Implementation Is Not So Easy

I round up the most relevant AI-in-finance news - the deals being done, who’s rolling out what, and what’s actually working on the front lines

OpenAI took an equity stake in Thrive Holdings this week...

…and announced they'll embed engineering teams directly inside portfolio companies. HSBC signed a deal with Mistral to develop custom AI models for financial analysis and translation. Both moves suggest that signing a frontier license and rolling it out firm-wide only gets you so far.

New research from UC Berkeley surveyed over 300 organizations with AI agents in production. The findings: 47% complete fewer than 5 steps. 70% use off-the-shelf models with zero fine-tuning. Deployments are basic. Use cases remain limited. And getting real value seems to require either specific implementation work or domain-specific scaffolding (two core design principles we preach at DealSage).

In This Week’s Issue:

News Digest:

OpenAI takes stake in Thrive Holdings to purpose-build enterprise AI
HSBC signs deal with Mistral for custom banking AI

From The Trenches:

What production agents actually look like, why they’re still pretty simple and how to get more out of them

Other M&A News This Week:

Busy week for deals including: OpenAI/Neptune, ServiceNow/Veza, Anthropic/Bun, Rezolve AI/Crownpeak, Tata/Commotion

Other Cool Stuff I've Read or Seen:

Meta buys Limitless, Marvell acquires Celestial AI, Anthropic/Google TPU deal, OpenAI "code red"

News Digest

OpenAI Takes Stake in Thrive Holdings to Purpose-Build Enterprise AI

OpenAI announced it's taking an ownership stake in Thrive Holdings and embedding research, product, and engineering teams directly inside portfolio companies. The initial focus: accounting and IT services. Two industries generating hundreds of billions in revenue but still running on workflows that haven't changed in decades.

This isn't a typical enterprise sales motion. OpenAI employees will work inside Thrive's companies to build AI systems, streamline workflows, and accelerate adoption. If those companies succeed, OpenAI's equity stake grows. The structure aligns incentives around actual outcomes, not license fees.

Thrive Holdings told Reuters they were struggling with off-the-shelf AI solutions that couldn't handle specific enterprise tasks. Generic tools weren't delivering. OpenAI's response: move beyond API access and get involved in the implementation.

For OpenAI, the logic is clear. They want to prove enterprise AI can deliver ROI at scale, and one way to do that is to own more of the implementation.

The details:

OpenAI takes ownership stake in Thrive Holdings (terms undisclosed)
Embedded teams from engineering, research, and product divisions
Initial focus: accounting and IT services

Why it matters: OpenAI is moving beyond selling API access. The model here is closer to "let us help rebuild your workflows" than "here's your license."

My take: This is OpenAI leaning into what a lot of people building in enterprise AI already suspect. The value isn't just in the model. It's in the scaffolding, the domain expertise, the structured workflows that turn raw capability into actual output. If OpenAI is embedding teams inside companies to make their own tools work better, that tells you something about where the implementation barriers are. And it tells you something about where the value accrues. Perhaps not just to the model providers, but to whoever builds the last mile.

Just look at Palantir, whose model of delivering AI solutions into commercial enterprises has exploded in recent quarters. Something DealSage’s co-founder Ben was intimately familiar with 🙂

HSBC Signs Deal with Mistral for Custom Banking AI

HSBC has signed a partnership with French AI startup Mistral to develop generative AI models for financial analysis, translation, and document processing. The deal, financial details undisclosed, is designed to create purpose-built AI tools for specific banking workflows.

Mistral, valued at nearly €12 billion, is seen as Europe's best chance at competing with US giants like OpenAI. The Paris-based company has signed contracts worth hundreds of millions with BNP Paribas, Stellantis, and AXA. For HSBC, the partnership means custom models trained for tasks like crunching complex documents in multiple languages and making financing decisions faster.

The deal is expected to expand into client onboarding and anti-money laundering checks. HSBC CEO Georges Elhedery emphasized the partnership will "work together to develop generative AI models" rather than just deploying off-the-shelf tools. That language matters. It signals HSBC is investing in custom implementation, not just rolling out generic capabilities.

The details:

Partnership for custom generative AI models across HSBC
Use cases: Financial analysis, translation, document processing
Future expansion: Client onboarding, AML checks
HSBC already using AI in corporate/institutional banking and wealth divisions

Why it matters: European banks are hedging their AI bets with non-US providers. And they're investing in purpose-built models for specific workflows rather than broad enterprise licenses.

My take: The language in the announcement is worth noting: "work together to develop generative AI models." That's different from signing an enterprise agreement and hoping for the best. HSBC is committing to building custom models for specific banking tasks. This is what more deliberate AI adoption looks like. Specific workflows get specific tools, built for the task. It takes longer, but the outputs are more likely to stick.

From The Trenches

What Production Agents Actually Look Like

New research from UC Berkeley surveyed over 300 teams with AI agents running in production. The findings offer a useful reality check on where agent deployments actually are.

The headline: 68% of production agents execute 10 or fewer steps before requiring human intervention. Nearly half execute fewer than 5. Teams are deliberately keeping things simple.

❝

“80% of production agents use predefined static workflows rather than open-ended autonomy. Organizations deliberately constrain agents to maintain reliability."

That's not necessarily a limitation. It's a design choice. The teams seeing results with AI in production aren't building autonomous systems that run unsupervised. They're building bounded workflows with clear human handoffs.

A senior leader at a financial institution told me about their AI transformation. They'd rolled out generalist AI tools firm-wide. Everyone got licenses. Leadership expected impact. The reality? As one finance transformation leader put it: "You actually have to train your people to use it, otherwise they're going to treat it like a Google search." The team struggled to identify use cases beyond basic summarization.

The UC Berkeley study offers some context. 74% of teams depend primarily on human evaluation rather than automated testing. 70% rely on prompting off-the-shelf models without fine-tuning. 79% use manual prompt engineering. Simple, controllable approaches are the norm.

The challenge isn't always the tools. It's sometimes tool-first thinking. Firms deploy Copilot or ChatGPT Enterprise expecting transformation, but without clearly defining what problems they're solving. Different teams adopt different tools. No integration strategy. No structure to unlock more complex workflows.

The result can be tool sprawl and pilots that don't go anywhere.

The firms making progress aren't necessarily the ones with the most sophisticated tools. They're the ones that picked one workflow, built structure around it, and committed to making it work.

Start with the problem. Then figure out which tool helps.

Implementation Is Not So Easy

OpenAI took an equity stake in Thrive Holdings this week...

News Digest:

From The Trenches:

Other M&A News This Week:

Other Cool Stuff I've Read or Seen:

The details:

The details:

Keep Reading