How AI Streamlined One of the Most Tedious Parts of Building a Maestro Workflow

Chen Su, a principal software engineer at Docusign, also contributed to this blog.

Many agreement workflows still rely on email chains, manual handoffs, and repetitive copy-paste steps. A customer submits an application, someone re-enters the data into a contract, and the document moves through review and signature. These manual steps increase the likelihood of inconsistencies and slow the process.

Maestro is designed to automate this workflow. It allows teams to build workflows where data moves through the agreement lifecycle without repeated manual re-entry. One of the most common patterns we see our customers rely on is the Web Form-to-eSign sequence:

• Web Form: Captures participant data (Name, Address, Account Number).

• eSign Template: Receives that data to pre-populate a signature-ready document.

While the concept is simple, the configuration was, until recently, one of the largest bottlenecks in our ecosystem.

The "invisible field" problem

The manual mapping of fields from a Web Form to a document had become one of the most common sources of user-reported errors. Customers often spent a significant amount of time mapping dozens of fields one by one, only to find the final document arrived blank due to a broken connection.

When these workflows failed, debugging required an engineer to jump on a troubleshooting call and manually inspect the metadata. As more customers began using Maestro, what had once been an occasional frustration became a major bottleneck.

We asked: What if Docusign handled 90% of this mapping automatically?

Why simple matching failed

The natural instinct for automation is semantic similarity: comparing the label on a Web Form ("First Name") to the label on a template ("FirstName"). 

This works only if both sides have meaningful labels. Many eSign templates do not. 

When users build templates, they often skip labeling individual fields because the document's printed text (e.g., "Applicant Name") provides enough visual context for a human. However, the system sees that field only as Text_b7605 or Field_003.

We analyzed real-world template data and confirmed: meaningful labels are the exception, not the norm. To the AI, the input wasn't difficult; it was genuinely meaningless.

The solution: a two-stage AI pipeline

To solve mapping, we first had to solve identification. We built a pipeline that transitions from visual "seeing" to logical "reasoning”:

Stage 1: Spatial OCR & field labeling

When a user opens the mapping panel, Maestro triggers a series of API calls to retrieve the template’s X/Y coordinates for every fillable field. We then pass the document image and these coordinates to our Field Labeling API.

This API uses spatial analysis to "read" the document just like a human would. It looks at the text surrounding Text_b7605 and identifies the words "Authorized Signature" nearby. Suddenly, the system knows exactly what that field represents.

Stage 2: Orchestration via Docusign Agent Fabric (DAF)

With meaningful labels on both sides, we pass the data to a GPT-4o agent orchestrated via the Docusign Agent Fabric (DAF). The agent acts as the "brain," performing fuzzy matching that standard logic would miss:

• "Home Address" (Web Form) → "Address" (Template)

• "Full Legal Name" (Web Form) → "Name" (Template)

Sequence Diagram of Maestro's Two-Stage AI Data Mapping Pipeline

Sequence diagram of Maestro's two-stage AI data mapping pipeline

UX challenges: managing the "20-second" gap

The two-stage pipeline introduced a technical reality: latency. The full process takes roughly 18–20 seconds. If this triggered automatically, it felt like the UI had frozen.

We made a deliberate UX decision to put the user in the conductor’s seat and added the “Map Me” button. By making this a manual trigger, we changed the psychology of the wait. Waiting 20 seconds for a task you initiated feels like progress; waiting 20 seconds for a page to load feels like a bug.

A dynamic progress bar keeps the user oriented, providing transparency into the AI's "thinking" process.

Maestro's AI-Assisted data mapping interface in action, featuring the dynamic progress bar that keeps users engaged during document analysis.

Maestro's AI-assisted data mapping interface in action, featuring the dynamic progress bar that keeps users engaged during document analysis

Impact and the road ahead

Since becoming generally available in March 2026, early data confirms our hypothesis: when mapping takes 20 seconds instead of 30 minutes, more workflows make it to publication that might have otherwise been abandoned due to configuration fatigue.

"Removing the manual mapping step isn't just a convenience; it's the difference between a workflow a customer can build themselves and one that requires a support call."

Phase two will expand this capability to other common permutations, including Web Form-to-Web Form and eSign-to-Web Form. The underlying architecture is built to generalize. Anywhere there is a source and a destination, Maestro will be there to bridge the gap.

Try it yourself

Log in to your Maestro environment today and look for the "Map Me" CTA in your next Web Form-to-eSign workflow.