Rethinking work for humans and agents

I have recently been reminded of an old story about how iron was discovered. Some travellers noticed that when they had finished cooking their chicken, the stones at the bottom of the fire had melted together to form a very hard substance that adopted the shape of the fire pit. They produced the first ever guide to creating iron, which started with some good chickens…

When a new technology arises, it captures the imagination of innovators, business leaders and technologists alike, dreaming of the ways in which the technology can change society and business. Unhindered by an understanding of the new technology or its real applicability, early pioneers experiment and try out different ideas, aiming to create a breakthrough. This was very visible with blockchain and the proliferation of tokenisation and crypto initiatives - which in most cases made no business sense.

Today we have AI Agents with much excitement around how these can be used to transform our ways of working and even replace humans at scale. Again, many ideas follow the ‘chicken’ example by trying to linearly project a limited understanding of a new technology onto a restricted view of the current world. This leads to use cases for AI that are simply inappropriate or restricted by thinking that is anchored in old ways of working.

Recent examples I have encountered include

• Delaying the adoption of AI until all the processes have been through Lean Six Sigma and “all waste eliminated”.
• Delaying the adoption of AI until the “current database has been cleaned” (despite this not containing the knowledge that AI will use).
• Focussing on implementing an existing supplier’s chatbot in a laser specific role without considering how the service process can be transformed by a knowledge based approach.

The key question is: how to identify and specify application areas for AI Agents that are valid, valuable and verifiable?

A framework for Agentic business redesign.

Let’s start with the good news – we don’t have to reinvent the wheel. There are a number of thinking models that focus on understanding the structure of a problem that provide half of the solution. We also have emerging wisdom on the limitations of AI and the strengths of humans.

Some relevant background thinking:

• Turner and Cochrane's "goals and methods matrix" (early 1990s) considers how well-defined a goal is with how well-defined the method to achieve it is.
• The Stacey matrix (certainty about cause and effect vs. agreement on goals) and Snowden's Cynefin framework (clear, complicated, complex, chaotic) base an approach on problem type.
• Thomas Davenport's Thinking for a Living considers complexity of work with degree of collaboration, giving transaction, integration, expert, and collaboration models.
• Hammer's 1990 HBR piece "Reengineering Work: Don't Automate, Obliterate" warns specifically about automating the existing process instead of rethinking it.

Considering these and then adding the lens of ensuring verifiable outputs and containing the cost of errors, we have developed a model to help specify AI solutions to real world business problem types.

The model has five elements:

LoOutcome clarityHi

Curate outcomes

Co-create components

Code a Solution

Curate workflow

Fit for purpose ?

HiProcess clarityLo

Validation

These elements are arranged to reflect the structure of the problem and the challenge as to whether a valid solution arises to this problem.

Code a solution - Clear Outcome · Clear Steps

In this case, we are applying AI to rapidly develop a new solution or reverse engineer an existing solution, enabling vibe coding for further development. Once we have captured the function and architecture of the solution, the benefit is derived from:

• Enabling rapid improvements such as new features
• Increase reliability and resolve bugs or security weaknesses
• Reduce the dependence upon ‘knowledge in the heads’ of human developers

Curate outcomes - Unclear Outcome · Clear Steps

In this case, we aim to create a new, creative or unknown outcome through experimentation. The process that we will follow is clear and could be established in a workflow with AI components at key steps in the process. This could be generating a (request for) proposal, video or image.

Curate workflow - Clear Outcome · Unclear Steps

In this case we know the outcome that is desired without having a clear approach to achieve this outcome. Our goal is to curate the process that will result in the desired outcome. This could be new drug discovery, data cleaning and migration, or addressing trouble tickets in customer service.

Cocreate components - Unclear Outcome · Unclear Steps

In this case the problem is complex and ambiguous and may even involve ‘feedback loops’ which defy linear modelling. We accept that we cannot resolve it in one step – we need to break down the problem into components and create building blocks that provide parts of a solution. This involves techniques such as fractal thinking and problem decomposition. Over time, we can migrate towards a more complete and valid solution.

Examples include modelling energy demand and supply, global financial markets or resolving geo-political disputes.

Fit for purpose?

This validation stage applies to all the above four problem types – is our solution an adequate response to the problem we set out to address? The criteria for this include:

• Validity – does the solution address, reduce or resolve the problem?
• Value – cost of the solution versus cost of problem / value generated?
• Verification – what are the risks? and the likelihood and costs of failure?

Summary

Agentic Redesign begins with an understanding of the problem, its structure and character. Only then can a solution be proposed and validated using AI in various ways.

Einstein was once asked – if you only had one hour to solve a problem, how would you approach this? His answer has reverberated through time – “I would spend 55 minutes on understanding the problem, and 5 minutes in defining the solution”.

High — Outcome clarity: can you define and check “done”? — Low

Unclear outcome · Clear steps

Curate outcomes

Make quality explicit with rubrics, examples and criteria.

Agent pattern

Enforce workflow with steps and an evaluator loop

Human role

Judges and edits

Autonomy

Low to medium

Examples

Proposals from templates, management reports, due diligence

“What does success look like?”

Unclear outcome · Unclear steps

Co-create components

Frame and decompose to create workable building blocks.

Agent pattern

Cowork or research assistant; parallel exploration of options

Human role

Leads the thinking

Autonomy

Low

Examples

Strategy development, new service design, organizational change

“What would make this clearer?”

Clear outcome · Clear steps

Code a solution

Drive AI to generate and test a repeatable solution.

Agent pattern

Deterministic workflow

Human role

Handles exceptions

Autonomy

High

Examples

ERP, CRM, invoice matching, expense validation, onboarding checklists

“Human monitors execution”

Clear outcome · Unclear steps

Curate workflow

Specify the outcome and constraints, not the steps. Let AI experiment.

Agent pattern

Goal-directed, verified

Human role

Sets goals and approves results

Autonomy

Medium to high, within guardrails

Examples

Support tickets, data reconciliation, vendor sourcing

“Can we write the acceptance test?”

High — Process clarity: are the “rules” and steps known and stable? — Low