revenue teams orchestrate end-to-end workflows instead of just inspecting them. Their announcements explicitly call out use cases like faster account handoffs, more reliable forecasting, and orchestrated plays across the full customer lifecycle. That’s not an accident, but a response to customers asking for execution, not just analytics.
For SalesTech builders, the lesson here is to treat summaries and insights as table stakes, and not an ultimate goal. The true benefit emerges when these insights are integrated into the system of record, enabling them to drive actions that the user can either approve, decline, or execute autonomously.
When agentic projects stall, it’s almost never because the LLM can’t produce a decent answer, but because the product can’t give a consistent, timely view of reality due to poor data quality.
A single, authoritative source for revenue data is often difficult to find across many organizations. Various systems across revenue operations like the CRM, BI tools, and even individual VP spreadsheets often present conflicting truths. The introduction of AI Agents doesn’t inherently resolve this data disparity. Instead, they operate by selecting one of these existing versions of truth to drive subsequent decisions.
Clari Labs research that Salesloft has distributed makes this pretty blunt. Even with unprecedented investment in AI, the majority of enterprises (87% of those not utilizing Clari and Salesloft) fail to meet their revenue goals. A significant contributing factor is data fragmentation: 55% report inconsistent pipeline signals due to disconnected data sources, and 48% acknowledge their revenue data is not yet suitable for AI. Furthermore, 42% cite a lack of formal governance frameworks to ensure data consistency and accountability. This is not an isolated problem; it represents the foundational challenge most businesses face.
What this looks like in practice is a forecasting agent rolled out on top of exactly that kind of mess. CRM stages are not aligned with how deals are actually run, opportunity amounts differ between CRM and finance, and activity data arrives late. The model itself performs fine given its inputs, but the numbers it surfaces do not line up with the CRO’s trusted view. Before long, the feature is quietly deactivated, and AI forecasting becomes a joke in the next planning cycle.
The path forward usually looks similar. Teams roll things back, standardize stage definitions, fix data contracts between CRM and data warehouse, and put basic governance in place. Only then do they re-introduce forecasting agents. The second time around, the conversation changes because the inputs are finally under control.
For those building AI systems, the difficult reality is that simply applying AI won't resolve issues caused by conflicting systems. If you skip the data readiness and governance work, your agents become amplifiers for misalignment. The lesson from early adopters is to treat data contracts, governance, and a shared definition of trusted metrics as prerequisites.
Lesson 4: The agent loop only works if your signals arrive on time
If you ignore the buzzwords, most agentic AI patterns in revenue systems are just a loop: observe, decide, act, learn. More than the underlying model, the teams successfully implementing this loop focus intensely on the speed at which real-world data is reflected within the system.
The loop looks something like this:
Continuous signals in: CRM updates, email and engagement data, conversation intelligence outputs, web intent, and product usage events.
Context stitching: combining those signals into a coherent picture of an account or opportunity instead of treating each event in isolation.
Decisions while the window is open: surfacing or executing actions before the buyer has already moved on.
Feedback: capturing what the agent did and what happened next, then folding that back into future decisions.
Successful early adopters of Agentic AI share a simple, yet crucial, habit: they relentlessly focus on reducing the latency between an event occurring and the AI agent's ability to observe and act upon it. Allowing hours to pass before an inbound signal (like a customer hand-raise) is visible to your system means you're engaging in sophisticated post-mortem analysis, not true Agentic AI.
Outreach’s work with IDC on agentic AI in revenue intelligence reinforces this. They highlight that revenue teams are using agentic AI for use cases like automated lead scoring, deal risk detection, forecasting, and coaching, and that the gains in forecast accuracy and conversion rates come when those agents operate on unified, real-time data instead of scattered signals. In other words, the loop is only as good as its freshest input.
For Agentic AI to move beyond mere reporting and become integral to workflows, the underlying data platform is a critical, product-defining choice. Infrastructure must be capable of continuous ingestion, maintaining queryability under heavy load, and delivering interactive-speed responses. This level of performance is essential, whether the application is in-product analytics, next-best-action prompts, or continuously auto-updating forecasts. Ultimately, any design effort will be undermined by slow or fragmented data.
SingleStore fits into this picture as the foundation that keeps the loop honest. It is designed for real-time, interactive applications where analytics is a core part of the product experience. By consolidating all events into a single database, you gain a practical advantage: high-concurrency querying and rapid data retrieval, ensuring your AI agents operate on real-time, up-to-the-minute information, instead of stale or conflicting data. Furthermore, native vector support and Approximate Nearest Neighbor (ANN) search capabilities bring AI retrieval and similarity functions directly alongside your existing operational and analytical data (learn more here).
For those building Sales Technology, it's clear that latency and data fragmentation are not just infrastructure challenges. They directly impact the user experience. An agent's perceived intelligence, even with the best prompt engineering, is fundamentally limited if it doesn't have access to real-time, fresh signals.
Lesson 5: Teams earn autonomy; they do not start there
Everyone gets excited about the idea of fully autonomous systems. The early adopters who are actually running agents in production tend to take a pragmatic approach.
I've found the distinction between copilot and autonomous approaches to be a valuable mental framework. Copilot patterns are where AI suggests actions and humans remain the orchestrators: recommended next steps, draft emails, proposed forecast adjustments. Autonomous patterns are where agents own steps end to end: automatically routing leads, updating fields, triggering playbooks, or rebalancing territories with minimal human intervention.
The ideal approach to autonomy treats it as something you earn. Start with copilot-style workflows in a narrow scope, instrument behavior, listen carefully when users push back, and only promote a workflow to higher autonomy once it has proven accurate, predictable, and well-understood results.
Don’t be the one org that rolls out an autonomous renewal playbook too early. The idea may be sound, but the data feeding it has to be fully trusted, and the guardrails precise. A few noisy misfires and the feature for anything AI-driven could take a hit. Have a baseline for why you are recommending this, and get RevOps and IT to agree on guardrails before turning autonomy back on for a smaller segment.
This matches what Clari Labs and IDC both point to in their own way: governance and growth have to move together. Organizations that succeed with agentic AI align CRO, RevOps, CIO, and data teams on which workflows are safe to automate, which signals must be trusted, and what good behavior looks like as autonomy increases.
For builders, the lesson is to design a trust ramp into your roadmap. Start with high-confidence, copilot-style workflows. Prove accuracy and usefulness. Then graduate the ones that earn trust into more autonomous patterns, with clear guardrails and shared ownership.
Conclusion: Build autonomy in layers, and let reliability be the glue
Agentic AI is moving revenue tech from “telling us what happened” to actually influencing what happens next. That shift is exciting. It is also unforgiving if your systems are not ready.
The early adopters who are pulling ahead are not necessarily the ones with the flashiest demos. They are the ones who:
Start with one painful workflow instead of trying to automate everything.
Treat insight as a means to execution, not an end in itself.
Fix conflicting truth before they blame the model.
Invest in real-time signals so the agent loop can actually run.
Earn autonomy over time, instead of flipping a “full AI” switch.
If you are early in your own journey, pick a single workflow where the ROI is visible and the blast radius is contained if the agent gets it wrong. Use that as a forcing function to improve your data and architecture so every new workflow is easier than the last.
The compounding advantage comes less from the first agent you ship and more from how much faster you can ship, and trust, the ones that follow.
-for-Real-World-Machine-Learning_feature.png?height=187&disable=upscale&auto=webp)
-Search_feature.png?height=187&disable=upscale&auto=webp)
