‘Lawyer-in-the-loop’ startup Wordsmith wants to bring AI lawyers to all employees

While Harvey AI aimed at lawyers themselves, Wordsmith is aimed more at the employees of a firm, with legal teams configuring the platform behind the scenes by connecting it to all their own data sources. Attorneys remain available when needed.

McNairn draws comparisons to the likes of TravelPerk, which gives SMEs a self-service business travel management platform that allows managers to define policies and approval processes. Employees make all their own reservations within these parameters.

“At TravelPerk one of the big steps [we made] was that we went from trying to accelerate the travel team by selling them slightly better tools, to essentially allowing the rest of the business to self-book,” McNairn told TechCrunch. “And then the travel team just managed and checked and made sure it was properly calibrated. And that shift of building tools just for operation, instead of building tools for the rest of the business to work more efficiently, is a huge change in the way you work.”

Companies can configure Wordsmith in two main ways: as an autopilot for simpler issues that don’t need expert oversight, and as a co-pilot, where a lawyer is always ready to give their stamp of approval before formal answers are given.

A typical workflow might involve someone in sales needing to vet a new contract, or perhaps a supply trying to close a deal who needs access to information like the company’s security posture — the kinds of questions that are fairly typical and where are the answers not much is likely to change. By asking Wordsmith, anyone can get the necessary information.

Other potential use cases might include someone issuing a company with a subject access request (SAR), whereby businesses in some jurisdictions are legally required to comply with requests related to access to personal data. In this case, Wordsmith could be configured to accept a submission and connect to a company’s ticketing system and respond with either the requested information or a response template outlining timelines and next steps — whichever if required by a company’s internal guidelines and procedures.

Model behavior

Wordsmith uses a combination of foundational large language models (LLMs), including OpenAI’s GPT-4 and Anthropic’s Claude.

“We use the right thing for the job,” McNairn said. “Some are very good at analyzing things like the logic within legal agreements, and some are very good at being extremely precise in helping us change language. Claude is very good at rationalizing through problems and OpenAI (GPT-4) is just fantastic, with different dimensions.”

Businesses have shown little reluctance to embrace genetic AI, which McNairn said the company is tackling in different ways. This includes allowing companies to specify that their data does not leave the EU. It also promises not to train its AI on companies’ data. Wordsmith configures a “private instance” for businesses, meaning it connects to data wherever they are (eg Google Drive or Notion) to refine a response using a company’s data, but that data isn’t used to model training for other companies.

“We use a technique called RAG (recovery augmented generation),” McNairn said. “So we don’t train on their data – we just use it when necessary. We recall it, use it to enrich the answer, and then give them an answer.”

High frequency

While strengthening in-house legal teams will initially be Wordsmith’s main focus, the company is also looking to work with law firms, as evidenced by its early tie-up with DLA Piper. In this case, DLA – a global multi-billion dollar legal force – is co-developing AI agents in partnership with Wordsmith, with a view to distributing it to its own clients.

In fact, they bring in their own technical knowledge to improve Wordsmith for very specific legal areas. It could become something they can sell as a new type of legal service, possibly at a lower rate.

“It’s higher frequency and lower cost to engage with business knowledge that way, rather than paying thousands of dollars an hour,” McNairn said. “Of [also] a much better way to show that they are progressive and want to adopt AI.”

This business model could work particularly well for small to mid-sized law firms, where Wordsmith could be engaged to take on larger jobs or take on more clients.

McNairn says that while this offering is still in the early stages of planning with DLA, Wordsmith will likely commercialize it soon. “It’s just not there yet,” he said.

With $5 million in the bank, McNairn says Wordsmith will now ramp up its recruitment in both Scotland and the US. The company has nine employees today, and while some are based in London and/or in the process of moving up, McNairn says he is keen for Edinburgh to become the company’s center of gravity.

“It’s the ecosystem thing that I’m very passionate about,” he said. “There were three unicorns I was involved in before this, and I just want to build something nice in Scotland.”