Photographers are among the most skeptical audiences for generative AI, yet a platform built for them just posted a durable adoption number that most enterprise AI pilots never reach. Pixieset, an all-in-one photography business service, launched an AI alt-text generator on Amazon Bedrock in early 2025 and reports that 35 percent of its applicable user base still uses the feature sixteen months after launch (AWS Machine Learning blog). In the first week alone the tool generated alt text for more than 750,000 photos, the team writes in the same post (AWS Machine Learning blog). The feature moved from concept to production in four months.
The skepticism Pixieset faced is not hypothetical. Photographers and artists have watched generative AI threaten their craft and flood their industry with synthetic work, the team notes at the opening of its post, which is precisely why a generic “add AI” feature would have backfired (AWS Machine Learning blog). Winning this audience meant choosing a task they already disliked rather than one they prized, a distinction that shaped every later decision.
The reason this matters is that it inverts the usual failure pattern. A 2025 MIT study cited by Pixieset found that 95 percent of enterprise generative AI pilots delivered zero measurable returns (AWS Machine Learning blog). Pixieset hosts more than 8 billion photos and is used by millions of photographers worldwide, so a feature that earns lasting use at that scale is a rare counter-example to the broader disappointment (AWS Machine Learning blog). The company also saw subscription upgrades in week one, which suggests the feature drove revenue rather than merely absorbing engineering cost.
The idea did not start in a roadmap meeting. Each year Pixieset runs a company-wide hackathon where teams experiment with new technology, and it was there that the team first saw what generative AI could do before bringing that curiosity back to an actual customer problem, according to the post (AWS Machine Learning blog). That origin matters because it set the framing: the question was never “what can this technology do?” but “where are our users losing time to tasks that are not creative?”
The problem Pixieset chose to solve was invisible scaffolding, not art. Most photographer websites had little to no alt text — the unique, accurate description for each image that makes it readable by search engines and screen readers — because writing one for hundreds or thousands of portfolio images is tedious and perpetually deferred, the post explains (AWS Machine Learning blog). Pixieset’s own product page confirms the feature is live today: its website builder advertises the ability to “automatically generate descriptive alt text for your website images” (Pixieset Website builder). Photographers knew they should add alt text to promote their businesses, but the sheer volume made it a task they kept postponing.
The cost of that neglect is concrete, not cosmetic. Without alt text, images are invisible to search engines, and a typical photographer’s portfolio contains hundreds to thousands of images that each need a unique description, the post points out (AWS Machine Learning blog). For a photographer whose business depends on being found online, that invisibility directly undercuts client acquisition, which is why the chore kept getting deferred rather than ignored outright.
The build itself stayed deliberately small. Pixieset runs an event-driven pipeline on Amazon EC2, AWS Lambda, and Amazon SQS; adding alt-text generation meant one more step that sends an uploaded image to Amazon Bedrock for inference by a multimodal large language model, then stores the caption for review in the website builder photographers already open daily, according to the AWS post (AWS Machine Learning blog). The integration was a single API call with no new infrastructure, GPU provisioning, or model hosting — which Graham Gibson, Staff Product Manager at Pixieset, names as the primary reason the team chose Bedrock (AWS Machine Learning blog). By using the fully managed service, Pixieset scaled from zero to 750,000 inference requests in the first week without provisioning a single server.
Reliability was treated as an adoption problem, not just an engineering metric. Because empty or failing captions erode trust and hinder adoption, Pixieset used Bedrock cross-Region inference to keep throughput up and added a fallback that retries a failed request on a second model of similar quality; the feature has had zero downtime since launch, the engineers report (AWS Machine Learning blog). The feature debuted on Anthropic Claude 3.5 Sonnet through Bedrock, selected for its multimodal image understanding, fast inference, and cost efficiency at scale (AWS Machine Learning blog).
The model choice also protected the team from lock-in. Because Bedrock exposes a unified API across providers, Pixieset could evaluate newer models without re-architecting, and the catalog has since expanded to include Claude Sonnet 5 as of June 2026, OpenAI’s GPT-5.5, Amazon Nova, and models from Meta, Mistral, and Cohere, the post notes (AWS Machine Learning blog). That GPT-5.5 is now available on Bedrock is independently confirmed by OpenAI’s own announcement that it brought GPT-5.5 and other capabilities to AWS, including on Amazon Bedrock (OpenAI on AWS). Advanced prompt optimization in Bedrock further lowers switching cost by adapting existing prompts to a new model automatically.
The discipline showed up early in what the team did not build. The same framing led Pixieset past flashier applications like AI-generated images, which would have encroached on the creative work photographers take pride in, the post explains (AWS Machine Learning blog). By ruling those out, the team protected the trust it needed for the metadata feature to land.
What kept photographers from rejecting the feature was control. Rather than applying AI captions across an entire site at once, Pixieset surfaces one suggested alt text at a time so a photographer can accept, edit, or reject it before expanding scope; only when comfortable do they switch on “auto-apply,” and every caption stays editable, the team writes (AWS Machine Learning blog). Ry Rainey, Staff Engineer at Pixieset, says the team talked a lot about how this first AI feature “must solve a real problem” rather than alienate the creatives it serves (AWS Machine Learning blog). That patience produced the headline number: 35 percent of applicable users eventually chose full automation on their own terms.
The interface reinforces that trust mechanically. Pixieset’s review screen shows a single suggested alt text per image with clear accept, edit, or reject options, so the photographer validates output quality on familiar turf before any portfolio-wide change (AWS Machine Learning blog). The team describes this as letting photographers build confidence at their own pace instead of being asked to surrender control on day one, which is how the eventual 35 percent figure was reached through product experience rather than assurance.
The reception confirms the team asked the right question. In feedback quoted in the post, one photographer called the “SEO for images” alt-text feature “great!! Saves me so much time,” and another said they updated their subscription to benefit from the innovation (AWS Machine Learning blog). The pattern — enthusiasm paired with a stated intention to keep reviewing the output — is exactly the trust the incremental rollout was designed to earn.
The lesson for builders is about triage, not tooling. Pixieset treated AI alt-text as a “must-have” table-stakes capability rather than a proprietary moat that deepens with unique data, which justified prioritizing speed to market over a year of differentiation that would not compound, the post argues (AWS Machine Learning blog). It also deliberately automated metadata — the invisible scaffolding that makes a portfolio discoverable — instead of anything touching the creative act, so users felt relieved of a chore rather than replaced, which is why the feature drives upgrades rather than cancellations.
That discipline is the real answer to the 95 percent failure stat. Most pilots collapse because they chase impressive technology aimed at work users value or fear, whereas Pixieset aimed at work users already wanted gone; the technology was almost incidental to the problem selection (AWS Machine Learning blog). When the chore is real and the control is retained, adoption follows even among the people least inclined to forgive AI.
For teams deploying on the same platform, the surrounding ecosystem has matured in ways that lower operational risk. OpenAI’s models and Codex are now available on Bedrock, and AWS has published operational guidance such as monitoring Codex usage on Amazon Bedrock with CloudWatch for teams scaling inference workloads (How to monitor Codex usage on Amazon Bedrock with CloudWatch). Security-focused model additions have also reached Bedrock, widening the range of workloads that can run there without leaving the managed environment (OpenAI adds Daybreak AI security models to AWS Bedrock).
The broader takeaway is that Pixieset’s result came less from model sophistication than from problem selection. A single Bedrock inference call, dropped into an existing pipeline, delivered a feature with 35 percent lasting adoption — but the technical simplicity was the lesser lesson, the team concludes (AWS Machine Learning blog). The harder one is knowing which problem to solve, whose work to automate, and how much control to hand over at once. For any organization staring at a generational AI mandate and a pile of unused pilots, that discipline is the part worth copying.
