Syntic
All posts

Published October 4, 2026 in Future of Work

From Chatbots to Digital Coworkers

By Syntic AI

In 2023, researchers from Harvard, Wharton, MIT and Warwick gave Boston Consulting Group consultants a business case built to trip up GPT-4. The spreadsheet looked complete. The interview notes held the details that changed the answer. Consultants working without AI got it right about 84.5% of the time. Those working with GPT-4 scored 60% to 70%, on average 19 percentage points worse. They also finished faster.

The consultants were not careless. They treated the system like a capable colleague: once it has proved itself, you stop checking. With people that instinct is reasonable. With AI it is a liability.

The move from chatbots to digital coworkers is usually told as a story about software: more memory, more tools, more autonomy. The harder change is on the human side. Teams run on practices built over decades: handoff notes, review norms, a sense of who is good at what, clear responsibility, ways to train juniors. Nearly all of them assume a human teammate. Bringing AI onto a team means rebuilding each one on purpose.

The coworker frame brings the wrong instincts

We judge colleagues by building a model of their competence. Someone who writes a sharp market analysis can probably be trusted with the next one. The BCG study, led by Fabrizio Dell'Acqua, named the reason that habit fails with AI: a "jagged technological frontier". Some tasks sit comfortably inside what the model can do. Others look just as hard but fall outside it, and the two are hard to tell apart. On 18 tasks inside the frontier, the 758 consultants with GPT-4 completed 12.2% more tasks, worked 25.1% faster and produced work rated more than 40% higher in quality. On the one task outside it, they did worse than people working alone.

The group given a prompt-engineering overview fell further (24 points) than the group without it (13), a weakly significant gap. Learning to use the tool was not learning when to doubt it.

A 2024 meta-analysis in Nature Human Behaviour by Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone at MIT widens the picture. Across 106 experiments and 370 effect sizes, human–AI combinations on average did worse than the better of the human or the AI alone. The pattern underneath matters more. Combinations tended to lose on decision tasks and gain on content creation. When the human alone was stronger than the AI, the pair tended to beat both; when the AI was stronger, adding a human tended to drag it down.

So "add a human in the loop" is not a safety plan by itself. Whether pairing helps depends on the task and on who is stronger at it, and a team has to know that before it designs the workflow.

Continuity is engineered, not remembered

A chatbot conversation ends when the tab closes. A coworker is expected to pick up tomorrow where it stopped today, which is why assistants now ship memory; Anthropic, for example, added project-scoped memory that users can view and edit to Claude's team plans in September 2025. Memory helps, but it is not continuity. Continuity is knowing what was decided, what is still open, what failed, and what the next person must not miss.

Anthropic's engineers hit this with long-running coding agents. Their November 2025 account notes that "each new session begins with no memory of what came before." They compare it to a project "staffed by engineers working in shifts" with no memory of the previous shift. The fix was not a bigger memory but a working record: a progress file, a feature list marked passing or failing, and descriptive commits.

Hospitals learned the same lesson. A 2014 New England Journal of Medicine study of I-PASS, a structured handoff programme for residents, covered nine hospitals and 10,740 admissions. After it was introduced, medical errors fell by 23% and preventable adverse events by 30%. The clinicians had not got smarter; the handoff had become a standard format someone could check.

For mixed teams, decide what a handoff between a person and an AI, in either direction, must contain: the goal, the constraints, decisions already made, what is not yet verified, and who to ask when something looks wrong. Then make that record readable and correctable by a person. A memory nobody can inspect is a rumour about the past.

Delegation is now a skill people learn

The BCG researchers also watched how successful consultants worked and saw two patterns. "Centaurs" split the work, deciding which parts to hand to the AI and which to keep. "Cyborgs" wove the AI into individual subtasks, starting a sentence and letting the model finish it, going back and forth constantly.

Neither style is always right. Delegation has become a skill separate from knowing the subject: breaking work into pieces the AI is likely to handle well, stating acceptance criteria up front, and keeping the pieces where a wrong answer is costly or hard to spot.

A good manager does this with a new hire, with one difference. A new hire learns your context over months and says when they are confused. An AI teammate changes when the model, instructions or tools change, and rarely says it is out of its depth. So the map of what it does well has to be written down and maintained by the team, not held privately by each person.

Trust has to be set per task, not per colleague

Human-factors research has a precise vocabulary for this. In a 2004 paper, John Lee and Katrina See argued that the goal is not more trust but calibrated trust, matched to what the system can actually do. Too much leads to misuse; too little, to disuse.

Experience does not protect people from miscalibration. In a 2023 study in Radiology, Thomas Dratsch and colleagues had 27 radiologists read mammograms with suggestions from an AI-based decision support system. Some suggestions were deliberately wrong. Inexperienced readers were about 80% accurate when the suggestion was right and below 20% when it was wrong. Radiologists with more than 15 years of experience fell from 82% to 45.5%.

Two common fixes disappoint. In a CHI 2021 study, Gagan Bansal and colleagues found that AI explanations made people more likely to accept recommendations whether or not they were correct. Zana Buçinca, Maja Malaya and Krzysztof Gajos tested "cognitive forcing" designs, such as asking people to decide before seeing the AI's answer. These reduced over-reliance, but the 199 participants rated the most effective designs least favourably. The interfaces that protect judgement are the ones people like least.

Self-reports agree. In a Microsoft Research survey of 319 knowledge workers and 936 real examples of AI use (Hao-Ping Lee and colleagues, CHI 2025), more confidence in the AI went with less reported critical thinking, and more self-confidence with more.

In practice, attach trust to task types, not to the teammate. "It is reliable at first drafts of customer replies" is a calibrated claim; "it is reliable" is not. And set review intensity by the stakes, not by how the last few outputs looked.

When the AI teammate is wrong, someone else takes the blame

A colleague who makes a serious mistake can explain it, apologise, be retrained or be let go. An AI teammate can do none of that in a way that carries moral or legal weight. The responsibility goes somewhere else, and often not to where the control was.

The sociologist Madeleine Clare Elish called this the "moral crumple zone". In a 2019 paper she looked at the partial meltdown at Three Mile Island and the loss of Air France Flight 447. In both, blame settled on the operators nearest the failure, though their control over the automated system was limited. The human absorbs the impact, and, unlike a car's crumple zone, protects the system rather than the person.

Mixed teams can build crumple zones without meaning to. Someone asked to "review" a hundred AI-drafted decisions an hour, without the time or context to check them, is not providing oversight. They are providing someone to blame. Real accountability has three parts: the named owner can change how the AI works, not just approve its output; the review load allows genuine checking; and post-incident reviews ask about the setup (instructions, tools, permissions, handoffs) before asking who clicked approve.

This is where the coworker metaphor stops helping. The case for treating AI as a role with an owner is made in The Era of AI Employees Has Started. The human side is simpler: the AI can share the work but not the responsibility, so decide who carries it before anything goes wrong.

Managing a team where some members cannot be coached

Managers of mixed teams take on three new tasks. The first is keeping the frontier map: a shared, current record of which tasks the AI does well, which badly, and which nobody has tested. The second is disclosure norms: when a teammate must say a draft or analysis was produced with AI, so reviewers know how hard to look. The third is reconfiguring instructions, tools and permissions when the AI's behaviour drifts. Coaching people and reconfiguring systems are different jobs; managers now do both.

A 2022 National Academies report on human–AI teaming concluded that AI will remain inadequate on its own in many complex and novel situations and will need careful human management; today's agents have only raised the stakes. How a whole company organises around this is the subject of What an AI-Native Company Looks Like.

The skills question is about the humans

The most uncomfortable evidence concerns the people on these teams over time. In August 2025, The Lancet Gastroenterology & Hepatology published an observational study by Krzysztof Budzyń and colleagues from four endoscopy centres in Poland. After the centres introduced AI-assisted polyp detection, the rate at which endoscopists found adenomas during colonoscopies done without AI fell from 28.4% to 22.4%. It is one observational study, and the larger trial it sits within may not show the same effect, but it is an early clinical signal of skills fading under routine AI support.

Software shows a similar pattern earlier in a career. In a randomised trial Anthropic published in January 2026, 52 mostly junior engineers learned an unfamiliar Python library, some with an AI assistant and some without. The AI group finished about two minutes faster, a difference that was not statistically significant, and scored 50% on a follow-up quiz against 67% for those who coded by hand. How they used it mattered: those who delegated code generation or leaned on it to debug scored lowest; those who asked for explanations or conceptual help retained much more.

A team's ability to oversee AI depends on human skills that careless AI use wears down. The juniors who will one day notice that the AI's answer is wrong are building that judgement now, in the very workflows where it is easiest to skip. Protecting it is a design decision: some work done unaided on purpose, explanation-first habits, and review assignments that teach rather than only catch errors.

The measure of a digital coworker

Chatbots were judged on their answers. Digital coworkers will be judged by their teams: whether handoffs carry what the next person needs, whether trust follows the evidence task by task, whether responsibility sits with someone who can act on it, and whether the people are getting better at their work or losing the ability to check it.

None of that appears on a model benchmark, and all of it is in the hands of whoever designs the team. The BCG trap case could only be solved by reading the interview notes that a tidy spreadsheet made look unnecessary. On a mixed team, someone still has to read the notes.

Sources