AI at Work: Why Faster Tools Can Make Teams Weaker

When a colleague starts using a writing assistant, the inbox gets shorter and drafts appear faster. That individual gain is easy to celebrate. Less visible is what happens next: meetings shorten, fewer people read full reports, and decisions are sometimes taken with less cross-checking. This tension — faster individual work but weaker team decisions — is what many organisations face with AI at Work.

Research shows a clear split. Generative tools tend to raise individual productivity for creative or text tasks, while experimental studies and a broad 2024 meta-analysis find that human+AI combinations do not always outperform the best single actor on decision tasks. The practical question is not whether the tools are clever — they often are — but how teams change their processes, measurements and responsibilities once those tools are introduced.

Two research strands explain why faster tools can have opposite effects. First, controlled experiments and a systematic meta-analysis (covering studies up to mid‑2023 and published in 2024) examined many tasks where humans worked with AI. That review reports a small negative pooled effect when a human+AI pair is compared to the best single party, with a standardised effect size of about −0.23 (Hedges’ g). In plain language: on average, teams using both a human and an AI did worse than the single best actor in those experiments.

Second, field experiments with generative AI show clear individual speed gains on creation tasks. One randomised study testing a writing assistant found time reductions around 40 % and typical quality improvements around 18 %. These positive individual effects coexist with the meta‑analysis result because the research samples different task types and conditions.

Automation bias — the tendency to trust automated recommendations too readily — is a key mechanism linking faster tools to poorer group decisions.

Automation bias is worth a quick definition: it is the tendency for people to follow an automated suggestion even when it is wrong, or to omit action because the system gave no flag. That bias compounds when teams stop challenging quick outputs or stop checking the rare but costly errors.

If a compact view helps, the table below summarises the core findings behind the headline tension.

Not all tasks react the same way to speed. Generative and creative work — drafting copy, preparing first versions of reports, summarising research — benefits most from faster tools. For these tasks, the AI is a force multiplier: it fills blank pages, offers structures, and lets people iterate faster. The 2023 experiment that measured time and quality is a concrete example of that pattern.

Decision tasks are different. These include classification, risk assessment, hiring shortlists, or medical triage. In such tasks the process often requires comparing alternatives, weighing trade‑offs, and combining diverse perspectives. Faster generation of a recommended choice can shortcut those deliberations. Two common consequences follow: teams anchor on the AI recommendation and stop exploring alternatives, or coordination costs fall so discussions disappear altogether.

Some practical examples:

• Customer support: an AI drafts responses faster, reducing backlog, but the team may overlook a growing pattern of incorrect solutions if nobody reads full tickets.
• Hiring: an AI ranks candidates quickly. If hiring managers accept the ranking without cross‑checking, subtle but important human fit factors can be lost.
• Clinical workflows: AI can speed up scans review, but overreliance on a machine rating can delay critical second opinions when the algorithm errs.

These examples show why speed by itself is not an unalloyed good. The effect depends on whether the team adapts its process and metrics to preserve critical checks and shared understanding.

Four mechanisms explain the paradox.

1. Measurement mismatch. Organisations often measure individual throughput — how many reports produced, how many tickets closed. Faster tools improve these metrics, so incentives push for speed. Team‑level metrics such as decision quality or error cost are rarely tracked with the same attention, so declines there go unnoticed.

2. Coordination loss. When a single person can produce a complete deliverable, fewer people collaborate on it. That reduces the number of perspectives and informal checks that used to catch errors. What looked like efficiency is actually a loss of collective cognition.

3. Calibration and overreliance. Automation bias makes users trust an AI’s output more than they should. If the AI is usually good, people stop checking it; when a low‑probability mistake occurs, the team may miss it. The meta‑analysis shows this effect is especially strong in decision tasks.

4. Erosion of skills. Regularly delegating complex judgement to tools reduces the opportunity for people to practise and learn. Over months or years this can lower the baseline human capability — which in turn makes human oversight less effective.

These mechanisms interact. For example, measurement mismatch encourages coordination loss, which accelerates skill erosion. The practical consequence is not that organisations must avoid AI, but that they must treat deployment as a socio‑technical redesign, not a plug‑in productivity boost.

Research points toward a set of robust, low‑friction measures that make speed gains safe for teams.

Start with measurement: compare three baselines before any full rollout — human alone, AI alone, and human+AI. This is the clearest way to see whether a combination really improves team outcomes. Many studies recommend exactly this because averages hide the important detail that combinations help in some settings and harm in others.

Next, redesign tasks. Break complex decisions into substeps and assign each partner the part it handles best. For instance, let AI pre‑score candidates, but require a human panel to review top scorers using a structured checklist. Use forcing functions: require a short justification from a human reviewer before final approval, or add mandatory review rounds for rare, high‑cost cases.

Train for calibration. Teach teams how and when to trust system outputs, and include exercises that expose overreliance. Interface signals that show uncertainty (confidence indicators) can help but are not a cure by themselves — they must be combined with policies and training.

Lastly, protect learning and accountability. Keep audit logs, measure error costs at the team level, and keep a schedule for human re‑training so skills do not atrophy. When possible, run incremental pilots with A/B comparisons and open reporting so decisions rest on data, not intuition.

Faster AI tools change what work looks like. They often make individuals quicker and sometimes more effective on creative tasks, yet they can weaken collective judgement and coordination on decision problems. The evidence shows there is no universal rule: effect sizes depend on task type, on whether AI or humans are stronger in baseline performance, and on the organisational processes around the tools. The right response is pragmatic: measure team outcomes, redesign work into clear substeps, require human checks where consequences are high, and invest in calibration and training. Treated as a redesign problem rather than a simple productivity hack, AI at Work can deliver net benefits without making teams less smart.