Blend Labs (BLND) Autopilot AI Agent Goes Paid as Pull-Through Rises

Blend Labs, Inc. (BLND) gave its first customer-side results for Autopilot, its AI agent for loan origination, on its August 6, 2026 earnings call. Lenders using Blend Labs Autopilot are seeing a preliminary 10 to 15% improvement in pull-through rates, the product has started charging and six lenders have signed contracts, but the company has not disclosed any Autopilot revenue [1].

What Blend Labs does and where Autopilot fits

Blend Labs, Inc. (BLND) is a cloud software company serving financial institutions: banks, credit unions and independent mortgage lenders. Borrowers fill out loan applications and upload documents on Blend's platform, and the lender's staff review and advance the file in the same system. The business covers mortgages, home equity loans and retail banking products such as account opening and consumer loans. Blend charges software fees mainly based on how much customers use the platform.

Autopilot is an AI agent embedded in that platform, and its users are the lender's review and operations staff. It reads the documents borrowers submit, runs calculations such as income, and checks the full loan file item by item against investor and regulatory lending rules. When something is missing, it asks the borrower for it and moves the file forward. When it is unsure, it hands the file back to the lending team for a person to handle [1]. Mortgage origination is Blend's main business, and Autopilot sits directly on that line.

From a qualitative remark to a paid product in a year and a half

On the February 27, 2025 earnings call, management first mentioned a solution called Doc AI. It said the tool automatically validates documents and flags discrepancies in lenders' post-closing work, improving accuracy and significantly reducing manual workloads. It gave no customer count, volume or cost figure at the time [3].

On May 7, 2026, the same document-checking work was upgraded into Autopilot, which added the ability to retrieve rules, run calculations, update the file and trigger follow-up steps. Adoption figures appeared for the first time: as of May 4, 65 lenders had activated it and 22 were running it in production. The product was free for the whole second quarter, and the head of finance said the incremental contribution was early and was not included in any outlook [2].

On August 6, 2026, management announced that Autopilot had become commercially available on July 1 and presented results from customer use for the first time [1].

What the latest numbers show

The figures fall into two groups. The first describes the scale of the trial. More than 65 lenders activated Autopilot during the four-month preview, and it has processed more than 45,000 cumulative loans since February [1]. Both figures come from the free period. The company did not say how many of the activated lenders converted to paid, and because activations already stood at 65 in May, that figure does not represent additions in the latest quarter.

The second group describes what lenders got from using it. Customers' pull-through, the share of applied-for loans that end up funded, improved 10 to 15%, and cycle time shortened by two to four days. Management separately estimates that Autopilot automates about four and a half hours of loan fulfillment tasks on average per loan [1]. Management called these preliminary data and did not provide the pre-improvement baseline, the sample size or the number of lenders behind them.

How the results could reach Blend's own financials

All of these results occur on the lender's side. For them to show up in Blend's performance, the relevant financial metric is Blend's platform revenue: contract revenue from Autopilot paid tiers and, later, revenue billed per funded loan. There are two paths, and both remain directional.

The first is direct fees. Six lenders, including the large mortgage servicer Onity, have signed contracts that include Autopilot; for now these are fixed-fee, one-year contracts [1]. The second is funded volume. The company said pricing will later shift toward billing per funded loan, so if customers' pull-through keeps improving, the same application volume would produce more funded loans.

The company has not disclosed Autopilot revenue, contract values or a per-loan fee. Commercial availability and all six signings came after the June quarter ended, and on both the May and August calls the head of finance cautioned investors not to put Autopilot revenue into their models for now [1][2]. On the cost side, management said only that Autopilot model costs were relatively low but growing in the second quarter, without quantifying them separately.

What is confirmed and what is not yet quantified

What can be confirmed is that lenders trialing Autopilot saw loans reach funding more easily, close faster and require less manual work, and that six lenders have signed and are paying for it. What has not been quantified is how much Blend itself gets out of it. Autopilot's paid revenue, contract values and the number of trial-period activated lenders that converted to paid have not been disclosed, so the size of the customer-side improvement cannot be converted into a revenue gain for Blend.

Application assessment

• Blend Autopilot Loan Origination | Business position: core business | Application stage: limited production | Coverage: single business unit | Value type: revenue growth

Sources

[1] Drillr · Blend Labs, Inc. (BLND) · 2026-08-06 · Earnings call

Quote: Based on our preliminary data of these loans that have gone through our system, our customers are seeing a 10 to 15% improvement in pull-through rates and two to four days of cycle time improvement.

[2] Drillr · Blend Labs, Inc. (BLND) · 2026-05-07 · Earnings call

Quote: As of Monday, May 4th, 65 lenders have activated Autopilot, 22 are running it live in production, and over 7,000 applications have already been touched by Autopilot since we moved to live production.

[3] Drillr · Blend Labs, Inc. (BLND) · 2025-02-27 · Earnings call

Quote: By automating document validation, flagging discrepancies and streamlining verification, it not only improves accuracy, but also significantly reduces manual workloads.