AI Collaboration Needs a Shared Workflow

Individual speed and team dependencies

The clearest gains from AI often sit inside work one person can change independently.

In a randomized field experiment across 66 firms, Microsoft 365 Copilot users spent roughly two fewer hours each week on email during the latter half of a six-month study. Their broader task quantity and composition changed little. Individual behaviours moved more readily than meeting behaviours that depended on other people changing their routines.

One person can draft an option or test another direction without coordinating a group. Teams still have to agree on the problem, compare credible options, resolve priorities, and carry the decision through implementation.

A randomized experiment at Procter & Gamble found that an individual working with AI matched the average output quality of a two-person team working without it during a one-day product-innovation task. AI also reduced the technical or commercial skew in proposals, which helped individuals contribute beyond their functional specialty.

The experiment studied a bounded task outside the extended work of an established team. Larger groups, iterative rework, implementation, and long-term business results sat outside its scope. Proposal quality within one day offers limited evidence about how an organization reviews a disputed decision and delivers it over time.

Role overlap needs explicit accountability

Cross-role participation rose sharply in Figma’s samples. Designers who reported participating in development increased from 21% in 2025 to 41% in 2026. Developers doing design work increased from 44% to 60%.

Those figures may reflect genuine collaboration, role substitution, duplicated effort, or a mixture of those conditions. The published totals combine them.

AI expands the work each role can contribute. A designer may produce working code, and a developer may create a credible interface direction. Expanded contribution needs a named decision owner and an assigned reviewer with the relevant responsibility.

This became a practical issue in my last role as soon as cross-role building got easier. People could move into adjacent work faster than our expectations about accountability could adjust. One leadership conversation reduced the issue to a named person who would carry accountability when the work failed.

Cheaper production increases the amount of judgment required. Among Figma respondents building AI products, 90% said design mattered at least as much as before AI, with 57% saying it mattered more. These figures describe perceptions within a vendor’s user base. They correspond with the practical pressure created when teams have more plausible artifacts to evaluate.

More output can also contain less genuine variety. In a preregistered experiment on short-fiction writing, access to AI-generated ideas improved average ratings and made the resulting stories more similar to one another. The work studied fiction, so its relevance to product teams is limited. It does support a plausible concern that a larger set of polished options may contain less variation than the count suggests.

Visibility and provenance

Figma reported that 76% of surveyed product builders completed at least half their work on what the company calls “the canvas.” Fifty-nine percent said most of their work happened there. The underlying question offered six responses across different combinations of canvas and code, plus a not-applicable option.

A shared workspace gives collaborators access to the same artifact and its surrounding feedback. The artifact may still omit why an AI-generated option exists, which sources shaped it, or which constraints remain unresolved. Human approval also remains a separate decision. Since Figma surveyed its own users, the effect of the canvas on team outcomes remains unresolved.

AI-assisted work needs enough provenance for another person to assess it. The record should include the task, source context, intended use, and unresolved constraints. The eventual decision belongs beside that information so another person can challenge the reasoning or continue the work without reconstructing its history.

At Assent, this changed how we thought about the shared artifact. Coded sources of truth, including Storybook-backed components and living prototypes, became increasingly useful for implementation. Figma remained important for ideation and review. The coded references reduced interpretation differences because Engineering could inspect the intended behaviour as it developed.

Adoption labels describe limited information

Figma grouped adoption into four patterns based on the pace of personal and organizational uptake. Its sample placed 36% in unified adoption, 27% in directive adoption, 20% in grassroots adoption, and 18% in nascent adoption. The total reaches 101%, presumably because of rounding.

The categories can support a team discussion. Figma’s public report omits the underlying questions and the cut points used to assign respondents. The published information is insufficient to reproduce the segments or connect aligned adoption with stronger performance.

Each category points to a different operating condition. Grassroots adoption can leave local expertise isolated. Directive adoption can separate leadership intent from daily work. Nascent adoption describes slow movement across the organization. Unified adoption indicates alignment without specifying the practices that produced it.

The practical work begins when an isolated experiment enters a repeatable team workflow with an owner, an assigned review, a documented decision, and an outcome measure.

Structure shared decisions

A 2026 Microsoft Research preprint shows how added process can reduce performance. Researchers compared approaches to AI use among 388 employees at one retailer. A protocol requiring paired, joint AI use was associated with lower document quality and substantially lower production than unstructured use.

Its scope limits generalization because session assignment, participant dropout, document scoring, and local context may have influenced the result. The study is most useful as a warning against turning every AI interaction into a group ritual.

Structure belongs around shared decisions, accountability, review, and measurement. People can complete individual drafting or exploration using the method that fits the task.

In my previous role, readiness and timing constrained collaboration more than idea generation did. Epics arrived with missing success measures and unclear scope boundaries. Permissions and data-lifecycle assumptions remained vague. The result was avoidable rework and additional alignment meetings.

We made readiness explicit through five stages: discovery, solutioning, validation, specification, and quality assurance. Design aimed to work at least two sprints ahead. Development could begin once the team had roughly 80% confidence and the remaining uncertainty was visible.

A workable AI collaboration model has five parts:

1. Make experiments inspectable: record the task, tool, source context, and intended use so another person can understand and repeat the workflow.
2. Assign an owner and reviewer: one person owns the recommendation, while someone with the relevant responsibility reviews it against an explicit standard.
3. Define the decision: record whether the work supports exploration, narrowing, approval, or delivery.
4. Preserve variety among options: compare independently developed directions before selecting one, especially when several artifacts begin from the same model output.
5. Measure a team outcome: time to decision and avoidable rework are useful starting points, while review load or viable options considered may suit other workflows.

Start with one live workflow. Document how AI-assisted work enters it, assign the decision owner and reviewer, and use an existing outcome measure to check whether the change improved team performance.