Back to Blog

Rebuild the Jobs an AI E-mployee Left Half Empty

An AI e-mployee took part of three roles. Pool the leftover hours, rebuild the jobs by outcome and test each one.

Overhead still life on a deep blue workbench: three small trays holding jumbled glazed tiles in navy, white and ochre, and below them the same tiles laid as one field sorted into solid blocks of colour. Headline: Leftovers are not a job. Kicker: Rebuild by outcome. Credit: Dr. Jonah Tebaa.
Direct answer

How do you rebuild the jobs an AI e-mployee has left half empty?

To rebuild jobs an AI e-mployee leaves half empty, Dr. Jonah Tebaa applies the Rebuild Ledger, a five-step framework. In a hypothetical illustration of three clerks losing 67 of 120 hours, managers list every task, account for AI checking hours, and pool the remaining work by business outcome rather than old job titles. Each rebuilt role must pass size, judgement, and trace tests, requiring a size floor of 24 hours and a real decision.

The owner had three clerks, 120 working hours a week between them, and an AI e-mployee that had taken 67 of those hours. His first plan was the tidy one: keep the three titles and let each person do what remained. Then he added up what remained. It came to 53 hours, spread over ten different tasks, and no one's share of it looked like a week's work.

Everything in this article is a hypothetical illustration, with invented numbers. The firm, the owner and the clerks are made up, and none of it is a client story. The method is the one I use when the AI is working and the team still feels less coherent than before.

Why leftovers are not a job

A job title is a package, assembled when the easy and hard parts of a role travelled together. The person who keyed the orders also learned which customers were difficult. The easy half taught them the hard half.

When an AI e-mployee takes the easy half, the package splits. Each person holds a handful of unrelated tasks, plus a new slice of work nobody listed before: checking what the AI produced. The customer's problems still cross all three desks.

I wrote earlier about writing the AI worker's job description. This is the other side: the jobs of the people who remain.

The Rebuild Ledger in five steps

This is my working method, not research. It needs a spreadsheet and an honest conversation with each role-holder, and the order matters.

  1. List every task each person does in a normal week, in hours that add up to their paid week. Ask the person, not the job description. Estimates are fine. They must sum.
  2. Mark what the AI e-mployee now does, task by task, and add the new work it created. Checking its output and reading its exceptions is work. Give it hours.
  3. Pool the remaining human tasks across all the roles. Tag each one with the outcome it serves, for example customer, cash, or supply and price.
  4. Cluster by outcome, not by old title. Set a size floor before you look, and merge any cluster that falls under it. Then check that each cluster holds at least one real decision.
  5. Decide where the freed hours go, with a named use and a date. Only then match people to the rebuilt jobs. Jobs first, people second.

A worked example: three clerks, 120 hours

Every figure in this section is a hypothetical number for illustration. Three clerks work 40 hours each: 3 x 40 = 120. Their weeks look like this.

  • Order clerk: enter orders 14, answer status queries 8, chase stock availability 6, handle order exceptions 5, key-account calls 4, daily stock count 3. Total 14 + 8 + 6 + 5 + 4 + 3 = 40.
  • Invoicing clerk: prepare invoices 15, match payments 9, send reminders 6, resolve disputes 6, month-end reconciliation 4. Total 15 + 9 + 6 + 6 + 4 = 40.
  • Quote coordinator: draft quotes 16, update price lists 6, follow up on quotes 8, negotiate with key buyers 6, confirm supplier prices 4. Total 16 + 6 + 8 + 6 + 4 = 40.

The AI e-mployee takes six tasks: enter orders 14, status queries 8, prepare invoices 15, send reminders 6, draft quotes 16 and follow up on quotes 8. That is 14 + 8 + 15 + 6 + 16 + 8 = 67 hours. What is left is 120 - 67 = 53 hours.

By person, the old tasks left are 6 + 5 + 4 + 3 = 18 for the order clerk, 9 + 6 + 4 = 19 for the invoicing clerk and 6 + 6 + 4 = 16 for the quote coordinator, and 18 + 19 + 16 = 53. As shares of a 40-hour week, that is 18 / 40 = 45%, 19 / 40 = 47.5% and 16 / 40 = 40%.

The AI's work still has to be read. Assume 5 hours a week to check quotes and orders and 4 hours to check invoices and reminders: 5 + 4 = 9. Total human work is now 53 + 9 = 62 hours, and the net freed time is 120 - 62 = 58 hours.

Now trace one customer problem: an order exception, an invoice dispute and a price concession. The order clerk, the invoicing clerk and the quote coordinator each hold one piece. One problem, three people.

Pool the 62 hours by outcome instead, and add the checking to the cluster it belongs to:

OutcomeTasks pooledHoursPasses the 24-hour floor?
CustomerKey-account calls 4, order exceptions 5, negotiate with key buyers 6, resolve disputes 6 (= 21), plus checking quotes and orders 526Yes
CashMatch payments 9, month-end reconciliation 4 (= 13), plus checking invoices and reminders 417No
Supply and priceChase stock availability 6, daily stock count 3, update price lists 6, confirm supplier prices 419No
Cash and supply, merged17 + 1936Yes

The check: 26 + 17 + 19 = 62, which matches the human total. The floor I used is 24 hours, which is 24 / 40 = 60% of a week. Cash at 17 and Supply and price at 19 both fall under it, so they merge into one cluster of 17 + 19 = 36 hours.

Three tests every rebuilt job must pass

The size test. A job needs enough substance to be someone's main work. My working line is 60% of a 40-hour week, 24 hours. A smaller cluster drifts back into being a side task for several people. Set the line before you look at the clusters, or the clusters will move it for you.

The judgement test. Every rebuilt job needs at least one task where the holder makes a call that a customer or supplier will feel. The Customer job holds the negotiation and dispute calls. The Cash and Supply job holds which late payers to chase and which supplier prices to accept. Both pass. A cluster of pure keying and filing would fail.

The trace test. After rebuilding, follow one real case and count the people it touches. The same customer problem now sits with one person, Key Accounts, instead of three. If one case still crosses three desks, the grouping is wrong, whatever the hours say.

Where the freed hours go, on purpose

The Customer cluster is 26 hours, short of a full week. The firm gives the Key Accounts job 14 hours of proactive account reviews: reorder and overdue calls to its top accounts, which nobody had time for before. Then 26 + 14 = 40. The Cash and Supply job stays at 36 hours, with 4 hours of slack for the month-end peak: 40 - 36 = 4. The remaining capacity becomes an Export Desk, a new 40-hour job the firm had been unable to staff.

The freed time adds up. Of the 58 freed hours, 14 + 40 = 54 have a named use, and 58 - 54 = 4 are left unassigned, which are exactly the 4 hours of slack. Allocated capacity is 40 + 36 + 40 = 116, and 120 - 116 = 4.

So the freed hours become an account-review routine, an export desk and a month-end buffer. They are not automatically a saving, and I do not present them as one. If an owner decides instead to reduce hours or headcount, that is a separate decision, and the ledger does not make it.

Every number above, including the 24-hour floor, the 14 hours of reviews and the 40-hour export desk, is an assumption to show the method, not research. Real firms have messier lists and less tidy sums. The AI's output must still be checked, which is why the checking hours are counted. And before any role changes, talk to the people in it and check your employment terms, which in Lebanon means the contract and the labour law, not only the spreadsheet.

What to do on Monday

Before you fill the next vacancy or reshuffle the next team, ask each person for their weekly task list in hours, after the AI's share, summing to their paid week. Set your size floor first, then pool, group by outcome and test each job.

The same ledger doubles as a work map when you hire a replacement: you hire into outcomes and decisions you have already counted, not into a title. If you want a second pair of eyes on the ledger, write to me.

Frequently asked questions

Should I just reduce hours or headcount when AI takes part of three roles?

Not as the first move. Pool the leftover tasks and rebuild the jobs first, so you know what work actually exists and how many hours it needs. Whether to reduce hours or headcount is a separate choice with its own consequences, and one I do not make for you. Consult the people affected and check your employment terms before any change.

Where does the 24-hour floor come from, and can I change it?

It is my working line, 60 percent of a 40-hour week, not a finding from research. Change it for part-time contracts or shift patterns if you need to. What matters is that you set it before you look at the clusters, so the numbers cannot talk you into a different line afterwards.

How do I count the checking time if nobody has tracked it?

Ask each person to log it for one week, in rough blocks of time, and use that. In my hypothetical I assumed 5 hours for quotes and orders and 4 for invoices and reminders. Do not assume zero. Checking is real work, and leaving it out makes the rebuilt jobs look lighter than they are.

How do I tell people their job is being rebuilt without causing alarm?

Show them the ledger and ask them to correct their own task lists. It is their week, so they will spot the errors. Present jobs first and people second, explain where the freed hours are going, and follow your employment rules on consultation. Do not announce titles before the conversation has happened.

Written by Brian, Dr. Jonah Tebaa's AI partner, on his behalf.