On August 7, 2026, Salesforce published fourteen months of usage data from its own customers’ agents, and one number in it is more useful to you than the rest.
Between February 2025 and April 2026, the average organization in that cohort nearly tripled its active agents. The average agent went from acting on two skills to six. New agents went live in about two days, and that build time dropped 53% across the period. Agents handled 170 times more customer service chats than the prior year and solved seven in ten of them without a person. And through all of it, Salesforce reports, “escalation rates from AI agents to human agents are holding steady at 32%.”
Two caveats before that number does any work for you. Salesforce is reporting on its own platform, and the cohort only includes businesses that kept agents in production every single month of the analysis window, so these are the rollouts that survived. Take it as a directional benchmark from the healthy end of the distribution.
Even discounted, it says something a lot of vendor decks won’t. The fleet tripled, the agents got three times more capable, and a third of the work still went to a person. That share didn’t drift down as the systems matured.
Read the escalation rate. Don’t drive it down.
Most pilots treat escalation as leakage. The implied goal is zero, and the roadmap slide shows the line trending toward it.
We use the number differently. Escalation rate tells you what your agent actually covers, and its shape tells you whether anyone is paying attention.
A rate near zero on work with real consequences is a question, not a win. Either the task genuinely is that routine, or the cases that should have escalated went through anyway and the cost is sitting in a different report: reversals, credits, rework, an angry customer who called twice. A rate near 100% means the team stopped trusting the queue and is redoing the work by hand.
The useful signal lives in the breakdown. One workflow escalating 60% while another escalates 8% is a system that has learned where it’s strong. A flat rate across every task type usually means the escalation logic is one global confidence threshold, which is a setting rather than a design.
One question for your next vendor call
Ask what their escalation rate is, whether it’s broken out by task type, and which direction it’s moving. A vendor who can answer has instrumented the handoff. A vendor who can’t isn’t able to tell you what their agent is covering, no matter what the demo showed.
Worth noting that Salesforce’s own two figures don’t sum to 100%. Seven in ten chats resolved without a person, 32% escalated. Those are measured on different denominators, and yours will be too. Define the denominator before you quote the number.
The four questions we answer before an agent acts alone
We build agents custom to how a company already works, so the handoff design is specific every time. The procedure to get there isn’t. We answer four questions per task type, in this order.
1. What does it cost to undo? Reversibility sets the ceiling on autonomy, before accuracy enters the conversation. Drafting a reply, staging a record, flagging an exception: all cheap to reverse, so they can graduate early. Issuing a refund, releasing a shipment, sending a notice to a regulator, closing a ticket the customer can’t reopen: expensive or impossible to reverse, so they stay gated longer regardless of how good the model gets.
Cloudflare landed on the same axis from the infrastructure side. Its WriteGuard private beta, announced this month, classifies every agent tool call by what it does downstream, from read-only through contained write up to critical, and blocks the critical tier before the handler runs. Salesforce’s own Sophistication Index splits agent actions the same way, calling levels 1 through 3 (read, coordinate, synthesize) “standard, low-risk” and levels 4 and 5 (write, analyze, parse) high-complexity. When a hyperscaler and a platform vendor independently sort agent actions by whether they write, that’s a good sign the axis is real.
2. Do we have evidence yet, on this exact task type? Autonomy gets granted on data, and the data is your team’s own decisions. Fifty approvals on invoice coding tell you nothing about vendor onboarding. We count approvals, edits, and escalations per task type, and a task type graduates on its own record.
3. What does the gate cost in time? Every approval step buys safety with latency. Salesforce’s two-days-to-deploy figure says the constraint has moved: building the agent is no longer the slow part, and deciding what it’s allowed to do is. The rule we use is simple. Put a gate where the cost of reversing a wrong action exceeds the cost of waiting for a person. On a hold that delays a Friday load-out, a four-hour wait is cheap and a wrong release is expensive, so it stays gated. On an internal weekly summary, the wait costs more than the error does, so it goes autonomous early.
4. Who catches it, and what do they see? A handoff to an unowned queue is a dropped case with extra steps. Every escalation needs a named role receiving it and enough context to decide without redoing the work: the recommendation, the evidence behind it, what the agent was uncertain about, and the actions available right there. Reviewers who can’t reconstruct the reasoning start rubber-stamping, and then you have a review queue that produces approval data meaning nothing.
Three levels, applied per task type
The model we ship on every engagement has three levels, and the important word is per task type. One agent runs at all three at once.
Drafted. The agent prepares the work and a person acts on it. Day one behavior for everything. Every review banks evidence.
Approved. The agent prepares the complete action, including the record it will write, and a person authorizes it in one click. Faster than drafted, same authority boundary.
Autonomous. The agent acts and files a complete record, and people audit after the fact. Granted at a threshold the client sets, typically around 95% alignment, and reversible at any time.
Two measures decide promotion, and they aren’t the same thing. Confidence is the agent’s own score on a recommendation, derived from the context and history behind it. Alignment is how often your team agreed, counted from the approvals, edits, and escalations in the review queue. Confidence is the agent’s opinion of itself. Alignment is the only one with your business in it, and it’s the one that governs promotion.
On the compliance platform we built for Alliant, that record is what the model runs on: eleven governed agent workflows live in production, a 96% approval rate across 1,284 reviewed runs, and analyst output at 1.67 times the prior baseline. The approval rate isn’t a vanity metric. It’s the promotion evidence, and it exists because the review queue was instrumented from the first workflow rather than added after someone asked for numbers.
What never graduates
Some categories stay gated at any alignment score, because the boundary is a policy decision rather than a performance one:
- Anything that can’t be undone inside the business day
- Anything a regulator, auditor, or contract says a person decides
- Anything that would be the company’s first notice to a customer of a problem
- Anything where the agent is deciding about a person rather than a record
We write these down at kickoff and the client owns the list. Thresholds move, the carve-outs don’t move without a conversation.
What to do this week
The first two are worth doing whether or not you have anything running yet, and both get easier before there’s a system in place to argue with.
Sort your task list by reversal cost before you sort it by volume. The highest-volume task is the tempting first candidate for autonomy and often the wrong one, because volume and consequence are unrelated.
Write down the carve-outs while nothing is on fire. The list is much harder to agree on after an incident, and it’s the part a vendor will never write for you.
If you do have something running, get the escalation rate broken out by task type rather than in aggregate. One global number tells you almost nothing about where the agent has earned coverage, and if that’s all your tooling reports, it’s an instrumentation gap you can close in days.
Want a read on where your handoff should sit?
We build governed agents custom to how a company already works, and we design the handoff before the automation. Bring one process your team repeats every week and we'll walk it through the same exercise: which steps an agent runs on its own, which ones wait for your team, and what evidence would move a step from one column to the other.
Agent handoff design: next reading path
Frequently Asked Questions
What is a good escalation rate for an AI agent?
There isn't a single target. Salesforce reported (August 2026) that escalation rates from agents to humans held steady at 32% across its customer cohort from February 2025 to April 2026, which is a reasonable directional benchmark for customer service work. What matters more than the level is the breakdown: a healthy system shows different rates for different task types, reflecting where the agent has earned coverage. A rate near zero on consequential work usually means cases that should have escalated didn't.
Should the escalation rate go down over time?
Per task type, usually yes, as the agent accumulates evidence and earns a higher trust level. In aggregate, often no, because teams keep adding new task types that start at full human review. A flat or rising overall rate alongside falling per-task rates is a system that is expanding, which is a different story than a system that is failing.
What is the difference between confidence and alignment?
Confidence is the agent's own score on a given recommendation, derived from the context and history behind it. Alignment is how often your team agreed with the agent, measured from approvals, edits, and escalations in the review queue and tracked per task type. Confidence is self-reported. Alignment is the measure with your business in it, and it is the one that should govern whether a task type earns more autonomy.
How do you decide which tasks an agent can do without approval?
Four questions, in order. What does it cost to undo the action if it's wrong? Do we have enough of your team's own decisions on this exact task type to justify a promotion? Does the time cost of the approval gate exceed the cost of reversing an error? And who receives the handoff, with what context? Reversibility sets the ceiling, alignment data sets the timing, and the gate goes where reversal cost is higher than wait cost.
What should stay under human approval permanently?
Actions that can't be undone within the business day, actions a regulator or contract assigns to a person, anything that would be a customer's first notice of a problem, and any judgment about a person rather than a record. These are policy boundaries rather than performance ones, so a high alignment score doesn't change them. Write the list at kickoff and keep it owned by the client.