← Playbook

Every company we assess has the same fifteen tasks. Here is the list.

Different industries, different words, same fifteen jobs. If you recognise more than eight of these, you have a straightforward automation project and you do not need anyone to tell you that.

2026-08-16/5 min read/Levelbrook AI Practice

We have now counted the work in enough companies to say something slightly deflating: the tasks are almost always the same. The nouns change — policies, orders, patients, matters, jobs, loads — and the shape does not.

Here is the list, in the order it usually shows up on the volume ranking, with the automation shape each one takes.

#The taskShapeGate
1Retyping a document into a system of recordextraction + confidence gateper-field
2"Where is my [order/claim/appointment]?"read + replynone needed
3Sending a document somebody asked forretrieve + attachnone needed
4Answering the phone outside business hoursvoice agent + calendar writeL3, recall
5Booking, rescheduling or cancelling an appointmentcalendar writeL2 → L3
6Small credits, refunds and adjustmentsreversible writeL2, ceiling
7Correcting an address or contact detailreversible writeL2 + history rule
8Opening a return, ticket, case or claimcreate actionL2 → L3
9Chasing somebody for a missing documentscheduled workflownone needed
10Copying data between two systems that do not talkorchestrationidempotency
11Building the same report every week by handscheduled generationnone needed
12Routing an inbound thing to the right personclassificationabstain on doubt
13Onboarding checklist that lives in someone's headworkflow + templateshuman sign-off
14Reconciling two lists that should matchdiff + exception queueexceptions only
15Answering the same five internal questionsgrounded Q&Anone needed

What the table is really telling you

Six of the fifteen need no approval gate at all. They are read-only or they generate something a human was going to check anyway. Those can ship in weeks and they are how you build organisational confidence cheaply.

Most of the rest are reversible writes. Which is exactly the population the approval queue exists for: real actions, real time saved, and a one-click undo if it goes wrong.

Almost none of them require a frontier model. They require good retrieval, a typed action layer, and somebody who counted. The intelligence is in the plumbing.

Not one of them is a job. Every one is a slice of several jobs — the slice nobody enjoys. This is worth being precise about with your staff, because the fear is real and the honest answer is better than the vague one.

The three that are usually mis-prioritised

#14, reconciliation, is almost always underrated. It is invisible — nobody complains about it because it happens at month end and then it is over — and it is often the single most expensive task on the list in hours. Ask specifically.

#11, the weekly report, is usually overrated. It feels automatable and it often is, but the volume is one per week. Two hours a week is 100 hours a year, which sounds fine until you compare it with the 630 hours in row two.

#4, the phone, is systematically underrated in service businesses because the loss produces no record. There is no row in any report for the person who rang, waited nineteen seconds, and called somebody else.

Do this before talking to any vendor Print the list. Put a tick next to the ones you recognise, an estimated volume per week next to each, and a "yes/no" for whether a mistake can be undone in one click. That sheet of paper is a better brief than most six-week discovery engagements produce.
Keep reading

This is what we do all day.

Support automation, AI phone agents, n8n back-office work, and the engineering loop itself — always behind a gate you control.