The virtual employee frame
Specify an automation as a role with six fields, inputs to budget, and predict the failure each missing field will cause before it ships.
On this page
An automation that does a job end to end is easier to specify as an employee than as software. Nobody hires a person with the brief "handles leads". A new hire gets a job description: what lands on their desk, what they produce, what finished looks like, who they report to, who they call when they are stuck, and what they may spend. An automation needs the same six fields, and one that ships without any of them fails in exactly the way that field would have prevented.
The brief that fails
"Set up an agent that handles our leads." Built literally, that is the agent from scripts, workflows and agents: three tools, a goal, the agency's 400 leads a month, about $0.06 a lead and $24 a month on a normal run. Three things go wrong in the first month, and none of them is visible from the brief.
Nobody reads its output. Four hundred replies go out. The first wrong one, a reply offering a service the agency stopped selling in the spring because the system prompt still lists it, is found three weeks later, when the prospect forwards it to the founder and asks for the price. Every reply sent in between was unread, because nobody was named to read them.
Nothing stops it at $400 a month. One Monday the CRM's email field is renamed. Every find_customer call now returns nothing, and the agent tries the domain, the company name, spellings of the name, and gives up after about 50 turns. By the last lesson's per-turn arithmetic that is 480,000 tokens in and 5,000 out, about $1.00 a lead: $400 a month where $24 was expected, noticed by whoever opens the invoice.
Nobody knows it is stuck. A lead that runs to 50 turns ends with no draft and no message to anyone. The brief promised no response time, so nothing is late and nothing alerts, until a prospect rings to ask why nobody answered.
The model did nothing unusual in any of the three. Each failure is a missing line in the brief.
The six fields
A role is six fields, each written so that someone other than its author can check it.
- Inputs. What arrives, from where, in what shape, at what volume, and what it refuses.
- Outputs. What it produces and where it writes, and what it may never do: send money, quote a price, delete a record.
- Definition of done. A state a person can check for one item by a stated time. Never "handled".
- Manager. A named person who reads a sample of the output on a fixed cadence and owns every change to the prompt and the gates.
- Escalation path. Which items go to which person, and how soon. Every item that can get stuck has a clock.
- Budget. Money per item and per month, a cap that ends a run, and the person minutes the role is allowed to consume.
The idea is older than models. Gerber's The E-Myth Revisited argues that a business runs on documented roles, not on whoever fills them; an automation is a role with nobody inside to improvise. Grove's High Output Management adds that a manager's output is the output of the organisation they manage; an automation joins that organisation the day it ships.
Filled in for the lead handler
The same agent, with the six fields written down. The two sums from how to use this course supply the budget line.
| Field | The inbound-lead handler |
|---|---|
| Inputs | About 400 leads a month from the website form, the general inbox and the partner-referral form. Message text up to 500 words; attachments are not read. |
| Outputs | A CRM row with five fields (name, company, service, budget band, source) and a reply sent from the sales inbox. It never quotes a price or a delivery date. |
| Done | Within 5 minutes of arrival, the CRM row is complete and either a reply has been sent or the lead is in the person queue with a reason. |
| Manager | An account manager. At launch, reads one sent reply in five, about 80 a month at 3 minutes; signs off every prompt change. |
| Escalation | No CRM match, or a request for a price: to the person queue, seen within 15 minutes in working hours. CRM or inbox unreachable: the owner is paged at once. Any lead not done at 5 minutes: an alert to the account manager. |
| Budget | Expected $0.06 a lead. A cap of 10 turns, which is 36,000 tokens in and 1,000 out, about $0.08, ends a run and queues the lead. People: 8 hours a month, $240. |
Run the renamed field through the filled-in role. The first lead after the rename hits the 10-turn cap, costs $0.08 instead of $1.00, and lands in the person queue with the reason "turn cap" and its CRM row unwritten; it is not done at 5 minutes, so the account manager is alerted. The incident costs an hour of leads handled by hand, not a month, and the worst month the budget allows is 400 × $0.08 = $32.
The six fields do not depend on the rung. Built as the one-call workflow, the budget line drops to $0.0054 a lead and the other five fields stay word for word the same.
What each missing field costs
Each blank field produces its own failure, surfacing somewhere you would rather it did not.
| Missing field | What happens | Who notices first |
|---|---|---|
| Inputs | It accepts whatever arrives: a spam run of 2,000 form submissions is 2,000 runs, about $118 | the invoice |
| Outputs | It writes wherever it can: a reply that quotes last year's day rate | a client, quoting it back |
| Done | "Handled" cannot be counted, so nothing is ever late | nobody, which is the problem |
| Manager | Nobody reads the output, and errors accumulate | a prospect, weeks later |
| Escalation | Stuck items wait without a clock | a prospect, by phone |
| Budget | Spend has no ceiling | the invoice, at month end |
Without the field, the first person to learn of the failure is outside the company or reading a bill. With it, the first person is the manager, within minutes.
The rest of the course fills these fields properly: module 1 writes done and the exceptions, module 4 the gates, the review sample and the escalation clocks (for this handler, a gate that from month 3 reads every reply before it sends, in place of the launch sample), module 8 the budget with an alarm on the rate of spend, and module 9 the owner, the weekly review and the retirement date. The turn cap itself, and why a loop needs three exits rather than one, is in agent loops and stop conditions.
In practice
When a vendor sells "an agent that handles" anything, ask them to fill the six fields for your volume, and listen hardest to three: who reads its output, what stops a run, and where a stuck item goes. "It learns" is not a manager, and "usage-based pricing" is not a budget; both answers leave the field to you. Inside the company, the rule is short: nothing ships without a named manager and a cap that ends a run.
The mistake that costs money is shipping a role with no budget and no manager. The two failures compound. Nothing stops the spend, and nobody is reading the output closely enough to notice that it has started.
Check yourself
1. Five of the six fields are filled in. Budget is blank, because "the expected spend is only $24 a month". Ship it this week?
Not without the cap. The budget field exists for the worst case, not the expected one. With no turn cap a looping lead costs $1.00, so one renamed field makes a $400 month that nothing reports before the invoice. A 10-turn cap is one setting, bounds a lead at about $0.08 and a month at $32, and sends every capped lead to a person.
2. In the launch months, a campaign triples lead volume to 1,200 a month. Which field breaks first, and what do you change?
The manager's. One reply in five is now 240 a month at 3 minutes, 12 hours, nearly 3 hours a week of an account manager's time, up from 1. Tokens rise to about $71 and the cap's worst case to $96, which the budget absorbs. Drop the sample to one in fifteen, back to 80 replies, only if the sampled error rate is low enough to pass module 4's test for sampling; otherwise add a second reviewer. Sum 1 is now $6,000 a month, so it still pays.
3. The escalation path reads "email the founder". Is that an escalation path?
No. It has no clock, nobody covers it when the founder is in a meeting, and it sends $30 problems to a $100 hour. Rewrite it by class: an unmatched lead or a price request goes to the account managers' queue, seen within 15 minutes; a CRM or inbox failure pages the automation's owner at once; only a contract question reaches the founder.