neuronalysis

What we do/In depth

For process owners and analysts

How we assess

How we decide, task by task, what an AI agent may do. This is the method we use in the analysis. You can use it without us.

← What we do

M1 · Jobs and tasks

We look at tasks, not jobs.

We break each role into task bundles, then single tasks, then steps. A role is rarely handed over whole. Most roles hold a few tasks an agent can do today.

In this example, five of the seven tasks can go to an agent at some step, and two stay with a person.

Each task gets its own verdict: the step an agent may start at (M5), or that it stays with a person.

Example · accounts payablemade up

Accounts payable clerk

Invoices in

Read and book invoicesspot check
Match invoice to orderspot check

Payments

Prepare the payment runapprove
Release payments over CHF 5,000person

Suppliers

Answer payment questionsapprove
Prepare a supplier callassist
Negotiate payment termsperson

M2 · How it runs today

Find out how the work is really done

Written process descriptions and the real process almost always differ. We look at the real one.

Process mining

Where systems keep logs, such as an ERP or a ticket system, we rebuild the real process from them: which steps, in which order, how often.

Task mining

With the consent of the people involved, we record which steps they take at their desktop.

Interviews and shadowing

We sit next to the people who do the work. The special cases are in no log, and they decide whether an agent works.

In a small company one or two workshop days with the owner and key people, plus exports from mail, accounting and CRM, are usually enough.

M3 · Scores

Every task, scored the same way

You get this table at the end of the analysis. One row per task, seven criteria, each scored 1 to 5.

TaskHow oftenFixed rulesData there Background neededMistake costEasy to checkRegulatedStart at
Read and book invoices spot check
Match invoice to order spot check
Prepare the payment run approve
Answer payment questions approve
Credit-limit exceptions person

blue: high is good for an agentred: high is bad for an agentmade up

M4 · Order

Two questions decide what comes first

What does a mistake cost, and can someone check the result quickly? Tasks where a mistake is cheap and easy to spot go first. Tasks where it is expensive and hard to spot go last, or never.

lowCost of a mistake →high

Approve each

Expensive. Easy to check. Payment run

Last or never

Expensive. Hard to check. Credit limits Payment terms

First

Cheap. Easy to check. Booking invoices Order matching

Assist only

Cheap. Hard to check. Call preparation
easyChecking the result →hard

What the seven criteria do

  • Cost of a mistake sets order
  • Checking the result sets order
  • How often it happens worth it?
  • Fixed rules or judgement possible?
  • Is the data there possible?
  • Background knowledge needed possible?
  • Laws and regulators limit

M5 · Autonomy steps

The agent starts small.

Autonomy means how much the agent may do without a person. It goes up one step when the numbers show it works, and down again when they don't. We agree the numbers first.

  1. Shadow mode

    Works next to your people

    does
    your people, as today
    agent
    same work, results compared
  2. Assist

    Prepares

    does
    a person
    agent
    drafts, never sends
  3. Approve

    Does it, a person approves

    does
    the agent
    checks
    every case
  4. Spot check

    Does it, a person spot-checks

    does
    the agent
    checks
    samples and all exceptions
  5. Autonomous

    Works alone

    does
    the agent
    checks
    monitoring, off switch

Up one step only on measured results

M6 · Fix the process first

Don't automate a bad process

Making a bad process faster with an agent is the most common mistake. We redesign it first.

Clear hand-over points

It is written down where the agent's work ends and a person's begins.

A path for exceptions

Every case the agent cannot handle has a defined way to a person. Exceptions often eat most of the savings, so we count them in shadow mode.

Structured inputs

Where a form or a fixed format removes guesswork, we introduce it before the agent.

M7 · People

Plan the people from the start

Who supervises the agents, who owns the process and who needs training are questions the analysis answers.

This is an operational risk as well as an HR topic. People who fear for their job give worse information about the work, and knowledge that is not written down leaves with them. We record it before anything changes.

NextStaff consultation, responsibilities and the law · Control and law →