It’s Thursday afternoon, and your senior data engineer has another Slack message from finance waiting. The request is small: a join, a rename, a recalculated column. The interruption isn’t small. Every one of these requests steals time from the platform work your team already committed to, and over time that routine prep work starts setting the roadmap.
That tradeoff is showing up everywhere. McKinsey’s State of AI in 2025 finds that while AI adoption is now nearly universal, 88% of organizations report regular AI use in at least one business function but only about a third have begun to scale AI across the enterprise. In many organizations, the blockers aren’t model quality. They’re workflow, data, and operating model issues. Which means the pressure lands right back on the IT and data teams already carrying modernization, governance, and AI enablement at the same time.
Meanwhile, the business still needs answers. Analysts sit on backlogs of prep requests they could probably handle themselves if the tooling were easier and the governance were clear.
Natural language is starting to change that. An analyst can describe the workflow they need in plain language and get back something they can inspect, adjust, and run. For IT, that creates a real opportunity, but only if the workflow is being built inside the same control plane as the rest of your data work. Otherwise, you’ve just made it easier for people to build ungoverned processes faster.
What natural language actually changes in data prep
Imagine an analyst typing: “Combine the Salesforce opportunity export with Marketo campaign data, dedupe by email, and flag anyone who opened more than three emails last quarter.”
Conversational workflow building
A natural language assistant can turn that into workflow steps: joins, filters, standardization, validation, and whatever else the request calls for. The analyst still reviews the workflow visually, checks the transformations, changes anything that looks off, and runs it when they’re satisfied. The result isn’t a black-box answer. It’s a workflow someone can inspect.
AI-assisted cleanup and transformation
Natural language also helps automate the kinds of cleanup tasks analysts spend hours performing manually:
- Schema alignment across inconsistent source systems
- Standardizing date and currency formats
- Cleaning free-text fields where customers entered “U.S.A.,” “USA,” and “United States” in the same column
- Identifying missing values, duplicate records, or outliers before downstream reporting
Natural language helps most when it handles that work inside the same governed environment as the rest of your stack, not in a side channel where data leaves the audit boundary.
Workflow documentation as a byproduct
There’s another practical benefit too. AI assistance can help describe workflow logic, surface anomalies, and generate documentation while the workflow is being built. Most teams do not have enough time to properly document solutions as they are created. If documentation gets captured during the build instead of as a follow-up task, there’s a much better chance it actually exists when someone else has to maintain the workflow later.
What changes for data and engineering teams
Natural language doesn’t eliminate the need for engineers or data scientists. It does allow shifting basic prep work closer to the people who understand the business context and free up specialists for work that requires their skillset.
That’s the real redistribution: fewer interruptions for joins, renames, cleanup, and light enrichment; more time for pipeline architecture, governance design, feature engineering, and the quality issues that actually require deeper technical judgment.
When it works well, analysts move faster and IT gets fewer small requests that derail bigger priorities.
Why governance gets harder, not easier
If natural language makes workflow building easier, more people will build workflows. That can be good, but it also expands the surface area IT has to manage. With the increased scale, governance becomes more about policy, permissions, training, review paths, and monitoring.
Early rollouts usually go wrong in expected ways. Teams enable AI-assisted workflow building before defining approved data sources. Audit logging is technically available, but nobody has decided who actually reviews it. Users are given more freedom, but there’s no escalation path for sensitive data or questionable outputs.
The teams that handle this well usually start with a contained use case, a clearly named governance owner, and a small group of trained users. Once this pilot is operating well, then it’s time to scale to other teams and use cases.
The real question for IT
For IT leaders, the question isn’t whether natural language can build a workflow. It’s whether the workflow it builds is safe, repeatable, and traceable.
A self-service prep capability that bypasses governance is a liability, not a productivity gain.
Two governance layers matter most: data governance and AI governance.
Data governance: what happens to the data
Workflows built by less-technical users still need to operate under the same controls as everything else. In practice, that usually means:
- Role-based access controls tied to your enterprise identity provider
- Audit logs for workflow execution, data access, and changes
- Lineage tracking that shows where data came from and how it was transformed
- Policy enforcement for sensitive or regulated datasets
IBM’s 2025 Cost of a Data Breach Report found that among organizations that experienced breaches of AI models or applications, 97% lacked proper AI access controls and that high levels of shadow AI added an average of $670,000 to global breach costs. Unmanaged AI use is already creating measurable security exposure.
AI governance: what happens with the model
If workflows depend on large language models, IT needs clear answers to a small set of questions:
- Which models are approved?
- Where is the data processed?
- Which prompts and outputs are retained?
- Who can invoke external models?
- How is usage monitored and billed?
A simple example makes the issue concrete. A finance analyst wants a workflow step that summarizes contract PDFs and pulls out renewal dates. In a governed setup, the analyst chooses from approved model connections IT already configured. The workflow runs, the activity is logged, and the data path stays visible. The analyst gets speed and IT gets control over the model inventory, data flow, and audit trail.
A practical framework for self-service analytics
A useful way to think about rollout is to match governance intensity to workflow risk. Self-service does not have to mean identical permissions for everyone. It should mean the controls are appropriate to the level of risk.
Tier 1: Low-risk operational prep
Routine joins, formatting, deduplication, and summarization on governed internal data. Standard access controls and audit logging are usually enough. This is the best place for most teams to start.
Tier 2: Cross-functional enrichment
Workflows that combine multiple business systems, unstructured documents, or external AI services. This tier needs more deliberate controls: approved data sources, approved model connections for AI-assisted steps, and periodic output review.
Tier 3: High-risk regulated workflows
Anything involving customer PII, regulated reporting, financial controls, healthcare data, or model-driven decisioning. This tier needs explicit approval before production, separation between builder and approver, full lineage, and active monitoring.
Many organizations over-govern tier 1 work and under-govern tier 3 work. Natural language doesn’t fix that by itself, but it does make a tiered model easier to apply once the boundaries are clear.
What this looks like in practice
There are a few ways the scenario at the top of this post can end.
The bad version is that the business gets tired of waiting and routes around IT. People build shadow workflows in spreadsheets or paste sensitive data into public AI tools because they need answers quickly. That’s already happening in plenty of organizations, and it’s usually worse than the original backlog problem.
The better version is that IT gives people a governed way to do more on their own. That requires a platform with three things in one place: access to approved data sources, governed AI model access, and audit and lineage capabilities, plus workflow-building tools analysts can realistically use. This example runs in Alteryx One.
The analyst who sent that join request opens Designer and uses Ask Alteryx, the chat-based workflow assistant formerly known as Alteryx Copilot, to begin. She describes combining Salesforce pipeline data with Marketo engagement history, and Ask Alteryx walks through the join logic, deduplication, and field standardization, proposing each step for the analyst to review before proceeding.
She can also use the AI Tool Palette inside Designer. This palette includes purpose-built tools for tasks like schema alignment across inconsistent sources (Schema Fit Tool), fuzzy matching and standardization (Precision Match Tool), data extraction from structured documents like invoices (Invoice Extractor Tool), and configurable LLM-based steps for custom prompting (Prompt Tool and LLM Override Tool). Each tool operates within the same governed environment as the rest of the workflow.
Behind the scenes, execution metadata is logged, lineage is preserved, and access policies still apply through centralized identity controls. AI-assisted workflows do not sit in a weaker governance tier. They are subject to the same review and audit expectations as everything else.
What IT has to configure
Alteryx One’s governance capabilities include SAML and OAuth authentication, identity provider integration with platforms like Okta and Azure AD, role-based permissions, audit logging, SCIM provisioning for automated user management, and lineage integration pathways with governance platforms such as Atlan and Collibra.
For AI model governance, Alteryx One supports connections to enterprise-approved LLMs including OpenAI, Anthropic, Google Gemini, and Cohere, as well as bring-your-own-model (BYOM) approaches for organizations with specific compliance requirements. LLM connectivity is tied to the Full User role and can be narrowed further with custom roles, so IT can decide both which models are available and which users are allowed to invoke them.
How to judge readiness before rollout
Before rolling this out, three questions will tell you a lot.
- What type of requests cause the most interruption? If most incoming requests are repetitive prep work like joins, formatting, cleanup, and light enrichment, natural language can probably absorb a meaningful share of that load. If the queue is dominated by brittle pipelines, tangled dependencies, or undefined business logic, the issue is with the underlying architecture.
- Which users can take on more responsibility? The best early users are analysts who understand their business context well but do not have the technical fluency to automate prep unaided. Natural language can help them move faster, but it doesn’t replace judgment. If someone can’t tell whether a join result is plausible, the workflow may still be wrong, just faster.
- How mature is your AI governance today? If you cannot clearly answer which models are approved, where data is processed, what gets retained, who can invoke external models, and how usage is monitored, those gaps will usually show up after adoption starts. If sanctioned tools aren’t available, employees turn to unsanctioned ones, and the audit trail disappears entirely.
Most organizations are not going to stand up a governed self-service model in a single move. They’ll start with one workflow category, one governance pattern, and one trusted group, then scale from there.
Getting started
Alteryx One supports the full workflow described above: natural language-assisted data prep, governed AI model access, and enterprise-grade controls — in a platform IT can manage and analysts can use without engineering support. There are a few ways to evaluate it:
- Free trial: Start a free trial and explore natural language-assisted workflow building firsthand. No engineering setup required to get started.
- Request a demo: If you’d prefer to see the governance controls and AI capabilities configured for your environment, request a demo to walk through it with someone who can address your specific IT requirements.
- Explore governance capabilities: The Alteryx One governance overview covers the access controls, audit logging, identity integration, and lineage capabilities available to IT administrators.