Half of all enterprises are moving away from renting AI through closed APIs and bringing open-source models in-house, according to Hugging Face’s chief executive in a TechCrunch interview[1]. His argument is that companies are tired of building on rented land. If you run a wealth management firm, that sounds like a debate for software companies. It isn’t quite. The tools you use to draft client emails, summarise meetings and speed up paraplanning are changing underneath you.

For about three years, AI at work has mostly meant chatbots. The latest releases are built around agents instead: systems that don’t only answer questions but carry out multi-step tasks across your business applications.

Agents act on their own across your software

The newest models are designed to take a goal and work through the steps to reach it across different pieces of software, rather than waiting for someone to prompt each step.

Digital Applied reports that while 75 per cent of businesses say they’re interested in these agent workflows, only 11 to 17 per cent have deployed them, and 40 per cent of agent projects get cancelled over deployment problems[2].

The gap between ambition and delivery is wide. Connecting an AI to your CRM, your document system and your email needs a different architecture from buying chat licences for the team.

With an agent, you’re giving a system permission to act for you. That brings security and compliance questions a standalone chat window never raised. An agent with read and write access across the whole client database is one bad output, or one compromised account, away from overwriting compliance records.

Why some firms are moving away from rented models

The case for self-hosting is mostly about control. For the past few years the default was to send your data to OpenAI or Anthropic, and relying entirely on closed APIs carries platform risk: prices, terms and access can change on someone else’s timetable.

The prices are moving in the other direction for now. Anthropic recently cut the price of Claude Sonnet 5[3]. Even so, the Hugging Face argument is that firms care more about control than raw capability for their core systems, and that narrow open-source models running inside a firm’s own environment are cheaper to run and keep data where the firm can see it[1]. That’s a vendor of open models making the case for open models, so it’s worth weighing accordingly.

The level of tool should match the problem

Most failed deployments come from buying a complex platform for a simple problem.

If your advisers need help summarising meeting notes, that’s Level 1 (education): an off-the-shelf tool with proper security and good prompting will do it today. If you want those notes to update the client record and draft a follow-up email, that’s Level 2 (integration), connecting two or three tools with something like Zapier or Make. Only when you need a specific multi-agent workflow working over your own data do you get to Level 3 (custom build).

My read

I don’t think most wealth firms need to choose between rented and owned AI this year. They need to stop treating every AI tool as a chat licence. The firms that will look sensible in two years are the ones that started at the lowest level that solves the problem, kept every agent’s permissions narrow, and kept a qualified person signing off anything that reaches a client or the file. The ones likely to struggle are those that bolted a broad-access agent onto their CRM because a release announcement made it sound easy, and then found that their whole workflow depended on one vendor’s pricing.

This article is for informational purposes only and does not constitute regulated financial advice or a compliance opinion. Consult a qualified compliance professional for advice specific to your firm.

Sources

[1] TechCrunch, ‘Hugging Face’s CEO on why companies are done renting their AI’, TechCrunch, 10 July 2026. Available at: https://techcrunch.com/2026/07/10/hugging-faces-ceo-on-why-companies-are-done-renting-their-ai/ [2] Digital Applied, ‘Why Agentic AI Projects Get Canceled (and How to Ship)’, Digital Applied, 11 July 2026. Available at: https://www.digitalapplied.com/blog/agentic-ai-project-cancellations-gartner-40-percent-2026 [3] Anthropic, ‘Claude 3.5 Sonnet’, Anthropic Announcements, 1 July 2026. Available at: https://www.anthropic.com/news/claude-sonnet-5