Does AI improve work?

Apply healthcare system sciences to understand AI in the workplace

Artificial intelligence now writes software, reviews contracts, analyses financial data and answers customer questions. Its technical capabilities are impressive and improving fast. But capability is not benefit. Important question: does AI make the actual work better - and how would we know?

In healthcare we've spent decades learning, painfully, how promising technologies can disappoint when they meet the realities of complex work. Electronic health records promised efficiency and, for many clinicians, delivered hours of extra documentation. Alerts designed to prevent errors produced fatigue and distraction. Driven by experiences like these, various disciplines – improvement science, human factors, clinical microsystems and safety science - were developed and adopted to study both the tools and the systems people work in. Their lessons apply to AI in any workplace.

Behind effortless results may be a lot of people untangling, patching and making it work.

Improvement science: did anything actually improve?

Consider a company that introduces an AI assistant to its customer service department. The system answers most questions accurately, responds almost instantly and handles hundreds of enquiries a day. On technical measures, it performs well.

But employees spend considerable time correcting its responses. Complicated enquiries take longer to resolve because the AI adds extra handovers. Customers still complain.

Improvement science offers a simple but powerful principle to apply in such situations: all improvement requires change, but not all change is improvement.

The Model for Improvement, widely used in healthcare, asks three questions. What are we trying to accomplish? How will we know that a change is an improvement? What changes can we make that will result in improvement? 

Answering them requires three kinds of measures:

Balancing measures capture costs that headline results conceal. An AI tool can save time for one employee while creating extra work for three others. In a clinic adopting an AI scribe, for example, the outcome might be time spent on documentation, the process measure how often clinicians must edit notes, and the balancing measure the number of errors that slip into the record.

Improvement science favours small, iterative tests of change: Plan-Do-Study-Act cycles. Rather than deploying a technology everywhere at once, teams test on a small scale, study the results, decide whether to adapt, adopt or abandon, and expand step by step. This suits AI particularly well, because its capabilities, limitations and patterns of use keep changing.

Human factors: people and machines working together

Now consider an AI assistant that reviews legal contracts. It flags problematic clauses with impressive accuracy. But after several months, lawyers begin accepting its recommendations without careful checking. What happens when it misses something important?

This is a classic human factors problem. Human factors examines how people interact with technologies, tasks and working environments, and has a long history in aviation, nuclear power, manufacturing and healthcare.

Anaesthesia is a good example. Modern anaesthetic machines incorporate sophisticated monitoring, alarms and automated functions. But safe care depends on clinicians understanding how the equipment behaves, interpreting its information, staying alert and responding well when something goes wrong. Automation introduces well-recognised hazards. 

These are not arguments against automation. They are reasons to evaluate the human–machine partnership.

"Human oversight" is often proposed as the safeguard. But requiring a person to approve an AI output offers little protection if that person lacks the time, information or expertise to judge it. Human oversight must itself be designed and tested.

Human factors specialists use observation, simulation and usability testing to study these interactions under realistic conditions: do people recognise errors, understand the technology's limits and know when to intervene? The result is a fuller picture of performance than is conveyed by technical accuracy figures.

Microsystems: where the work actually happens

Perhaps the most transferable contribution from healthcare is the concept of the clinical microsystem.

A clinical microsystem is the small, functional unit that delivers care: an operating theatre team, an emergency department, a primary care practice, a ward. Each has its own people, routines, relationships, information flows and culture, and outcomes depend on how these elements work together. A skilled clinician and an excellent device don't guarantee good care if communication, coordination or handovers fail. Related work-system models, such as SEIPS, make the same point: people, tasks, tools, environment and organisation interact, so changing one changes the others.

This idea translates readily to other work settings. A bank has loan assessment teams, a law firm has case teams, a software company has development teams. Each is a workplace microsystem.

Consider an AI coding assistant. It may generate correct code and make individual developers faster. But what happens to the team? Does it increase the burden on code reviewers? Does it change how junior developers learn? Does it create defects that appear only in testing or live operation? Does it strengthen collaboration or encourage people to work in isolation?

These questions point to an important principle: the correct unit of evaluation is the team using AI – not the tool, and not the individual user.

Safety science: understanding both success and failure

Traditional safety investigations ask what went wrong and how to prevent it happening again. A complementary approach, Safety-II, tries to explain how things usually go right despite difficult and variable conditions.

Hospital staff constantly adapt to interruptions, missing information, unexpected events and shortages. Much of the time, care succeeds because people compensate for weaknesses in the system. Safety science calls this the gap between work as imagined – described in policies and guidelines – and work as done, with its shortcuts, informal channels and workarounds.

This gap is important for AI in two ways. First, a system designed around official workflows may struggle with how work really happens. Second, an AI system may look reliable precisely because employees quietly correct its mistakes. Its outputs look good, but the corrective work is invisible. Conversely, staff may discover valuable uses the tool developers never anticipated.

Evaluators therefore need to examine how results are achieved, not just the results themselves.

Evaluating AI-enabled work

Together, these disciplines change what we evaluate, what we measure and how we judge success. We can ask whether AI performs, but we need to know whose work improves, who inherits new burdens, who is accountable, and whether the benefits justify the full costs.

They suggest a multi-level framework:

Level	What we need to know
Technology	Does the AI perform accurately and reliably?
Human interaction	Can people use it effectively and recognise its errors?
Team or microsystem	How does it change workflows, coordination and the distribution of work?
Organisational safety	Can the organisation detect problems, recover and learn?
People and trust	Who benefits, who bears the risks and who is accountable?
Implementation and value	Are the benefits sustained and worth the costs?

No single method answers all these questions. Technical testing establishes what AI can do. Simulation shows how people interact with it. Workplace pilots reveal effects on teams and processes. Larger studies assess effectiveness and value. Ongoing monitoring is essential: healthcare has seen widely deployed prediction tools perform far worse in real hospitals than their developers reported.

As AI becomes more "agentic" – using tools, taking actions and interacting with other systems with less human review – these questions become more pressing. Moving from doing the work to supervising it, the risks of complacency and skill loss grow.

Conclusion

Healthcare system sciences offer proven ways to establish whether and how AI is improving work: define the aim, measure what matters including unintended consequences, study people and machines together, evaluate the team, and look for the hidden work behind the results. The ultimate test is whether AI helps people, teams and organisations perform safely, reliably and sustainably.

References and further reading

1. Batalden PB, Davidoff F. What is “quality improvement” and how can it transform healthcare? Quality and Safety in Health Care. 2007;16:2–3. https://doi.org/10.1136/qshc.2006.022046
2. Berwick DM. A primer on leading the improvement of systems. BMJ. 1996;312:619–622.
3. Langley GJ, Moen RD, Nolan KM, Nolan TW, Norman CL, Provost LP. The Improvement Guide: A Practical Approach to Enhancing Organizational Performance. 2nd ed. Jossey-Bass; 2009.
4. Nelson EC, Batalden PB, Godfrey MM, editors. Quality by Design: A Clinical Microsystems Approach. Jossey-Bass; 2007.
5. Carayon P, Schoofs Hundt A, Karsh BT, et al. Work system design for patient safety: the SEIPS model. Quality and Safety in Health Care. 2006;15(Suppl 1):i50–i58. https://doi.org/10.1136/qshc.2005.015842
6. Parasuraman R, Manzey DH. Complacency and bias in human use of automation: an attentional integration. Human Factors. 2010;52(3):381–410.
7. Goddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association. 2012;19(1):121–127.
8. Green B. The flaws of policies requiring human oversight of government algorithms. Computer Law & Security Review. 2022;45:105681.
9. Hollnagel E, Wears RL, Braithwaite J. From Safety-I to Safety-II: A White Paper. The Resilient Health Care Net; 2015.