News

• AI usage is far from even when it comes to managers vs. regular workers
• Workers are facing more work or increased complexity due to AI
• Redesigning work is more important than upping quality or quantity

New PwC research has claimed 19% of UK workers now use AI every day at work, up from 15% last year, but despite some progression, there are still clearly some biases when it comes to who uses AI and how they use it.

For example, while 85% of senior execs and 73% of managers have used AI, only 35% of non-managers have used it.

But even if more knowledge workers were to use artificial intelligence as part of their daily routines, PwC argues that AI isn't actually making work any easier.

Even though employees claim AI has helped improve the quality of their work (70%), use more of their skills (63%) and bring value to the workplace (58%), the report actually argues that nearly half (44%) of workers have seen their workloads increase following the use of AI, with a similar number (45%) worried their job's complexity has also increased.

With this growing complexity in mind, PwC UK Chief Innovation and Technology Officer Claire Reid explained that employers and employees need to "rethink how work gets done" instead of just increase quality or quantity.

"The real prize isn’t just doing more work; it’s redesigning work so that people and AI together can deliver better outcomes," Reid summarized.

Still, the discrepancy in use among different levels of seniority is hard to ignore, and PwC could have the answer. Many fear negative consequences if they experiment with AI, are confused about their employer's policy or simply need more training. A clearer strategy and greater communication could be enough to change this and enable more workers (and therefore employees) to benefit from AI's more positive impacts.

"Organisations need to rethink work around the right problems, give people the skills and confidence to use AI well, and put the right safeguards around it," Reid added.

An AI agent approves a supplier payment. The invoice matches the purchase order, the amount sits within budget and the supplier is already in the system. Every check passes, but the decision is still wrong because the contract expired yesterday.

That’s the risk enterprises now face. An agent can make a decision that is entirely logical within the information it has and still be wrong because the business has changed around it.

For the past few years, the enterprise AI debate has focused mainly on the models. Which is smartest and which is the most efficient?

That focus made sense when AI was mostly an assistant. A copilot would find and analyze information, or summarize a dataset before handing the result to a person.

But human review is not a perfect safeguard. People can defer to confident-sounding outputs or rubber-stamp recommendations, allowing a bad answer from an assistant to feed directly into a consequential decision.

Agentic AI raises the stakes further. An agent can update customer records, approve requests, trigger workflows or make significant business decisions without a person standing between the model and every action.

As that happens, we have to ask questions that go beyond which model is best and start asking about the environments around those models. What data can the agent access? What rules govern its behavior? What is it allowed to change? And does it have enough context to understand what a sensible decision actually looks like?

Those questions become more urgent when a system can turn a flawed answer into a real-life action, changing records, approving transactions or sending a process in the wrong direction before anyone notices.

So, how do you reduce that risk?

While it may sound counterintuitive, part of the solution is to give AI greater access to your business.

If an agent is going to act on your behalf, it needs visibility into your organization's goals and rules, as well as the current state for the task at hand. That doesn’t have to mean giving it unrestricted access to everything, but ensuring it can see the right information in real time, rather than waiting for its next training update.

But there’s an important distinction here. When I say give an agent “access”, I mean giving it greater visibility and context, not necessarily greater authority.

It may need to understand the customer, the transaction, the workflow, the rules around it and what has already happened. That doesn’t mean it should be free to change all of those things.

That’s the balance enterprises need to get right. Broad context, but with narrow authority.

If agents need broad context to make good decisions, where does that context actually come from?

This is where operational context starts to become the differentiator. Far too many enterprises are still training AI agents in a piecemeal way, hoping they can simply feed them more and more company documents and, over time, they will absorb the context required to run and manage parts of the business.

The problem is that this approach can only get you so far. A document might tell an agent what a rule says, but in most businesses, rules have exceptions. They're also regularly updated, which means old documents can quickly become misleading.

And that’s a problem because if AI is learning from outdated material, or from examples where those rules were applied differently, it can end up building the wrong understanding of how the business actually works.

So if piecemeal training only gets you so far, what’s the alternative?

The solution is what I call putting “context at the core”.

That means placing AI within your core operational platform and grounding it in the systems where the enterprise already records commitments, applies rules and carries out transactions.

Instead of working from fragments of the organization, the agent gets a fuller picture of what is happening and what should happen next. Rather than simply receiving a list of rules to follow, AI agents can see how those rules are actually being applied across the business.

That gives them much greater context about which rules apply in particular situations, whether those rules are still current and whether the agent is actually allowed to act.

That doesn’t mean every AI application has to live in the same place. Large organizations will always have a wider technology estate, with different applications and services working together.

What’s more important is having a trusted core that brings together the data, rules, permissions and history AI needs to understand how the organization actually works.

That is a much stronger starting point than having lots of isolated AI tools across the enterprise, each working from its own partial snapshot of the business.

My own industry is a good example of why this is so important. Insurance is highly regulated and data-intensive, and the right decision can depend on policy terms, customer circumstances, local rules and exactly where a claim or account stands at that moment.

Take a household claim after a storm. An agent working on it might need to understand the policy wording, effective dates, endorsements, billing status, repair estimates, fraud indicators and the latest activity on the claim. If the policy changed yesterday, an extract taken last week could already be out of date.

That does not mean pouring every available document into a model. Too much irrelevant or contradictory information can make an agent less reliable. The goal is to retrieve the smallest set of current, authoritative facts needed for the task from the systems where those facts are maintained.

Insurance makes that need especially clear, but the same principle applies anywhere AI is acting across complex enterprise processes.

If context gives an agent a better understanding of what is happening, governance determines what it is actually allowed to do about it. That is why the two need to sit together. In practice, context at the core means putting context and control in the same place. Governance has to sit inside the workflow, not around it.

First, give the agent enough context to understand what’s going on, but be very clear about what it can do with that information. It might be able to read a record but not change it, recommend an action but not approve it, or act on its own only up to a certain point.

Then make sure you can see what it’s doing. Important actions should leave a clear record of the information the agent used, the decision it made and any changes it triggered. Security controls should also stop emails, attachments and other external content from being treated as trusted data or instructions.

You also need to keep testing in case models or workflows change. Enterprises need to keep checking how agents behave and be clear about when a person needs to step in.

And don’t hand over too much too quickly. Start with answers, move to suggestions and only then to actions. Read-only access and dry runs can show teams how an agent behaves before it is allowed to make changes inside enterprise systems.

The models underneath all of this will keep changing too. The goal is to build context and permissions into your operating environment so they stay in place whichever model you use next.

Enterprises that succeed with agentic AI start by limiting execution rights and expanding autonomy only after the model proves it interprets company logic correctly.

We've featured the best AI website builder.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit