AI budgets per employee: limits are not enough—measure completed work

Update, 20 October 2026: count experiments and separate credit pools

Jirka Herník’s 19 October avatar experiment includes a useful public cost breakdown: $4.81 for the final visual render and $6.08 for testing and tuning, totaling $10.89. Images and his own time are excluded. This is one account’s spending on one experiment, not HeyGen pricing or our benchmark.

For a content pipeline, that is more useful than quoting only the successful render. I would keep one job ID, every attempt beneath it and a clearly marked accepted variant. Voice spending already included in tuning must not be added again. The same accounting applies to dubbing, product videos and advertising variants.

Notion’s 19 October announcement adds a separate budgeting issue: Business and Enterprise receive 1,500 monthly Standard credits for Auto and 500 Select credits for manual model choices or Custom Agents. New customers move immediately; existing customers on 2 November. The post describes AI Command Center as coming soon.

My administrative recommendation is to separate everyday personal work from automated runs before the change. Included credits are not a cash cost per task. Check actual consumption per accepted result and your workspace’s rules; vendor routing savings do not guarantee the same reduction for your company.

Test what happens when a budget runs out. The expected outcome should be “awaiting owner decision,” not endless retries or silent switching to another paid account. Budget, quality and recovery behavior belong in the same operating report.

Update, 10 October 2026: agent runtime is not a billable hour

In AI s rozumem on 9 October, Filip Oborník raises the problem of pricing and compensation when people work with several agents in parallel. The public episode description explicitly leaves the question open. It is a useful reason to extend this article: tool consumption, human effort and the value delivered need separate records.

The following is my proposed pilot, not a measured result from the video. Pick one process for two weeks, such as incoming lead triage in n8n or Make. Give each enquiry one ID and attach every agent attempt, correction and human approval to it. A retry must not become another “completed task” in the report.

For each case, record direct API and tool spend, active human time spent briefing and reviewing, elapsed time to acceptance, and any reopening caused by an error. Two agents running concurrently for an hour do not represent two hours of human labour. Waiting time matters to service speed; labour cost should reflect actual human involvement.

A useful pilot calculation is:

Cost per accepted case = (API + tools + human handling cost + allocated setup and licence costs) / accepted cases.

Include failed attempts in the numerator. Allocate shared licences and setup using a rule agreed in advance; do not count the same spend once from the invoice and again from a token estimate. Track routing accuracy, corrections and turnaround alongside cost. Otherwise, an apparently cheaper workflow may simply hand more work to sales.

This is neither a client rate card nor an automatic employee ranking. First group enquiries by comparable difficulty. Agree project prices and compensation around scope, acceptance criteria and responsibility for corrections. Operational data should show whether the automation works, rather than reward the number of agents launched.

AI tools are moving from an interesting employee benefit to a normal operating cost. A Lupa.cz survey of Czech technology companies reports that Apify pays roughly CZK 3,000 per employee for AI tools and uses an internal cap of CZK 20,000 per developer per month. Some companies report much higher spending.

These figures are not a universal benchmark or an official price for “how much AI should cost.” They are reported values from one editorial survey. They are still a useful signal that monthly seat pricing is no longer enough as the only financial control.

From licences to variable consumption

GitHub announced usage-based billing for premium Copilot requests in April. Claude Code, Codex, model APIs, and agent platforms increasingly have similar economics.

Autocomplete is a short interaction. An agent that reads a repository, plans changes, runs tests, uses a browser, and iterates belongs to a different cost category. Using a strong model for architecture is not the same as using it for a routine purchase.

A company budget therefore needs to separate at least three layers: basic tool access, variable model and agent consumption, and the cost of integration, logs, validation, and human review.

Why a hard limit is not enough

A cap prevents an unlimited bill, but it can create two bad extremes. A low cap discourages people from using AI on work where it would genuinely help. A generous cap rewards activity even when it produces no accepted result.

A better metric is cost per completed and accepted task.

For coding agents, measure the cost of a change that passes tests and review. In support, measure a correctly classified and resolved ticket. For CRM enrichment, measure a record that passes validation. For document extraction, measure an invoice whose totals, supplier, and accounting checks are correct.

A cheaper model that needs three repair loops can cost more than a stronger model that finishes correctly on the first run.

Route work instead of banning strong models

I would not manage this with a rule that says strong models are forbidden. Define classes of work.

The cheap layer handles tagging, initial routing, simple JSON, deduplication, and normalisation. The main worker handles routine email, CRM enrichment, invoices, support, and content drafts. The strongest model handles expensive exceptions: a difficult bug, disputed contract, ambiguous document, or a plan that other agents will execute.

Track the model and effort, tokens and cache hits, tool calls and iterations, the reason for escalation, the validation result, and the cost of the finished task—not only the request.

Budget by risk and value

One monthly allowance for every role is easy to administer and often wrong. A developer running long agents has a different profile from someone summarising one meeting per day. Sales enrichment across hundreds of leads differs from legal analysis of a few sensitive contracts.

Set budgets by team and use case, then add a small personal experimentation allowance. People need room to test tools. Production workflows need an owner, metric, and audit trail.

Minimum company setup

Define approved tools and the data classes they may receive.
Set limits per team or project, not only one global cap.
Route tasks across cheap, main-worker, and specialist layers.
Give agents stop conditions, maximum steps, and task budgets.
Store the validation outcome, not only the token bill.
Regularly retire workflows that consume money without proving value.

My take

CZK 3,000 or CZK 20,000 is not the answer to how much a company should spend on AI. It is a useful starting point.

Good governance neither hands out unlimited accounts nor bans capable models. It separates routine work, main work, and expensive exceptions. It measures the cost of an accepted outcome, adds validation, and knows when an agent must stop for a person.

AI coding and automation are starting to look like cloud operations: limits are necessary, but without unit economics, observability, and responsibility, they are only numbers in an admin panel.

Sources: Lupa.cz: what AI costs Czech technology companies, Jirka Herník: what companies pay employees for AI, GitHub Blog: Copilot is moving to usage-based billing, and GitHub Docs: Copilot billing.