AI Should Start With the Task, Not the Technology

3 minutes ago

7 min read

AI is already capable of supporting both judgement-led and evidence-led decisions.

A manager can use AI to generate options before making a judgement call. An analyst can use it to clean data, explore hypotheses or interpret results. In both cases, however, the process remains fundamentally human-led and AI-supported.

The human still needs to decide whether the output makes sense, whether important context is missing and whether it is fit for purpose.

This article deals with something different:

What happens when we start handing the task itself to AI?

That is the move towards AI-led automation.

And it should start with the task, not the technology.

AI changes tasks before it changes jobs

Much of the debate around AI has focused on whether it will replace jobs.

But jobs are really bundles of tasks.

Organisations employ people because there is work that needs to be done. Related tasks are grouped into roles based on the skills required, how the organisation is structured and the talent available.

AI does not necessarily replace that whole bundle at once.

It changes individual tasks first.

That makes a more useful organisational question:

Which tasks should AI perform, which should remain human, and how should the two work together?

This also explains why productivity gains can be real without jobs disappearing overnight. Some activities become faster, some are automated, others remain human, and the shape of the job changes around them.

Start by understanding the work

Before choosing an AI solution, organisations first need to understand the workflow.

What actually happens from beginning to end?
Where are the bottlenecks?
Where do errors occur?
Which activities take disproportionate time?
Which parts genuinely require human judgement?

Workflow or value-stream mapping helps make those questions visible.

Otherwise, we risk automating a poor process and simply performing the wrong work faster.

How predictable is the task?

For practical purposes, FYT finds it useful to distinguish between two broad types of work.

Tasks with a relatively clear correct outcome

Some tasks are governed by established rules, calculations or accepted outputs.

Examples might include:

reconciling transactions;
calculating standard formulas;
validating information against defined rules;
categorising routine transactions;
updating records.

Traditional software, rules-based automation and machine learning have long worked well in these environments because the expected outcome can usually be specified and checked.

Machines are also naturally suited to repetitive work where consistency matters.

Tasks with several acceptable outcomes

Other tasks are more open-ended.

Responding to customers, preparing marketing copy, summarising an issue, drafting a proposal or brainstorming ideas may all have several good answers rather than one objectively correct response.

These activities traditionally relied much more heavily on human language, judgement and context.

Generative AI has changed that boundary.

It can now perform many tasks that we once assumed required a person.

The distinction is not perfect, but it forces an important question:

What kind of task is this, and what kind of AI — if any — fits it?

DESIGN: Start with the business case

Once a potential task has been identified, the next question should not be:

“Which AI should we build/buy?”

It should be:

“What are we trying to improve?”

How much time does the task take today?
What does it cost?
How reliable are the current outcomes?
What quality level is acceptable?
Where do mistakes occur?

That gives us a baseline.

Without one, claims of “productivity improvement” are difficult to verify.

It also connects directly to Critical Problem Solving: define success — and failure — before investing heavily.

What does AI need to improve, by how much, and what result would tell us that the business case is not working?

That helps protect organisations from confirmation bias and the temptation to keep defending an initiative simply because time, money or reputation has already been invested.

BUILD: Use as much customisation as the task requires

Not every AI deployment needs to be unique.

For standardised tasks with established outputs, an off-the-shelf solution may be entirely adequate.

Financial statements, routine summaries, document classification or other activities governed by well-understood conventions may offer relatively little benefit from reinventing the process.

If a general-purpose system can perform the task reliably enough while reducing time or cost, that may already be a strong business case.

Not every AI deployment needs to create competitive advantage. Sometimes doing existing work faster, cheaper or more consistently is enough.

The question changes when the task reflects something the organisation does particularly well.

Perhaps the organisation has developed a distinctive way of assessing risk, responding to customers, writing proposals, diagnosing problems or making operational decisions.

In these cases, generic capability may not be enough.

The opportunity is to help AI emulate and scale the organisation's best practices.

That may involve providing:

proprietary knowledge and data;
policies and operating rules;
examples of strong and weak outcomes;
expert decisions and their rationale;
preferred formats and standards;
relevant workflow and customer context.

The technical method may vary. What matters is the principle:

Use generic AI where the task is generic. Add organisational context where the way you perform the task is part of your advantage.

If an off-the-shelf solution already achieves the required quality, reliability and economics, further customisation may add little.

But where the organisation wants to preserve and scale its best moves, its own knowledge and examples become much more important.

TEST: Does it actually beat the current process?

Generative AI makes it easy to produce something that looks impressive.

That can create a dangerous temptation:

build it, try a few examples and launch.

A system intended to perform organisational work should instead be tested against the process it is replacing or augmenting.

Three questions matter.

Is it better?

Does it actually improve the outcome we care about — time, cost, quality, customer experience or something else?

Is it reliable enough?

How consistently does it produce an acceptable result?

The answer depends on the task. A minor error in marketing copy is very different from an error in payroll, compliance or safety.

How does it fail?

AI does not just need to work on average.

We need to know:

what kinds of errors occur;
how often they occur;
whether they can be detected;
what happens when confidence is low;
and which cases should return to a person.

This gives management a much better basis for deciding whether — and how far — to automate.

The question is not whether AI works. It is whether it works well enough for this task, at this level of risk, to create a positive business outcome.

DECIDE & DEPLOY: Automation is a management decision

Once the system has been tested, deployment becomes a management decision rather than a purely technical one.

Leaders need to understand the gains, risks and trade-offs.

That includes decisions about:

the reliability threshold;
which cases can be automated;
which still require people;
when exceptions should be escalated;
how errors will be detected and handled;
and whether the business case still holds.

This may feel slower than simply switching on a new tool.

But once AI begins taking actions repeatedly across a workflow, poor assumptions can also be repeated at scale.

The discipline before deployment is what makes scaling safer.

Deployment changes both workflow and workforce

AI deployment does not affect only the task being automated.

It changes the work around it.

Perhaps employees no longer create the first draft and instead review AI-generated work.

Perhaps routine cases disappear and people increasingly handle exceptions.

Perhaps junior employees lose some of the foundational work through which they previously built experience.

Or perhaps one person can now handle a volume of work that once required several.

So AI deployment also requires organisations to think about:

roles and responsibilities;
workload;
skills;
approvals and controls;
learning and development;
and ultimately job design.

This brings us back to the earlier point.

AI changes tasks first. Jobs change because the bundle of tasks inside them changes.

Controlled automation before full autonomy

The current excitement around AI agents creates another temptation: move quickly towards systems that can plan and execute multi-step work with limited intervention.

That may make sense for some tasks.

But greater autonomy also means giving the system more freedom to decide how an objective should be achieved, interact with other systems and take action without human approval.

That increases the importance of boundaries, permissions, monitoring and exception handling.

For many organisations, a more sustainable progression may therefore look more like cruise control than fully autonomous driving.

The organisation defines the destination.

People decide which parts of the journey can be delegated.

The system performs those tasks repeatedly and at scale.

People monitor the outcome and take back control when necessary.

As reliability and confidence improve, more may be delegated.

Automation is something organisations earn through understanding, not something they simply switch on.

The objective is not AI. It is better work.

This final chapter brings the broader framework together.

We started with the idea that organisations can make decisions through judgement, evidence and automation, often using all three in combination.

We then introduced Critical Problem Solving as the discipline of solving the right problem, questioning assumptions and changing course when the evidence demands it.

The Analytics track showed how evidence creates a learning cycle:

Data → Insights → Decisions → Outcomes → Learning

AI-led automation extends that logic into execution.

Once we understand the task, define the business case, choose the right level of AI, provide the context it needs, test it against reality and understand its failure modes, we can begin handing over parts of the work.

But the important questions remain human:

What task are we trying to improve?Why should it be automated?What does success look like?How reliable does it need to be?What happens when it fails?And what does this change for the people and processes around it?

The sustainable advantage therefore does not come simply from adopting AI faster.