For the last few years, most conversations about artificial intelligence have involved some variation of the same party trick.

Ask it something.

It answers...

Write an email. Summarise a report. Find an anomaly. Explain a spreadsheet. Suggest what somebody should do next.

Impressive, certainly.

But there is a fairly important distinction between telling someone what to do and actually doing it.

A satnav can tell you to turn left. It doesn't normally grab the steering wheel.

And this is why the development of Agentic AI within SAP deserves rather more attention than another announcement about artificial intelligence.

The interesting development isn't simply that SAP's Joule can become more intelligent.

It's that AI agents are increasingly being designed to participate in workflows and take actions across business processes.

That sounds like a small distinction.

It isn't.

First, what actually is Agentic AI?

Traditional generative AI largely operates on a request-and-response model.

You ask.

It answers.

Agentic AI introduces another capability: agency.

Instead of merely producing information for a human to consider, an AI agent can be given an objective and potentially coordinate a series of steps required to achieve it.

Imagine the difference between these two instructions:

"Which supplier invoices require attention?"

and:

"Identify the supplier invoices requiring attention, determine why, route the relevant cases to the correct people and initiate the appropriate next steps."

The first produces information.

The second starts resembling a colleague with access to your systems.

And that distinction becomes particularly significant when the system involved is SAP.

Why SAP changes the equation

There are plenty of places where an AI making a mistake is irritating.

If an AI writes a slightly strange LinkedIn post, somebody deletes it.

If it creates a poor image, somebody generates another one.

If it incorrectly summarises a document, hopefully somebody notices.

ERP is different.

SAP can sit underneath finance, procurement, inventory, manufacturing, HR, supply chain and countless other processes which businesses actually depend upon.

So putting increasingly autonomous AI into this environment creates an interesting asymmetry.

The more useful the AI becomes, the more consequential its actions can become.

This is why businesses probably shouldn't think about Agentic AI purely as an AI project.

It is also a process, data, integration, permissions and governance project.

The slightly uncomfortable question: what exactly are we automating?

Automation has always had one slightly inconvenient characteristic.

It doesn't automatically make a process good.

It makes a process automatic.

There is a difference.

Imagine a company has developed a purchasing process over fifteen years.

Different departments use slightly different procedures.

Supplier information contains duplicates.

Approvals have accumulated because nobody remembers why they were introduced.

One integration occasionally fails.

Several users have permissions inherited from jobs they stopped doing three years ago.

Humans compensate for these imperfections surprisingly well.

They know that Dave normally checks that.

They remember that this supplier is listed twice.

They know that the system says X, but in this particular situation you actually do Y.

This invisible human correction layer exists inside almost every organisation.

AI doesn't necessarily inherit that institutional knowledge simply because it has access to the software.

Which creates an important principle:

Before asking whether AI can automate a process, it may be worth asking whether you would actually want that process automated exactly as it exists today.

Agentic AI may therefore expose something businesses have been able to ignore for years:

the difference between the documented process and the real one.

Data becomes more important when AI can do something with it

Bad data is hardly a new problem.

But historically, bad data often created inconvenience.

A person spots something odd.

They investigate.

They correct it.

They continue.

Now imagine an AI agent using that same information as the basis for subsequent actions.

The economics of bad data change.

A duplicated supplier isn't merely an untidy database record.

An incorrect classification isn't merely something that makes reporting annoying.

An outdated customer record isn't simply an administrative nuisance.

They become potential inputs into automated decisions and workflows.

This is one of the paradoxes of AI.

The more sophisticated the intelligence sitting above your systems becomes, the more important the boring foundations underneath it become.

Data quality suddenly isn't boring.

Neither are integrations.

Neither are permissions.

Neither is process design.

Unfortunately, nobody has yet invented an exciting keynote presentation called "Getting Your Master Data Right."

Perhaps they should.

Permissions become a very different question

There's another distinction worth considering.

When somebody can see information, we think about access.

When somebody can change something, we think much more carefully about authority.

AI agents potentially blur those two worlds.

If an agent is going to participate in a procurement workflow, what can it initiate?

What requires approval?

What can it change?

When should it escalate to a human?

What happens when confidence is low?

Who is accountable for the resulting action?

And crucially:

what shouldn't the agent be allowed to do?

These aren't reasons to avoid Agentic AI.

They're reasons to design it properly.

Businesses have spent decades developing segregation of duties, approval hierarchies and access controls for humans.

Introducing software capable of taking increasingly autonomous actions means those principles don't disappear.

If anything, they become more important.

Integration suddenly matters rather a lot

Very few enterprise processes exist entirely inside one beautifully contained system.

A customer order might touch ecommerce, CRM, SAP, warehouse systems, logistics providers, payment services and reporting platforms.

Procurement might involve supplier portals, SAP, banking systems and external approval processes.

The AI agent may be intelligent.

The environment surrounding it may be considerably less elegant.

And an agent cannot magically repair a broken interface simply by being clever.

This creates another counterintuitive consequence of Agentic AI:

AI could make integration capability more valuable, not less.

The temptation with every new technology wave is to assume the old problems disappear.

Usually they don't.

They simply become the infrastructure upon which the new technology depends.

Humans aren't necessarily disappearing from the process

Agentic AI is sometimes discussed as though businesses face a binary choice:

Human does job → AI does job.

Reality is likely to be considerably more nuanced.

Some actions can be autonomous.

Some require approval.

Some should only generate recommendations.

Some require escalation when defined conditions occur.

Some processes may remain predominantly human.

The interesting design question therefore isn't:

"Can AI do this?"

It is:

"At which points should AI act, and at which points should a human remain involved?"

That's a much better question.

Because maximum automation and optimum automation aren't necessarily the same thing.

A Formula 1 car isn't impressive because the driver has nothing to do.

It's impressive because technology and human judgement have been allocated to the places where each contributes most.

Enterprise AI should probably be approached with similar thinking.

So what should SAP organisations be looking at now?

You don't necessarily need to begin with an enormous AI transformation programme.

A more useful starting point may be understanding the environment AI would eventually be expected to operate within.

Look at five things...

Processes — Are your real workflows understood, or merely the ones shown in documentation?

Data — Can automated actions safely rely on the information inside your systems?

Permissions — Is it clear who and what can initiate, approve and execute actions?

Integrations — Are the connections between SAP and surrounding systems reliable enough for increasingly automated workflows?

Governance — Where should AI act autonomously, where should humans approve, and how will actions be monitored?

None of these questions are particularly futuristic.

That's precisely the point.

The future of AI may depend on some very unfashionable work

There is something slightly amusing about the current AI revolution.

We are spending enormous amounts of money creating extraordinarily sophisticated intelligence.

And then discovering that its effectiveness may depend upon whether somebody correctly maintained the supplier master data.

But perhaps that's how technological change usually works.

The breakthrough attracts the attention.

The infrastructure determines whether it works.

SAP Agentic AI represents an important shift because AI is moving closer to execution.

For organisations running SAP, that creates considerable possibilities.

But it also changes the question.

For years we've been asking:

"What can AI tell us?"

Increasingly, the question will become:

"What are we comfortable allowing AI to do?"

And before answering that, there is one question worth asking first...

Is your SAP environment ready for AI to act?

Spottech works with organisations across SAP, S/4HANA, integration, AMS and enterprise systems. If you're exploring Agentic AI and want to understand what it means for your existing SAP environment, contact Spottech to discuss it.