Eighteen thousand sellers, two agents, no human in the qualifying loop
Siemens says Agentforce now touches every one of its roughly 2,500 unqualified inbound leads a month, across 132 countries, feeding a sales force of 18,000 people. That comes from Salesforce's September 15 announcement, and it is a genuinely large number for a technology most companies are still running as a pilot in one region or one product line. Two agents do the work. One writes the personalized outreach a lead gets. The other decides whether that lead is qualified and which of the 18,000 sellers should own it.
The second agent is the one that should keep a sales operations leader up at night. Deciding who gets a lead used to sit with a sales development rep who could hear a hesitation or recognize a company name before sending it to the right person. Now that judgment sits inside a model making the call across a lead volume no team that size could review by hand even once.
At a hundred leads a month, a bad routing call gets caught because someone on the floor notices a pattern and mentions it in a Monday meeting. At 2,500 leads a month spread across 132 countries and however many languages Siemens sells in, that Monday meeting stops working as a control. Nobody sees enough of the queue at once to notice a pattern forming.
The incentive to move fast is real
Salesforce's release pairs the deployment with a number that has nothing to do with Siemens specifically. General industry analysis cited in the announcement puts aftermarket service business at roughly six times the growth rate of new equipment sales and about four times the margin. That is context the release borrows from the wider industrial sector, not a result Siemens has reported about its own book of business, but it explains why a company with Siemens' footprint would rather respond to every inbound lead than risk losing a service contract to inbox fatigue.
Marc Benioff's line on the deal was direct: "Siemens and Salesforce are showing what it means to become an agentic enterprise at industrial scale." Siemens AG President and CEO Roland Busch framed it from the product side: "By embedding our digital twin into the commercial workflow, we are putting a virtual engineer in the hands of service technicians and salespersons." Both quotes describe ambition. Neither describes what happens when the qualifying agent gets a lead that does not match its training, a distributor asking about a product line Siemens stopped selling in that market, say, or a buyer whose procurement process does not resemble the ones the model has seen most.
What misrouting looks like when nobody is watching
The failure mode worth worrying about rarely announces itself. A slow bias creeps in that nobody notices right away, the qualifying agent consistently underrating leads written in a language it handles less confidently, or funneling anything that resembles a large account to the same handful of top sellers while smaller but real opportunities sit in a queue. None of that trips an alarm. It shows up months later as a soft win rate in one region, and the first instinct is usually to blame the sellers or the market instead of the routing logic behind it.
Buying norms make this harder, not easier. A qualification model tuned mostly on how deals get done in North America and Western Europe is going to apply the wrong test to a market where procurement runs through a formal tender process, or where a first inquiry legitimately comes in informal and gets refined later. Static routing rules already struggle with this, as why lead routing rules stop making sense covers, and an agent making the same call at machine speed across 132 countries does not fix that problem. It just hides it behind a more confident sounding decision.
Quality control cannot be a person reading transcripts anymore
The old version of quality control was spot checking. A manager pulls ten calls or ten emails a week and reads them. That does not scale to an agent making a qualification and routing decision on every one of 2,500 monthly leads. What has to replace it looks closer to a design review for agent behavior done before launch, paired with ongoing statistical monitoring after launch, conversion rate and cycle time broken out by country and by language, checked against a baseline rather than watched only in aggregate.
It also means someone has to own the authority to stop the qualifying agent from routing anything while a suspected pattern gets investigated, which is a harder governance question than a press release lets on, since a pause on an agent touching every inbound lead has a real cost in missed pipeline the moment it gets flipped on. Siemens and Salesforce have not said what that escalation path looks like, only that the agent now handles all of the volume. The handoff between the agent and the seller who receives the lead is the other place this gets tested, because a seller who gets handed a bad routing call needs an easy way to send it back rather than working a lead that was never actually qualified.
Buyers evaluating something like this for their own sales org should also ask which edition and configuration of the platform the qualifying agent runs on, because governance differs by Agentforce edition in ways that determine what audit trail even exists to check later.
The number to ask for is not seller count
Eighteen thousand sellers and 132 countries make a strong press release headline, but they are not the number that tells you whether this deployment works the way Siemens says it does. The number that matters is routing accuracy audited by country and by language, tracked over time, with a defined threshold for when a person reviews the qualifying agent's decisions again. Salesforce's announcement does not include that number, and it is fair to assume Siemens tracks something like it internally even if it never becomes public.
Anyone else sizing up an Agentforce deployment at this scale should ask for that number before signing, not after. A readiness check before scaling an agent past pilot is the place to write down what routing accuracy actually means for your own leads, and who gets paged when it slips, well before the volume grows large enough that nobody can read the queue by hand anymore.
