The subtraction gate: Why AI agent deployments fail (and what to measure in the ones that work)
November 19, 2026

Overview: Successful AI agent deployments require more than a strong business case. Leaders need to define the agent’s role clearly, decide which existing processes will stop when it goes live, and measure performance across the agent, handoff, and human layers. Angela’s “subtraction gate” is the key discipline: if old meetings, spreadsheets, reports, and workflows continue in parallel, the promised efficiency and ROI can quickly disappear. Long-term value comes from pairing productivity gains with better precision, stronger frontline management, and clear measures of success.
In a recent conversation with GeekWire, I talked about why some AI agent deployments fail even when the technology works and the business case looks sound. The problem is often operational: organizations add the new workflow without deciding what old work should disappear, then struggle to prove that anything has actually changed.
Proving the ROI case for AI agents is only half the job. The deployments that quietly fail usually don’t fail because the technology can’t work. They fail because leaders didn’t decide what needed to stop, and they didn’t keep measuring the right things after launch.
If you want AI agents to become real members of your revenue team (not an experiment running alongside your “real process”), you need two disciplines: (1) plan the deployment like a hire, across the full revenue workflow, and (2) enforce a “subtraction gate” so the old way doesn’t survive in parallel. Then you need a measurement framework that makes it obvious at the agent layer, the handoff layer, and the human layer whether you’re actually improving outcomes.
As we explore how to measure the ROI of AI agents, proving the business case is only half the job.
One of the most common mistakes leaders make is treating an AI agent as a deployment. That framing pushes teams toward tools, configuration, and “launch plans,” while skipping the hard work of defining the role the agent is meant to play inside the revenue team.
The better approach is to think about the agent like a new hire:
When leaders define the role clearly, the agent becomes easier to integrate into the operating system of the go-to-market team. It stops being an experiment and starts being part of how revenue work actually gets done.
Just as important: plan for the full revenue workflow. Too many teams over-index on outbound prospecting and miss where agents can support the rest of the revenue journey: opportunity management, deal progression, and even retention and the broader customer experience. If you design only for the top of the funnel, you’ll get localized efficiency but miss the bigger outcome: a system that improves how revenue teams operate end to end.
Defining the agent’s role is only the beginning. Explore a practical five-step approach to implementing AI agents across RevOps workflows, from planning and governance through rollout and measurement.
Here’s the operational concept most teams skip, and it’s the reason deployments can fail:
The subtraction gate is the discipline of deciding, before launch, exactly which process, meeting, spreadsheet, or report gets retired when an agent goes live.
Most organizations spend real time redesigning workflows: where agents fit, what they do, what good output looks like, and what efficiency or effectiveness gains they expect. What they don’t do is plan the stop.
I’ve seen more deployments fail because teams didn’t intentionally decommission what the agent was supposed to replace.
When nothing stops, the old work persists. The team runs the new agent workflow and the old workflow, which means the deployment can’t show value. You can’t claim time savings if you keep doing the original work. You can’t claim better outcomes if you keep reinforcing the old habits and reporting lines. The agent becomes an additive layer of complexity instead of a transformation lever.
Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, largely because they fail to prove their value.
Decommissioning often fails for a predictable reason: leaders treat it as an ops task rather than an executive communication task.
A CRO may delegate decommissioning to RevOps: ending a meeting, retiring a spreadsheet, removing a report. But if the business executive doesn’t communicate clearly to the sales organization what is ending, when it’s ending, and what is replacing it, the “old system” stays alive.
And it’s not usually because people are resisting change in principle. It’s because pressure rises (it always does). When things get tight, frontline managers and teams relapse to familiar tools and rituals, especially the shadow spreadsheets that feel safe in the moment. If those artifacts aren’t explicitly retired, they become the fallback, and the agent becomes optional. That’s how a deployment fails without anyone ever declaring it a failure.
The subtraction gate has to be outlined at the beginning and reinforced after go-live. It’s the executive’s responsibility to make the stop real and to ensure it doesn’t creep back in.
If you want to know whether an AI agent is working, you need a measurement system that matches how value is actually created. I recommend tracking metrics across three distinct layers:
At the agent layer, you’re measuring whether the agent is operating and operating at meaningful scale.
Look at metrics like:
This layer answers: Is it running? Is it running where it matters?
The handoff layer is where most teams learn the most and where they often misdiagnose problems.
If agents are producing outputs like emails, research, meeting preparation, and summaries, track what percentage of that output is actually getting used.
The key interpretation: if the acceptance rate is low, you likely have a quality problem, not an adoption problem. When teams label that as “adoption,” they tend to push enablement and compliance. But if the output isn’t good enough, no amount of training will make people use it. Fix the quality first.
This layer answers: Is the output trusted enough to be used in real work?
The human layer is not optional, and it shouldn’t be reinvented. The conversion metrics sales leaders already track still matter:
Those metrics don’t go away when you introduce an agent. What should change is the expectation: as agent precision and effectiveness increase, those human-layer metrics should improve over time.
This layer answers: Are we actually improving outcomes, not just activity?
A major reason AI initiatives struggle to prove ROI is that leaders over-focus on productivity gains and doing work faster without connecting those gains to better outcomes.
Time efficiency can be deceptively hard to “cash.” I often compare it to getting a raise: you expect to feel the difference immediately, then you look up six months later and wonder where it went. Time works the same way. If you do low-value activities more quickly, you’re still doing low-value activities, and you won’t see the ROI you hoped for.
When I think about ROI, I separate it into two layers:
Precision is what sustains the case. It’s the reason your win rate, connect rate, and conversion metrics should trend upward as the agent improves. When you combine productivity and precision in a disciplined scorecard, you can get to unit economics: what each pipeline dollar costs you, what each closed-won dollar costs you, and whether the new operating model makes business sense.
Many leaders can’t confidently answer whether their agent strategy is working because they never defined success in a measurable way upfront. The result is a “hand-wavy” program. People feel busy, outputs exist, but nobody can point to a shared definition of “good,” a clear scorecard, or an agreed set of thresholds that indicate progress or risk.
That uncertainty makes deployments inefficient. If you don’t know where you’re going, you can’t manage the system toward an outcome. And if you can’t define what “good” looks like, you can’t tell the organization what to stop, and you can’t protect the value you’re trying to create.
The goal is not to measure everything. The goal is to define a small set of metrics across the agent, handoff, and human layers that creates confidence: we know what success looks like, we know how we’ll track it, and we know how we’ll respond when a number moves in the wrong direction.
Sustained success doesn’t come from the redesign alone. It comes from leadership on the front lines.
I’ll say it plainly: the frontline manager is the control point.
Organizations often redesign the workflow, launch agents, and even define success. Then they fail to build a leadership framework that ensures the new process holds. Without a leadership rhythm like inspection, coaching, and reinforcement at the frontline level, the system will drift. Under pressure, the team will revert to old behaviors, old tracking mechanisms, and old shortcuts.
If you want momentum to last, invest in frontline leadership as part of the deployment plan, not as an afterthought.
Everyone worries about dirty data, and it’s a legitimate concern. But there’s a major opportunity teams miss. The cleanest data is the data a human never has to touch.
For sustained success, lean into conversational intelligence and agent-driven capture of what’s said in real customer interactions. Don’t rely on people to manually retype critical context into Salesforce or other systems. Pull it directly from their words, structure it, and populate it automatically. That creates a “gold context layer” that improves the reliability of downstream workflows and supports agent performance over time.
The organizations that sustain success don’t treat AI as a headcount reduction exercise. They redesign the job.
As a leader, you have to decide what the role becomes when agents take on meaningful portions of the workflow. If you don’t redesign the job, you’ll end up with mismatched expectations, unclear accountability, and a team that’s unsure how to spend the capacity the agents created.
I expect the market to go through a correction. Many organizations moved fast, treated the space like the Wild West, and didn’t build enough governance. They didn’t define what would stop. They didn’t define what metrics would prove success. That will create a lull, a moment where teams realize they launched technology without building an operating model.
Then we’ll see a rebound: more thoughtful adoption, better governance, clearer success definitions, and more disciplined deployments that actually hold up over time.