London AI firm V7 expands from image data labeling into workplace automation

V7 Labs, an AI company known for producing tools that help companies label and organize data for computer vision tasks, is expanding into the competitive arena of business process automation.

Today the company announced it was debuting a V7 Go, a software platform that relies on large AI models to automate white collar business tasks.

The company said the platform is already being used by a number of customers, including an asset management firm that is using it to extract data from confidential memos; a biotech company that is using it to mine information from scientific research papers; and a freight company that is using it to automate the processing of shipping documents.

Many consider business process automation the next big category poised to be disrupted by generative AI, as the technology moves from simply creating new content—such as composing emails or images—to carrying out the steps needed to complete a task. A growing number of venture-capital-backed startups are trying to create AI software that can act as an agent for people, carrying out tasks such as scheduling or purchasing. Adept AI, Elemental AI, and Interloom are among the others trying to grab a piece of the market.

Alberto Rizzoli, V7’s cofounder and CEO, told Fortune that the company decided to move into workplace automation because of its potential market size and after discovering that many of V7’s existing customers had started using V7’s existing software, called Darwin, to automate tasks beyond just labeling image data.

The existing market for business process automation was estimated at $10.2 billion in 2023, according to figures from Cognitive Market Research. But many see the existing players in the space—which include so-called robotic process automation firms, such as UiPath, and low-code application builders, such as Appian—as ripe for disruption because their existing software is not based on generative AI models.

Rizzoli said that many businesses want to use AI to automate aspects of work but lack the machine learning engineers to do so. So V7 wants to build simple software tools that essentially use AI to build AI, creating systems that can reliably carry out tasks.

With V7 Go, users can drop the files they want to use into the software platform and tell the software in plain English what to do, Rizzoli said.

The software uses AI models to automatically create a searchable tabular index—essentially a spreadsheet—of the information contained in the files. It also creates a spreadsheet-like interface listing the steps it intends to carry out in order to complete the task.

In one example, Rizzoli dropped a 40-page engineering report about a diesel engine along with the questionnaire for a spec sheet with 30 questions into V7 Go. In seconds, V7 Go completed the questions in the spec sheet automatically from the information found in the document. The software’s process can then be edited by a user through the spreadsheet-like interface to improve the steps the AI has devised for completing the task.

Simon Edwardsson, V7’s cofounder and chief technology officer, says that the company has tested its AI-powered information extraction engine on 1,000 documents and compared its accuracy against two other document processing software tools, one from Instabase and the other from Eigen Technologies. V7 Go was 100% accurate compared to 66% for Instabase and 42% for Eigen, he said.

Rizzoli also said that in its own tests V7 Go was 34% better than using OpenAI’s GPT-4 model alongside retrieval augmented generation (or RAG), where a search is performed on a specific database and then an AI model is asked to summarize or reason about the results of that search.

Under the hood, V7 Go uses a series of AI models to extract information from files, including computer vision AI that identifies specific objects in images and language-based models that can understand text. Large language models sit on top of these more tailored AI models, developing the steps needed to complete the task. The company said that users can choose to swap in different LLM models from companies such as OpenAI, Anthropic, or Cohere to power some of these processes.

Tuza, a U.K. company that helps businesses adopt payment solutions, was among V7 Go’s beta testers. Its CEO, Ed Hardy, says that product saved the company weeks of engineering time by allowing commercial teams to quickly build workflow automations without needing any help from machine learning experts. “Any basic Excel user can leverage Go from sign-up,” he said.

But another early beta tester, Francesco Bellanca, who had tested the V7 Go when he was head of product for Paris-based business media company Sifted, said he found V7 Go’s spreadsheet-like interface unfamiliar and difficult to use. He said he wanted an interface that supported more free-form, flowchart-like process descriptions and which also allowed processes that only take place if certain conditions are met. Rizzoli said the version of V7 Go the company is releasing today has been updated to support the kinds of conditional logic requirements Bellanca mentioned.

V7, which is based in London, has about 350 customers for its existing image labeling software platform, according to Rizzoli, including candy company Mars, biotech company Genentech, and forklift manufacturing firm Kion. The company initially became known for helping biotech and medical companies develop AI software to automatically classify images of cells and classify medical images.

It employs about 80 people and raised a $33 million Series A venture capital round in 2022 that was led by Radical Ventures and Temasek, the Singapore government-backed fund. Other investors included Air Street Capital, Amadeus Capital Partners, and Partech.

This story was originally featured on Fortune.com