Digital labour and the human-agent ratio

What is digital labour?

Digital labour is the habit of describing AI agents as a supply of work you provision, rather than as software you license. You do not buy ten seats of a tool, you add capacity to a task, and you talk about that capacity the way you talk about staffing.

The phrase came from a vendor. Salesforce launched Agentforce 2.0 on 17 December 2024 as the digital labor platform for enterprises, and Marc Benioff said in the announcement that it lets any company build a limitless workforce. The human-agent ratio is the number that travels with the framing: how many agents you run per employee. It appears in vendor decks and advisory posts, and there is no agreed definition sitting behind it.

Say plainly what the vocabulary is doing before you argue with it. Calling an agent labour moves the spend from the IT line to the operations line, where it gets compared with the cost of a person instead of the cost of a licence. That comparison is the point of the word. It is a budgeting and selling device first, and a measurement second, if at all.

What counts as one agent

The ratio falls apart on its first question. Take an order-intake flow that reads the mail, finds the customer, checks the price, creates the order and answers the sender. You can register that as one agent with five steps, or as five agents that call each other. Both are defensible, one of them doubles your ratio, and neither changes what the flow does. A scheduled script that has run since 2019 becomes an agent the week it moves onto an agent platform, without a line of its logic changing.

Your invoice will not settle it, because the platforms do not bill per agent. Microsoft's Copilot Studio documentation, updated in August 2026, says usage is billed in Copilot credits, charged for model tokens, for tools and knowledge, and for the harness itself. Salesforce sells Flex Credits at 500 dollars per 100,000, an Agentforce action costing 20 credits, or 2 dollars per conversation. UiPath licenses unattended robots per concurrent runtime, so ten robot definitions that never run at the same time need one runtime and cost what one robot costs. All three count work done rather than workers.

Three questions decide your ratio, and they belong on paper before you quote it: does a multi-step workflow count once or once per step, does a renamed scheduled script count, and does an agent that ran twice last quarter count?

Two companies, the same ratio

Two Belgian wholesalers, sixty office staff each, both handling around 1,400 sales orders a month. Both report one agent for every two people, so thirty agents. Both numbers are honest.

The first company. Their thirty are twenty-one scheduled scripts that ran for years and got renamed during a platform migration, plus nine assistants from a build week, of which two get opened in a normal month. Orders going through untouched went from 55 to 57 percent, and cost per order from 5.80 to 5.75 euro. That is 70 euro a month against a platform bill in the hundreds.

The second company. Their thirty are the steps of four flows that run every day: order intake from the mailbox, matching supplier confirmations, credit checks on new accounts, and chasing missing delivery notes. Untouched orders went from 55 to 78 percent, and cost per order from 5.80 to 4.10 euro. That is 1,400 times 1.70, so about 2,380 euro a month against roughly 900 euro of credits, a bit under 18,000 euro a year.

The ratio was identical and told you nothing about either company. Cost per order separated them in one line.

The half of the framing that earns its keep

Something real sits underneath the sales language, and you can take it without quoting a ratio. With per-seat software you decide in January how much capacity you own until December. With agents billed per action, capacity follows demand: you spend more in the week the year-end invoices land and less in July, with no purchase order in between. That is why a spend cap belongs in the design from day one.

It also changes what you scope. A licence project asks which department gets the tool. A capacity question asks which task runs short of hands in March, and whether that task is one an agent can do at all. The second question is the better place to start.

The questions the framing forces you to answer

Once you talk about agents as labour, the management questions that come with staff arrive too, and this is where the framing does its most useful work.

Who owns the output. An agent that sends a quote sends it in your company's name, so somebody has to be answerable for it by name. That is what an agent registry records: an owner, a purpose, a risk level and what the agent can reach. A system with no owner is the failure the registry is built to prevent.

What a unit of work costs. Take a supplier query answered in six Agentforce actions. At the list price of 500 dollars per 100,000 Flex Credits, an action is ten cents, so the conversation is about sixty cents. The person doing the same query in eight minutes at a loaded 45 euro an hour costs six euro. That reads as a tenfold gap until you add the cases that end with a person anyway: at one in five, the average becomes sixty cents plus a fifth of six euro, so about 1.80.

Who does the onboarding and the review. A new colleague gets shown the exceptions, corrected for a few weeks and checked afterwards. An agent needs the same under different names: test cases before it goes live, a weekly sample of its output read by a person, and a date when that sampling gets reviewed rather than quietly dropped.

What happens when it fails. Decide in advance which errors retry, which stop the flow, who gets called, and how the case reaches a person with its history attached. An agent that fails silently at three in the morning is worse than a step nobody automated, because nobody is waiting for it.

What is actually observable about jobs

Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab track US payroll records from ADP. In their update of 12 August 2026, covering data through June 2026, they state that they do not see widespread, economy-wide job displacement associated with AI. What they do see is concentrated: employment of workers aged 22 to 25 in the most AI-exposed occupations sits about 19 percent below where it would be had it kept pace with same-aged workers in less exposed occupations, a gap that widened from around 15 percent a year earlier, through reduced hiring rather than people being let go.

Marc Benioff said on a podcast in early September 2025 that Salesforce had taken customer support headcount from 9,000 to about 5,000, that he needed fewer heads, and that half of support interactions were then handled by agents. That is one firm, selling the product, describing its own operation.

For scale, Eurostat reported on 11 December 2025 that 20.0 percent of EU enterprises with ten or more employees used AI technologies in 2025, up from 13.5 percent in 2024. Set that next to Salesforce's survey of 200 HR leaders from 5 May 2025, where 80 percent expected humans and agents to be working together within five years: one of those numbers records what happened, the other records what executives expect.

Counting agents versus measuring work completed

Put the two side by side on one dimension: what each tells you about the business.

A count of agents tells you how many things exist on a platform, under a definition you chose and can change. It rises when you split a workflow, rename a script or leave a dead pilot on the list, and it falls when you consolidate, which is usually an improvement. It has no unit a finance person recognises and it cannot be compared with another company's.

Work completed is cost per case, cycle time and quality. Those move only when something real happens, they are the same numbers you used before agents existed, and they compare against last quarter. A business does not buy labour, it buys work done.

So if a vendor quotes you a human-agent ratio, ask what they counted, and then ask what happened to their cost per case.