The differentiator is applied AI capability: knowing where to use it, how to design the work around it and how to determine whether it has actually created value.
At Mason Analytics, we call this Hiava™: the capability to design AI around the work, rather than simply add AI tools to the way the work is already done.
For many organisations, the first practical challenge posed by generative AI was relatively narrow: how do we prompt this thing effectively? Prompting still matters. Clear instructions, relevant context and well-defined outputs can materially affect what an AI system produces.
But the more important business question is now broader: how should this work actually be designed?
A good prompt may improve an individual output. It cannot determine whether AI is being applied to the right problem, what evidence should inform it, where human expertise is required or whether the resulting work is commercially useful.
In practice, applying AI intelligently means being able to answer a series of harder business questions. Eight are particularly important.
1. Are we solving the right problem?
AI projects often begin with the technology: Where can we use AI? This encourages businesses to automate whatever appears easiest rather than address what matters commercially.
The starting point should be a defined problem. Is the business trying to reduce production time, improve customer understanding, increase consistency, remove repetitive work or create something that was previously unaffordable?
The problem should be specific enough to assess. “Use AI to improve marketing” is not a useful objective. “Reduce the time required to turn customer research into an evidence-based creative brief” is.
This is also where domain expertise matters. A technically competent prompt cannot compensate for a poorly understood problem. Someone still needs to understand the customer, available evidence, commercial constraints and desired outcome well enough to know what the AI should actually be helping to achieve.
Without that, a business can demonstrate considerable AI activity without creating much value.
2. Is AI the appropriate solution?
Not every inefficient process requires AI.
Some problems are better addressed by removing unnecessary steps, clarifying responsibilities, improving conventional software or stopping the activity altogether. Starting with the business problem rather than the technology makes that possibility easier to see.
AI is particularly useful for certain forms of work involving language, images, pattern recognition, prediction, synthesis or variation at a scale that would otherwise require significant human effort. But it also introduces limitations, including inconsistent outputs, fabricated information, bias and reduced transparency.
The relevant comparison is therefore not AI versus perfection. It is AI versus the current method, including the relative cost, quality, speed and risk of each.
Sometimes AI will be transformative. Sometimes a spreadsheet or a redesigned process will be the better answer.
3. Are we redesigning the workflow or automating the existing one?
This is where business use of AI needs to move beyond prompt engineering.
A good prompt may improve an individual output. It does not determine where AI belongs in a process, what information it should receive, which tasks should remain human, how its output should be checked or what should happen next.
Adding AI to a poor process can simply make the same problems occur faster.
Businesses therefore need to look at the complete workflow: what information enters it, where decisions are made, where delays or unnecessary effort occur, what the customer ultimately receives and how quality is assessed.
AI may remove stages, combine tasks or change the order in which work is completed. It may also make previously impractical work commercially viable.
Research into real-world AI implementation shows why this matters. Performance depends not simply on the accuracy or capability of an AI system but on how it fits into the surrounding workflow. A capable system can still fail to improve performance if it creates additional work, interrupts established practices or provides information at the wrong point.
This shift from tool use to work design is central to Hiava™. The objective is not to insert AI into as much work as possible. It is to design a better way of doing the work.
4. Does the AI have the right context and data?
General-purpose AI begins with general knowledge. Yet businesses often expect it to produce specific, differentiated work without supplying the information required to do so.
Useful output may depend on customer research, brand positioning, product information, previous performance, technical constraints, internal knowledge and examples of what acceptable work looks like.
This is another reason domain expertise matters more than prompting technique alone. Effective instructions depend on knowing which context is relevant, which evidence is credible, which constraints matter and what a good result should contain.
More information is not automatically better. Poor, outdated or unrepresentative data can produce confident but defective outputs. The OECD identifies data quality and appropriateness as important characteristics of effective AI systems, while NIST recommends understanding an AI system’s context, intended purpose and likely consequences before deployment.
The quality of the information and expertise surrounding an AI system places a ceiling on the usefulness of what it produces.
5. Where is human judgement essential?
“Human in the loop” is frequently treated as a sufficient safeguard. It is not.
People can accept AI recommendations too readily, overlook errors or inherit biases contained in AI output. Experimental research has shown that biased AI advice can influence subsequent human judgement, even after the advice itself is no longer present.
Human involvement therefore needs a purpose.
That may include defining the problem, supplying contextual expertise, challenging assumptions, judging creative or brand suitability, handling exceptions or taking responsibility for consequential decisions.
The important question is not whether a person touches the work somewhere along the way. It is whether the right person is making the right judgement at the right point, with sufficient expertise and information to improve the result.
6. How will we evaluate the output?
Plausible output is not necessarily accurate, distinctive or useful.
Evaluation criteria should therefore be considered before implementation rather than after an impressive demonstration.
Depending on the task, these might include factual accuracy, relevance, consistency, customer response, brand fit, production quality, error rates or time saved.
The level of scrutiny should also reflect the consequences of failure. A low-risk internal drafting tool does not require the same evaluation as an AI system producing customer-facing advice or informing significant decisions.
Testing should include realistic cases, difficult examples and foreseeable failure conditions, not merely examples chosen because the system handles them well.
NIST’s AI Risk Management Framework places measurement alongside governance, contextual mapping and active risk management for this reason.
7. Can it work within the real business?
A successful experiment is not the same as a functioning implementation.
The AI must fit existing systems, responsibilities, approvals, security requirements and working practices. Employees need to know when to use it, what information they may provide, how its outputs should be assessed and where responsibility ultimately lies.
Adoption matters as much as technical capability. A system can perform well under test conditions and still fail because it adds friction, duplicates existing work or does not fit how people actually perform their jobs.
Implementation therefore requires both technical integration and deliberate work design.
The question is not simply does the AI work? It is does this way of working work?
8. Has it improved anything that matters?
Ultimately, AI adoption should be judged against the business problem it was intended to solve.
Relevant measures might include production cost, completion time, error rates, conversion, output capacity, customer satisfaction or the proportion of employee time redirected towards higher-value work. The appropriate measure depends on the original objective.
A baseline matters. Without knowing how the previous process performed, it is remarkably difficult to establish whether the new one is genuinely better.
Businesses should also look beyond the most obvious improvement. A process can become faster while generating additional review work, reducing quality or simply transferring effort somewhere else.
The objective is not to prove that AI was used successfully.
It is to establish whether the business now performs better.
Competitive advantage lies in application
As AI tools become more capable and widely available, access to the technology becomes less distinctive.
That does not mean AI stops creating competitive advantage. It changes where that advantage comes from.
It moves towards organisations that can identify worthwhile opportunities, bring genuine expertise to the problem, provide the right information, design effective human-AI workflows and evaluate the results against something that matters commercially.
This is the capability we mean by Hiava™.
The progression matters. Prompting is a useful skill. Workflow design is an organisational capability. The latter is considerably harder to acquire and replicate.
AI can make work faster and expand what a business can produce. It cannot independently decide which problems are worth solving, supply missing commercial understanding or determine whether an output deserves to be used.
Having AI is becoming standard.
Knowing how to turn it into better work and better business performance is not.
Sources
National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0).
National Institute of Standards and Technology. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile.
OECD. (2024). OECD AI Principles.
Wenderott, K. et al. (2024). Effects of artificial intelligence implementation on efficiency in medical imaging workflows: A systematic review. npj Digital Medicine.
Vicente, L., & Matute, H. (2023). Humans inherit artificial intelligence biases. Scientific Reports, 13.
Buçinca, Z. et al. (2022). Who goes first? Influences of human–AI workflow on decision-making. Proceedings of the ACM on Human-Computer Interaction.
