Built to Work Alone, Asked to Work With Us

Built to Work Alone, Asked to Work With Us

A series on Agents and Biases

Part 1: The test that leaves the human out

Many of the benchmarks we use to judge AI agents share a quiet assumption: the agent works alone. A task goes in, an output comes out, and the result is marked right or wrong. It is a sensible way to compare tools. It may be an odd way to judge something that will spend much of its time inside a team.

Work has rarely been organised that way. A planner checks with the client before booking. An analyst asks which question matters before running the numbers. A site engineer flags a problem before the concrete is poured. My sense is that much of the value sits in those small exchanges: knowing when to ask, when to report back, and when to hand a decision to someone else.

Now an AI agent joins that work. It can search, draft, code and plan quickly. But the questions that matter in a team are different from the ones a benchmark asks. Does the agent tell you what it is about to do, or simply do it? Does it remember the correction you gave it ten minutes ago? Does it know which decisions are yours to make? And when you step in and do part of the work yourself, does it adjust, or carry on with its original plan?

These are less questions about intelligence than about being a good colleague.

If people and agents do better together than agents do alone, are we measuring, buying and designing for the wrong worker?

Part 2: What the evidence shows

Researchers have started to test this directly, by putting real people and AI agents on the same task and comparing the result with the agent working alone, or with people working without AI.

1. Together beats alone, but not on every task

In one study, 99 people worked alongside an agent built on GPT-4o on travel planning, writing a literature review, and analysing data in tables.¹ Compared with the same agent working on its own, the team's result was preferred in 86% of travel plans, 74% of data analyses and 66% of literature reviews.

A UK government pilot with 500 participants found a similar pattern for workplace tasks: people with access to an AI model scored 19% higher on quality and produced 61% more points per minute than those without.² Separate tests found that non-experts using frontier models had 4.7 times higher odds of writing a feasible lab protocol than those using the internet alone.³

The gains were uneven. In the workplace pilot, one task required open-ended strategic proposals, and AI access made no statistically significant difference. The authors suggest AI helps most with closed-ended analysis and synthesis, and still lags on open strategy. They also describe their estimates as preliminary, with a high degree of uncertainty.²

2. Agents do the work well, but collaborate poorly

The same collaboration study looked at where the agents went wrong. Communication problems appeared in 65% of sessions with real people: agents acted without telling users, or asked again for information they had already been given. Situational awareness problems appeared in 40%, and planning problems in 39%.¹

Collaboration also has a cost. In tests with simulated users, collaborative agents finished fewer tasks within the step limit than agents working alone, in one case 75% against 87%.¹ More turns mean more chances to stall.

The positive side: when they did deliver, the best collaborative agents produced better work than autonomous ones,¹ and satisfaction was highest (4.06 out of 5) on data analysis, where the agent handled most of the code and people mainly posed new questions.¹

3. People stay in the loop by choice

Users were not passive reviewers. They voluntarily took 21% to 32% of the actions themselves, working directly in the shared task rather than through the agent.¹ In a smaller follow-up, when people could work alongside the agent at the same time instead of taking turns, 70% preferred it, and satisfaction rose from 3.55 to 3.95 in travel planning and from 3.55 to 4.25 in data analysis.¹ Satisfaction was lowest (3.06) on the literature review, where users brought their own expertise and strong preferences.¹

One group of researchers argues that this is what many AI evaluations miss: they score AI on how well it imitates or replaces people, rather than on how well it works with them.⁴

The gap may lie less in what the agent can do than in how the work between person and agent is designed.

Part 3: Designing the work for two

If this evidence holds, the question for organisations shifts from "how capable is this agent?" to "how well does this pair work?" That changes what gets tested, what gets bought and how roles are drawn.

Measure the team, not the tool. Pilots could score three versions of the same work: person alone, agent alone, and the two together. In these studies, the team beat the agent alone, while agent-only scores could not show the communication failures. In practice, this could be a three-way comparison on one workflow over a few weeks, tracking quality, time and rework.

Design the handover, not only the output. Communication was the most common failure, and acting without telling users was a recurring pattern. Organisations could define when an agent must confirm, report progress or stop. A short "ask before" list is one example: anything that spends money, goes outside the organisation, or changes a shared record.

Let the task decide the split. The clearest gains came on closed-ended analysis. Strategy showed no measurable uplift, and satisfaction dropped where personal expertise and preference ran high. A practical start is to map a workflow's steps from closed to open, let the agent lead on the closed end, and keep people leading on the open end with the agent in support.

Questions for the two-person desk

1. Collaboration adds turns and checking. At what point does the better result stop paying for the extra time?
2. When an agent prepares a decision and a person approves it, who owns the outcome?
3. Will people keep stepping in at the same rate once they trust the agent, and should they?

I write about AI, AI transformation, the future of work, behavioural economics, and user research. Mostly questions, occasionally answers. If you are thinking about any of this, I would like to compare notes.

References

1. Shao, Y., Samuel, V., Jiang, Y., Yang, J., and Yang, D. (2024, revised 2026). Collaborative Gym: A Framework for Enabling and Evaluating Human-Agent Collaboration. ICLR 2026. arXiv:2412.15701
2. UK AI Security Institute (2 February 2026). AI and the future of work: measuring AI-driven productivity gains for workplace tasks. AISI
3. UK AI Security Institute (2025). Frontier AI Trends Report. AISI
4. Haupt, A., and Brynjolfsson, E. (2025). Position: AI Should Not Be An Imitation Game: Centaur Evaluations. ICML 2025. PMLR

LinkedIn version

If people and AI agents do better together than agents do alone, are we measuring the wrong worker?

Many AI benchmarks share one assumption: the agent works alone. A task goes in, an output comes out, and it is marked right or wrong.

[PLACEHOLDER: your practice context, e.g. what prompted this topic (a client conversation, a project). Vijay to confirm before publishing.]

These are observations I would consider when designing AI transformations or integrations for knowledge work, because the real question is less about which tool to buy and more about how the work is redesigned around it.

1. Together beats alone, but not on every task. In one study with 99 real users, the team's result was preferred over the agent alone in 66% to 86% of cases across three tasks. A UK pilot with 500 people found AI access raised quality by 19%, but made no significant difference on open-ended strategy.
2. Agents do the work well, but collaborate poorly. Communication failures appeared in 65% of real-user sessions, such as acting without telling the user.
3. People stay in the loop by choice. Users voluntarily took 21% to 32% of the actions themselves.

In practice, this means measuring the pair, not only the tool, and designing the handover as carefully as the output.

For teams adopting agents, the transformation may be less about replacing roles and more about deciding which steps the agent leads, which the person leads, and how they hand over.

My question: collaboration adds turns and checking. At what point does the better result stop paying for the extra time?

If you are piloting agents with your teams, I'd like to hear how you are splitting the work.

#AgentsAndBiases #Innodrive #AITransformation #FutureOfWork #HumanAICollaboration

First comment: The full article, with references and three hypotheses for designing work around people and agents: [BLOG LINK]

Slides (4 x 1080 square, Cover + Parts 1 to 3): Issue8_cover.png, Issue8_part1.png, Issue8_part2.png, Issue8_part3.png. Created and delivered in the Claude session; to be added here manually (direct upload to Notion was blocked from that session).

Research notes (internal)

Input (internal)