Scaling an Agent From 1 to 100 Conversations a Day

You scale an AI sales agent's conversation volume the way you scale anything else that can quietly break if you rush it: in deliberate steps, checking a specific number before you add the next batch of conversations, not after. The jump from a pilot running five or ten conversations a day to one running a hundred is not simply more of the same traffic. Two 2026 studies of how AI agents behave over longer and more varied interactions found that an agent's reliability does not stay flat as conversations get longer or more complicated. It decays, in some cases sharply, well before the decay would show up in a simple daily total. Getting the first-response setup solid before you add volume is what keeps that decay from ever reaching a live lead.

Why does an AI agent's reliability drop as conversation volume climbs?

The conversation count itself is not the hazard. What more volume usually brings with it is the hazard: more conversations that run longer, branch into more back-and-forth steps, or wander into territory the original script never planned for. A 2026 study across nine models and four task families, Shubhra Mittal's "How Fast Do Agents Rot?", found that an agent's odds of finishing a multi-step task follow a geometric pattern set by one number: the odds it gets any single step right. That number sits well below perfect even for the strongest models tested, which means the odds of getting every step right in a row keep shrinking as the number of steps grows, with no point where the decline levels off on its own. On the paper's own agentic tool-use task, every model fell from near-perfect success to close to zero within sixteen steps, and the paper's projected reliability on a hundred-step, production-length task dropped to 0.24 from 0.42 at the shorter length most public benchmarks actually test.

What the sixteen-step finding means for a sales conversation

A conversation that stays simple, confirming a time, sending a link, answering one factual question, rarely gets anywhere near sixteen steps, and this risk barely touches it. A conversation that drifts into negotiating terms, handling a real objection, checking a pricing exception, and looping back twice to confirm details does get there, and that is exactly the kind of conversation a sales team starts routing to the agent once it moves past the easy leads in a pilot and begins sending harder ones through at scale. The useful response is not avoiding complexity altogether. It is counting steps per conversation type as volume grows and watching the harder categories on their own line, since one blended success number across all conversation types will hide the category that is already falling apart behind the categories that are not.

Does conversation complexity matter more than conversation count?

Yes, and a second 2026 paper measured how much more. Khanal, Tao, and Zhou's reliability framework ran ten models across 23,392 episodes on 396 tasks split into four duration buckets and three work domains, scoring each model on what the authors call a Graceful Degradation Score: how much of its short-task reliability a model keeps once the task runs longer. The score depended heavily on the kind of work being measured, not just how long it ran. For a software-engineering-style task, the score dropped from 0.90 down to 0.44 as duration grew. For a document-processing-style task, closer to what an agent does when it pulls an account detail or checks a stored contract term, the score barely moved, from 0.74 to 0.71.

Why a service-style conversation decays faster than a document-style one

The gap between those two numbers is a clue, not just a figure to file away. A document-processing task has a fixed shape: the information already exists somewhere, and the agent's job is to find it and report it correctly. A service conversation does not have that fixed shape. It branches based on what the person on the other end says next, and every branch is a fresh chance for an earlier small mistake to compound instead of staying contained to the one step where it happened. A sales conversation sits closer to the branching end of that spectrum than the fixed-lookup end, which is the paper's own warning in plainer terms: a volume increase aimed at harder, open-ended conversational work degrades faster than the same increase in volume aimed at simple lookups, even when both get called "more conversations" on the same weekly report.

The two numbers worth writing down before you scale

• 0.90 to 0.44: the drop in graceful-degradation score for the task type closest to open-ended, branching conversation as duration grew, in Khanal, Tao, and Zhou's study.
• 0.74 to 0.71: the same score for a fixed-shape lookup task over the same duration increase, which barely moved.

What should you change about the agent's script as volume grows?

Narrow the script's scope before you widen it. The instinct once volume climbs is to let the agent take on more kinds of conversations so it earns the investment faster, but the research above points the other way: keep the script's scope tight, let volume climb inside that scope first, and only then expand scope on purpose, checking the numbers again each time you do. Testing a script change properly matters more here than it did during the pilot, because a change that looks like an improvement on fifty conversations a day can hide a real problem that only surfaces once the agent is fielding the longer, harder conversations that come with the next fifty.

Split your success numbers by conversation type as you grow, not only as a lagging line in the monthly report. A booking confirmation that takes two exchanges and a payment-plan negotiation that takes twelve are not the same workload wearing different clothes, and blending both into one weekly success rate is exactly how a real problem in the second group hides behind a healthy-looking average pulled up by the first.

When does a human need to take over, and how do you decide?

Decide the escalation rule before volume forces the decision on you in the middle of a conversation that is already going wrong. The rule should name, in advance and in writing, the specific situations that hand off to a person: a dollar figure above a set threshold, any mention of cancellation or a formal complaint, a question the script was never written to answer, or simply more exchanges on one conversation than your own data says the agent reliably handles before it starts guessing. How long a deployment actually takes to get right includes the time spent building this rule and testing that it fires correctly on real conversations, not just the time spent writing the first draft of the script.

Building the escalation rule before the volume arrives, not after

Write the rule down while the pilot's small volume is still running, while a missed handoff only touches a handful of conversations and someone can review every one of them by hand that same day. Waiting until volume has already reached fifty or a hundred a day means the first real test of the rule happens at a scale where a mistake can reach dozens of people before anyone reviews the transcript.

What is a realistic way to stage the ramp from pilot to 100 a day?

Pick a volume step you can fully review by hand that same week, not a volume step that sounds impressive in a planning meeting. A jump from ten conversations a day to thirty, where someone can still read every transcript before the next step, is safer than a jump from ten to eighty that outruns anyone's ability to check it. Hold each step long enough to see the harder conversation types show up inside it, not only the easy ones that fill up a small pilot first, then raise the ceiling again once the numbers split by conversation type from the step you are currently on look steady.

Resist the shortcut of doubling volume on a fixed calendar regardless of what the numbers say. The research above points to conversation complexity, not raw count, as the thing that drives reliability down, so a calendar-based ramp that ignores the mix of easy and hard conversations can walk straight past the point where quality was already slipping. A ramp gated on the numbers from the step you are on, rather than on a date picked in advance, runs slower in the short term and costs considerably less than discovering the same problem once it reached a hundred conversations a day instead of thirty.

How many conversations can one agent actually scale to before adding more agents stops helping?

There is no fixed ceiling on conversation count by itself. The ceiling comes from how much of that growing volume is the harder, branching kind of conversation shown decaying fastest above, and from whether anyone is watching the split between easy and hard conversations closely enough to catch the moment it tips. Gartner's December 2025 sales forecast makes a version of the same point from the buying side of the market: it projects that AI agents will outnumber human sellers by a factor of ten by 2028, while fewer than 40% of sellers will say the agents they already have improved their productivity. Gartner's own sales analyst, Melissa Hilbert, put the reason plainly: beyond a certain point, more AI does not mean more productivity.

For a small sales team, the practical version of that warning is straightforward: running one agent well inside its tested limits on hard conversations beats running three agents each doing a mediocre job across the same mix of work. Adding a second agent, or widening the hours and channels the first one covers, should follow a deliberate decision to take on more scope, not a reflex to push more raw volume through the same unchanged script.

What should you check every week as volume increases?

Three numbers, reviewed together rather than filed away separately: the success rate split by conversation type, the count and the stated reason behind every human handoff, and the step count on whichever conversations ran the longest that week. A weekly review that tracks all three catches the moment a harder category starts slipping days before it would show up as a customer complaint, which is the entire point of reviewing weekly instead of waiting for a monthly rollup to flag it after the fact.

The staff time that added volume is meant to free up matters just as much as the agent's own numbers. The Bureau of Labor Statistics counted 2,666,000 customer service representatives employed nationally in 2025, at a median wage of $21.53 an hour, in a role the agency projects will shrink 5% by 2035 as this kind of work automates further. That is roughly the shape of the work an agent scaling past a hundred conversations a day is meant to be standing in for, which is also why the weekly review should ask not only whether the agent is keeping up with volume, but whether a person is still reading the handoffs closely enough to catch what the agent cannot.

If you want a second pair of eyes on where your own agent's scope and escalation rule stand before you add the next batch of volume, your free Systems Plan covers what to check before you change anything.

Methodology

The reliability-decay figures above come from two 2026 papers that measure how AI agents perform as a task gets longer or more complex, rather than how often they succeed on one short attempt: Shubhra Mittal's study of nine models across 10,664 analyzed trajectories (arXiv:2609.01660), and Khanal, Tao, and Zhou's study of ten models across 23,392 episodes and 396 tasks (arXiv:2603.29231). The sales-market figures are Gartner's own December 2025 sales forecast, including the analyst quote attributed to Melissa Hilbert. The labor figures are the US Bureau of Labor Statistics' published occupational data for customer service representatives. None of the numbers above are Conversion System client results; the escalation thresholds and the weekly review structure in this post are a starting point to adapt, not a fixed formula, since the right numbers for any one business depend on what its own agent is actually being asked to do.