Source-linked AI summary
Fits and Starts: Enterprise Use of AutoML and the Role of Humans in the Loop
Anamaria Crisan, Brittany Fiore-Gartland
TL;DR
Enterprise AutoML use remains insufficiently understood, including how visualization is used with it. The paper analyzes interview transcripts through selective coding and identifies usage scenarios, continuing human effort, and infrequent visualization use. It notes that the findings’ diversity and sparse observations limit their generalizability.
Problem
Few studies examine whether and how data visualization is used with AutoML in real-world settings, leaving enterprise use insufficiently characterized.
Source: §1 · p. 1Method
The study uses selective coding to organize interview transcripts around an existing framework of data-science work.
Source: §3.2 · p. 5Results
The study identifies three routine enterprise AutoML scenarios and finds that considerable human effort remains necessary while visualization is used infrequently and minimized in data-science work.
Source: §7 · pp. 13–14Takeaways & Limitations
The findings have implications for redesigning visualization’s role in AutoML and for tools that balance automation with human agency.
Source: §7 · p. 14Diverse experiences, some findings based on a handful of observations, and reliance on an existing framework constrain the findings’ generalizability.
Source: §6.3 · p. 13
Abstract
from arXiv · showhide
AutoML systems can speed up routine data science work and make machine learning available to those without expertise in statistics and computer science. These systems have gained traction in enterprise settings where pools of skilled data workers are limited. In this study, we conduct interviews with 29 individuals from organizations of different sizes to characterize how they currently use, or intend to use, AutoML systems in their data science work. Our investigation also captures how data visualization is used in conjunction with AutoML systems. Our findings identify three usage scenarios for AutoML that resulted in a framework summarizing the level of automation desired by data workers with different levels of expertise. We surfaced the tension between speed and human oversight and found that data visualization can do a poor job balancing the two. Our findings have implications for the design and implementation of human-in-the-loop visual analytics approaches.
1 INTRODUCTION
Organizations are adopting AutoML amid shortages of technical expertise, but enterprise systems still require human intervention and must balance speed with human agency. Interviews reveal varied AutoML use, substantial human effort, and limited use of visualization, motivating a framework for automation levels suited to workers with different expertise.
1 INTRODUCTION
- Few studies examine whether and how data visualization is used with AutoML in real-world settings, despite its potential role in human-machine collaboration.Enterprise AutoML systems remain limited in flexibility and do not yet span the full data-science workflow.Source: §1 · p. 1
- Interviews with 29 people across organizations of different sizes examined current or planned AutoML use and visualization’s role in the human-in-the-loop.The study focuses on visualization as a way to integrate people into automation.Source: §1 · p. 2
- AutoML use requires considerable human effort, complicated by work handoffs across expertise levels and a data-knowledge gap participants believe AutoML is widening.Participants also found visualization difficult and time-consuming to create, often slowing automation with limited benefit, while useful collaboration tools were lacking.Source: §1 · p. 2
- The study contributes enterprise use cases, a three-scenario summary, and a framework of desirable automation levels for people with different technical expertise.The authors position these contributions as resources for designing human-in-the-loop tools that balance automation and human agency.Source: §1 · p. 2
2 RELATED WORK
Prior work describes AutoML as expanding from model selection toward broader data-science workflows, while noting that systems still depend on human labor. Research also offers human-in-the-loop interaction and visualization approaches, but enterprise use across data-science work remains insufficiently understood.
2.1 AutoML in Data Science
- AutoML has broadened from automating hyperparameter tuning and model selection to encompassing multiple data-science steps, though enterprise systems still require substantial human labor.Fixed-structure pipelines are common but can be brittle across data types and tasks; variable-structure pipelines learn process networks for datasets and objectives.Source: §2 · p. 2, §2.1 · p. 2
2.2 Automation and the Human-in-the-loop
- Human-in-the-loop research frames automation as a balance between speed and human guidance, with interaction modes ranging from user-driven to cruise-control and autopilot.Prior work also studies feedback to AutoML and the information people need to interpret its outputs.Source: §2.2 · p. 3
- Enterprise studies of AutoML use remain few, and earlier work on practitioner settings focused on collaboration or trust rather than AutoML’s broader intersection with data-science work.The cited prior studies include practitioner interviews, ethnography, and controlled experiments that did not recruit current practitioners.Source: §2.2 · p. 3, §2.4 · p. 4
2.3 Data Visualization and AutoML
- Visualization systems support exploration and comparison of models and pipelines, but fixed-structure assumptions can make variable pipelines difficult to compare visually.Existing systems are described as a cruise-control mode between slower user-driven and faster, less transparent autopilot interaction.Source: §2.3 · p. 4
2.4 Situating our Research
- This study broadens enterprise research by examining how AutoML is situated across organizational data-science processes and how enterprise users view visualization-mediated human involvement.Prior enterprise studies addressed narrower themes, including collaboration and trust.Source: §2.4 · p. 4
3 METHODOLOGY
The study used semi-structured interviews to examine enterprise AutoML use and visualization, refining its inquiry as themes emerged. The analysis applied selective coding to interview transcripts using an existing data-science-work framework as a scaffold.
3.1 Interviews and Data Collection
- Semi-structured interviews explored organizational data-science work, current or planned AutoML use, and visualization tools; 29 participants were recruited through snowball sampling.Interviews lasted about 60 minutes, were recorded and transcribed, and covered a broad range of topics.Source: §3 · p. 4, §3.1 · pp. 4–5
3.1.1 Sensitization to Emerging Concepts.
- Initial coding and debriefs surfaced expertise differences, click-based versus code-based solutions, tool switching, trust, collaboration, and predictive modeling as themes that refined interview questions.The researchers also began with sensitizing concepts from prior work on data-science work and workers.Source: §3.1.1 · p. 5
3.1.2 Participant Characteristics and AutoML Use.
- Participants included 17 analysts and 12 managers; 8 were active AutoML users, 10 were experimenting, and 11 were knowledgeable but not using or planning to use it.They represented organizations ranging from fewer than 100 to more than 50,000 people and a range of industries.Source: §3.1.2 · p. 5
3.2 Selective Coding Process
- Selective coding organized transcripts around four higher-order and fourteen lower-order data-science processes from an existing framework, rather than generating a new code set.The researchers used constant comparison and report achieving theoretical saturation; they also used tensions between observations and the framework to motivate later critique and modification.Source: §3.2 · pp. 5–6
- Interview statements were coded as explicit process references when terminology mapped directly to the framework, or as implicit references when context supported the inference.Figure 3 illustrates this annotation approach.Source: §3.2 · p. 6, Figure 3 · p. 6
4 RESULTS
The results section first presents general attitudes toward AutoML and automation, then examines AutoML across higher-order data-science processes.
4 RESULTS
- The results are organized from general attitudes toward AutoML and automation to its intersection with higher-order data-science processes.Source: §4 · p. 6
4.1 Attitudes Toward Automation
Participants viewed AutoML with cautious optimism: it can improve productivity and widen access, but concerns about tool fit, human oversight, and unsafe decisions tempered enthusiasm.
4.1 Attitudes Toward Automation
- Participants saw AutoML as a productivity aid, but its use or exploration was more widespread in larger organizations for reasons the authors say remain unclear.Participants at larger organizations cited automating work and triaging toward more pressing problems.Source: §4.1.1 · p. 6
4.1.1 Improving Productivity.
- AutoML can speed routine work for data scientists and lower barriers for less-technical workers, though participants differed in how much expertise safe use requires.Participants saw automation of model tuning and selection as a way to shift effort toward verification and correction.Source: §4.1.1 · p. 6
4.1.2 The Importance of Integrating Tools.
- Recursive data work and tool switching limited visualization’s role, while some participants expected less visualization as organizations scale.Visualization remained in use for static reporting or personal understanding, but participants described it as poorly integrated or abandoned in communication.Source: §4.1.2 · p. 7
4.1.3 Concerns of Automating Bad Decisions.
- Participants feared blind trust in AutoML could enable consequential mistakes, and some worried non-experts would need substantial guidance to use it safely.Regulatory explainability and high-impact financial decisions heightened concerns; participants also described restricting access to data scientists.Source: §4.1.3 · p. 7
4.1.4 Limiting the Human-in-the-Loop.
- Participants saw human-in-the-loop approaches as valuable at the right points, while also wanting to minimize manual steps and pursue “good enough” results that surface issues early.They described AutoML as enabling rapid, low-code prototyping and “fail fast” use, but said poorly timed applications could feel intrusive.Source: §4.1.4 · p. 7, §4.1.5 · p. 7
4.2 AutoML in Data Science Work
Across the data science pipeline, participants described persistent human work around AutoML, with desired automation and oversight varying by task and expertise. They also identified unmet needs in preparation, governance, and communication, alongside rapid prototyping uses.
4.2.1 Data Preparation Continues to be a Rate Limiting Step for Automation.
- Data preparation remained labor-intensive, from combining sources and profiling data to assessing whether it was ready and useful for analysis.One participant estimated that 40–50% of their team’s time went to bringing data together; rapid AutoML trials could help assess partially cleaned data, but participants still wanted better-integrated preparation tools.Source: §4.2.1 · p. 8
4.2.2 Use of AutoML in Analysis Varies by Data Science Role.
- Desired control over AutoML analysis varied with expertise: experts favored customization, while less-technical users often needed guidance, low- or no-code tools, and visualization.Participants also described experts as a potential bottleneck for analysts, while warning that less-technical users may struggle to interpret or refine models and that trust in the analyst matters alongside trust in AutoML.Source: §4.2.2 · pp. 8–9
4.2.3 ‘Good Enough’ Rapid Prototyping to Bootstrap Anal-
- Participants used AutoML for rapid prototypes to assess data or start client conversations before committing to deeper development or investment.They valued this “good enough” use for surfacing issues early, although participants said AutoML’s generalizability across different customer problems remained challenging.Source: §4.2.3 · p. 9
4.2.4 Inadequate Support for Governed and Managed De-
- Governed deployment and monitoring remained human-intensive, and participants reported that existing tools did not adequately support AutoML governance.The volume and variety of data and models complicated oversight; participants called for monitoring support, particularly where financial or regulated models face strict controls.Source: §4.2.4 · p. 9
4.2.5 Correction and Repair of Deployed Models.
- Repairing deployed models depends on human oversight, but analysts may lack the expertise to diagnose problems and choose corrective actions.Participants described AutoML as brittle and reliant on iterative correction that can require both technical and domain expertise.Source: §4.2.5 · pp. 9–10
4.2.6 Communication and Collaboration Build Trust in Au-
- Communication and collaboration require additional work: visualization tools may be poorly integrated, results need translation for business users, and workflows need to support coordination across roles.Participants described a post-analysis presentation gap and challenges carrying recommendations through to market implementation; they also sought workflow awareness, handoffs, and tools usable across organizational processes.Source: §4.2.6 · p. 10
4.2.7 Summary.
- The authors conclude that no end-to-end AutoML solution currently covers the full scope of data science work, leaving workers to bootstrap tools and visualizations into existing processes.They argue visualization guardrails should reflect users’ expertise, context, and expected automation, while serving diverse people across organizational workflows.Source: §4.2.7 · p. 11
5 INTERPRETATION OF FINDINGS
The findings distinguish three AutoML uses shaped by workers’ expertise and the consequences of errors, and argue that desired automation varies across data-science work while human oversight remains necessary.
5.1 Three Usage Scenarios for AutoML Emerge
- The three scenarios are routine-task automation, rapid prototyping, and democratized model creation; the last calls for heavy guidance because users may struggle to detect or correct errors.The scenarios depend on technical expertise and the magnitude of consequences associated with errors.Source: §5.1 · p. 11
5.2 Varying the Level of Automation Across Data Science Processes
- The proposed eight-step workflow varies desired automation by process and expertise: high-expertise workers can automate routine work and prototype, while less-experienced workers need more support and domain expertise remains relevant downstream.Human oversight is required throughout, with “cruise control” proposed as a mode for overseeing and steering systems; Figure 1 needs future validation for generalizability.Source: §5.2 · p. 11
5.3 Eliciting Tasks for Visualization Design
- Visualization design should account for workflow-wide effects on data and tasks, since mismatched task elicitation can produce tools that are dismissed as poorly integrated into end-to-end work.Prototyping may serve model development, rapid failure discovery, or customer-facing exploration, while monitoring may allow high automation until human action is needed.Source: §5.3 · p. 12
5.4 Modifications to our Analytic Framework
- The analytic framework adds collaboration under communication, governance under deployment, and Guidance after communication, with human-machine, human-human, and organizational forms.Collaboration involves multidirectional exchanges; governance concerns launching, monitoring, and refining production models.Source: §5.4 · p. 12
6 DISCUSSION
The discussion frames enterprise AutoML as human- and organization-involved data work, not simply model automation. It identifies mismatches in automation and visualization support, alongside limits on how broadly the findings can be generalized.
6 DISCUSSION
- “Cruise-control” automation may be adopted in enterprise settings, but diverse expertise, complex tooling, and brittle workflows impede it; visualization is rarely used to support it and may be removed.The authors suggest visualization may be misspecified because researchers lack an understanding of how automation is used and where human-in-the-loop interaction occurs.Source: §6 · p. 12
6.1 Implications for Automating Data Science Work
- The authors argue for augmenting rather than automating data science, since end-to-end automation obscures the human labor needed and leaves inadequate support for human-machine collaboration.They argue that making human engagement explicit brings workers’ needs into consideration.Source: §6.1 · p. 12
- Collaboration and trust among data workers matter to data work; participants reported that AutoML could erode trust across expertise levels by enabling “citizen data scientists” to automate bad decisions.The authors argue that “humans-in-the-loops” better describes AutoML’s place within wider data-science and organizational processes.Source: §6.1 · pp. 12–13
- Participants were primarily data-science experts, so the authors call for follow-up with non-data-scientists who work closely with them.Some participants were managers overseeing mixed teams.Source: §6.1 · p. 13
6.2 Implications for Data Visualization Systems
- A visual workflow overview may help with handoffs, but visualization systems may not integrate with existing tools, scale to organizational data, adapt to workflows, or serve differing expertise levels.The authors recommend surfacing these ecological-validity threats early.Source: §6.2 · p. 13
- Guardrails should fit expertise, context, and expected automation: preparation needs more human labor, whereas deployment monitoring may need human response chiefly when events such as model drift arise.The authors suggest well-designed guardrails may support trust and collaboration, and call for research on dynamic, adaptive guardrails.Source: §6.2 · p. 13
6.3 Limitations and Future Work
- Sparse observations and reliance on an existing framework are the primary limitations; the findings need generalizability validation, which may be difficult given AutoML’s novelty.The authors suggest developing a survey instrument to probe enterprise AutoML use and potentially inform later mixed-methods research.Source: §6.3 · p. 13
7 CONCLUSION
The conclusion identifies routine automation, rapid prototyping, and democratized access as enterprise AutoML scenarios, while emphasizing human effort and oversight. It also reports infrequent visualization use and points to implications for visualization’s role in data science.
7 CONCLUSION
- Enterprise AutoML use spans routine-work automation, rapid prototyping, and democratized access, but still involves complex handoffs and considerable human effort, making oversight necessary for safety and effectiveness.Visualization can play an important role with AutoML, but participants used it infrequently and actively minimized it in data-science work.Source: §7 · pp. 13–14