Digital Workers on the Org Chart Come With a Metered Salary

An article from Interrupt Agents, "Your 2026 Org Chart Will Include Digital Workers", describes a collision that HR and IT departments are hitting at the same moment. An agent is not a larger text box. It takes a goal, runs several steps, calls tools and stops when it judges the job done. Managing that has more in common with managing a fast, tireless junior colleague than with managing a licence seat.

The organisational argument

The article's central claim is that agentic work does not fit the usual split between buying software and hiring people. A tool sits in a licence inventory and gets renewed once a year. An agent needs an owner, a scope of authority, an escalation path and a way to reconstruct what it did last Tuesday afternoon. The article describes that responsibility landing on IT, HR and business leads at once, with nobody holding the entire picture.

The deployment tempo is the least difficult part of the problem. Standing up an agent has become routine inside most organisations that try it; deciding who answers for it has not. The technical build is a matter of configuration, while the accountability question has no vendor to call.

The gap between productivity and profit

The real gap sits between individual output and organisational profit, and it is worth reading as a sequence rather than a contradiction. Individual output rises first, because it requires no coordination with anyone. Money appears later, and only where the process around the individual changed. An agent that drafts a report faster, inside a review chain built for human speed, saves the drafting hours and nothing else.

What separates the two paths is the step that comes after the individual one. A faster draft handed back to the same review chain is absorbed by that chain, which was rarely the bottleneck to begin with. An agent wired into the next system changes who signs off, what gets checked and where the output enters the process that actually produces revenue. Redesigning the operating model around AI is that second piece of work, and it is a decision about the process rather than about the model.

There is a trust floor underneath. Managers being asked to sign off on an agent's output are being asked to extend a level of trust they would not hand a new hire in the first month. That gets reported as resistance. It is also a fair description of what agents currently do: fast, confident, and wrong in ways that look plausible until someone checks.

What one agent task costs

Cost is where the org-chart discussion becomes concrete. Agent runs are input-heavy, because every step re-sends the accumulated context. Take a modest workflow of 30 steps with an average context of 6,000 to 7,000 tokens per step. That lands around 200,000 input tokens and 40,000 output tokens for one completed task. That is not a large agent. It is a normal one.

At list API prices, the same task prices out like this. Gemini 4 Argon is missing from the table because Google has not published a per-token list price in the material available to us.

Vendor list prices as published on 19 October 2026, per million tokens; the per-task column applies the 200,000 input / 40,000 output assumption above.

The spread between cheapest and most expensive is a factor of twenty for identical work. Two models on the table ship open weights under MIT — DeepSeek-V4.1-Flash-1017 and GLM-5.3 — which matters for any European buyer who has to explain where the data went.

Scaling the bill

One task tells you very little. A thousand tasks a day, twenty working days a month, is a plausible volume for a mid-sized back-office workflow. On GPT-6.1 Sol that is $16,000 a month. On Grok 4.7 it is about $12,800. On Mistral Large 4 at the preview rate it is roughly $4,400. On DeepSeek-V4.1-Flash-1017 it is $2,160.

Twelve thousand dollars a month is a hire. Two thousand is a subscription. Both figures buy the same 30-step workflow, with error rates this table does not measure and cannot.

Prompt caching changes these numbers, sometimes substantially, and no list price above includes it. Neither do retries, failed runs, or the human hours spent checking what the agent produced. The table is the floor of the bill, not the invoice.

The European deadline that moved

Anyone in the EU putting an agent anywhere near employment decisions should know the compliance calendar changed. Employment, worker management and recruitment sit in Annex III of the AI Act, which makes systems used there high-risk. The Digital Omnibus on AI pushed the deadline for standalone Annex III systems from 2 August 2026 to 2 December 2027. Embedded high-risk products follow on 2 August 2028.

Sixteen months of extra runway is not a repeal. An agent that allocates tasks, scores performance or ranks candidates falls inside that scope, and the obligations arriving in December 2027 include logging, human oversight and conformity assessment. Building the audit trail while the agent is still small costs less than reconstructing it later from a pile of database rows.

Separately, obligations for general-purpose models under Chapter V of the AI Act have been directly enforceable by the EU AI Office since August 2026. Those bind the provider. If you build your own agent stack on top of someone else's model, the deployer duties are yours, and a vendor's compliance statement does not move them across.

What the org chart needs on the record

Workday's Agent System of Record, which folds agent capacity into headcount planning and skills inventories, is one answer to the ownership question: give digital workers a record the way human workers have one. For EU customers that record becomes part of an employment file, so retention periods, access rights and the legal basis for processing it are easier to settle before the first agent goes live than afterwards.

Three things are worth writing down before an agent gets a name. Its scope of authority, meaning what it may do without asking. Its escalation path, meaning who it asks. Its budget, in tokens or in currency, with a hard stop attached.

The cost figures above rest on one assumption you can check yourself in an afternoon: how many steps your actual workflow takes. Log the input and output token counts on ten real runs before signing anything. That is the number the procurement conversation should start from, and it is usually larger than the demo suggested.