Welcome to AI for FIs, from Dixon Strategic Labs. Each week, this newsletter curates critical developments in agentic AI and explains why they matter for your FI.
Forrester reported this week on the companies that have gotten AI agents into production, and what sets them apart: they coordinate their agents, rebuild the work around them, and give each one its own identity. Gradient Labs is one that did it, with an AI voice agent that runs whole collections calls for fintech customers.
Carnegie Mellon released a free tool that rates a company's AI program across eight areas and hands back a single maturity level and a short list of what to fix next.
Further down, agents are moving into credit decisions, drawing new tools built to watch them, and writing most of the code at the labs that build them.
Tech.eu · June 1, 2026
A Gradient Labs agent set up to verify documents for a loan application, step by step, with each tool call defined. The structure around the model is the harness. Source: Gradient Labs, Tech.eu.
Gradient Labs, founded by former Monzo staff, builds agents that run entire regulated workflows. A Lending Agent handles a borrower from a missed payment through collections calls to a repayment plan. A Disputes Agent works a case from intake to chargeback. A third runs identity and document checks.
The agents work across channels, including voice, with compliance rules for FCA Consumer Duty and the EU AI Act built into the system. Gradient says its voice agent is running in production at scale, handling hundreds of thousands of calls a month for fintechs. Its latest funding is aimed at expanding into the US.
Gradient Labs wraps a general-purpose model in a harness that makes it dependable enough for regulated work.
Quick primer: What is a harness?
The model, a system like Claude or Gemini, is the brilliant new hire on day one. They are capable but unproven. Left alone, they forget what happened an hour ago and have no idea which rules apply.
No FI would hand that person the phones and full account access on day one. First come the training, the limited access, the checklists, and a manager reviewing the work.
The harness is all of that, built in software around the model.
It turns raw capability into a worker a credit union can trust with a member's money. The model can be swapped for a smarter one next year. The training and the guardrails are the part that took months to build.
A model running without a harness turned up in the UK this year. West Midlands Police used raw Copilot output in its briefing materials, and it invented a football match that never happened. That invented detail went into the official intelligence behind the fan ban, because no one checked what the model made up against what was real.
A good harness is a good thing.
Why it matters: Gradient is a working proof that hard, regulated, member-facing work (lending, disputes, identity checks, collections) can run in production. It is one vendor at a handful of fintechs, but it’s work most institutions still assume only people can do.
Forrester · June 3, 2026
Forrester's chase-and-catch framing: agentic technology races ahead while enterprise trust and governance lag behind. Source: Forrester.
Forrester's State of Agentic AI 2026 reports that three-quarters of enterprise leaders are adopting agentic AI, while only a small minority run it in production, past what Forrester calls "agentish" chatbots.
Forrester names three concrete things the companies that reached production did differently:
Set up how the agents work together first. Before adding agents, they built a shared list of every agent and clear rules for passing work between them, so the new agents and the old software work together.
Rebuilt the work around the agent. An agent added to a process built for human speed only saves a few steps. The bigger wins came from rebuilding the process to fit the agent.
Gave every agent its own identity, like an employee. Its own credentials, access to only what it needs, a full log of what it did, and a named person who owns it.
Forrester also names what stops most companies. More than half say agents are sprawling past their governance, even with the tools like the NIST AI risk framework in place. And logging every action for an audit costs too much.
Why it matters: By Forrester's count, most companies are stuck running pilots, and only a small minority have agents in real production. What separates the few is the governance work above.
Carnegie Mellon SEI · June 8, 2026
The five levels of maturity, from experimenting at the bottom to running AI reliably at scale at the top. Source: Carnegie Mellon SEI.
Carnegie Mellon's Software Engineering Institute, the group behind CMMI, the maturity standard software teams have used for decades, released an AI Adoption Maturity Model with Accenture on June 8. It targets the gap the Forrester piece describes: knowing when AI work is ready to leave a pilot for real use.
The model sorts an organization onto a five-step ladder, from just experimenting at the bottom up to running AI reliably across the whole business at the top. An organization rates itself across eight parts of the business, like strategy, staff, data, risk, and operations, lands on a level from 1 to 5, and gets back a ranked list of the gaps to close first. The top step is not meant for everyone, so the model has each organization set a target level that fits its business.
A company is mature when the AI it runs is reliable, well-governed, and repeatable. How much AI it runs matters less than how carefully it runs it.
Why it matters: Many leadership teams can only guess at how far along their AI is. This settles it with a credible benchmark.
Accenture and SEI walk through the model in a free webinar today, June 9, at 1:30 p.m. Eastern.
📡 On the Radar
Experian built an Agent Operating System into its “Ascend” credit platform. AI agents will make credit decisions and run fraud checks on it, starting with early adopters later this year.
AI fintech Aveni raised £12 million to build tools that check how AI agents treat customers. Watching what agents do and keeping them on their best behavior is becoming its own product.
Microsoft launched Scout, an always-on agent across email, files, and chat in Microsoft 365.
A senior OpenAI employee told the Financial Times that "chat is dead", and OpenAI is rebuilding ChatGPT around agents that take action (like personal financial management!).
Anthropic said a coordinated pause on self-improving AI would be a good thing, reporting that Claude now writes more than 80% of the code merged into Anthropic's own codebase.
In a moment where agentic AI is rewriting operational rules, I help credit union leaders use hands-on experimentation to sort out what this means for their strategy. Drop me a note at [email protected].
How this newsletter is made: Brent curates the research and writes the analysis. Claude helps with drafting and editing. Published on Beehiiv. ⚡ Alakazam ⚡.
Know someone who should be reading this? Send them the subscribe link.
