The model is no longer the bottleneck.
KPMG's 2026 AI in Finance survey names slow adoption, not capability, the top barrier to ROI. The model is no longer the bottleneck, the collaborative layer is.
Hi folks,
KPMG's 2026 AI in Finance survey landed on Monday. 93% of US finance teams will scale AI within 18 months, and the top barrier among those not getting ROI is "slow organisational adoption and change management." My agent-testing week says the same thing in a different vocabulary. The intelligence is here. The collaborative layer is the work.
The one thing
The hard part of AI in finance is no longer the model. It is everything around it.
- KPMG surveyed 1,013 senior finance leaders across 20 countries. 93% of US firms will deploy or scale AI in finance within 18 months. Half are planning multi-agent systems. (KPMG)
- For the minority unsatisfied with ROI, the top barrier is "slow organisational adoption and change management". KPMG's own framing: "AI success depends as much on managing people as it does on managing technology." (KPMG executive summary)
- Biggest obstacles to better day-to-day AI use: lack of clear, role-specific use cases (64%) and hands-on practice environments (61%). Both are firm-level work, not model work.
What this means in plain terms.
- Anthropic's finance agents (launched 5 May) already produce what an AU mid-market analyst would, in minutes. More below.
- The bottleneck moves to everything around the model: output format, prompt-level constraint, the fund's own data, review discipline, who actually adopts.
- Every one of those is buildable, measurable, and the actual wedge.
In the mix
- Blackstone and KKR in early talks with Google to give portfolio firms access to Gemini. Same shape as Anthropic's $1.5B JV with Blackstone, Hellman & Friedman and Goldman the day before. AI is becoming a portco-benefit at the largest managers, not only a deal tool. (Bloomberg)
- Anthropic's Agent SDK starts metering on 15 June. Programmatic Claude usage moves off Pro and Max subscriptions into a separate metered credit pool at API rates. If anyone on your team is running automations off a personal account, the cost shape changes in four weeks. (InfoWorld)
- FSB's first global report on private credit vulnerabilities. Headline framings: limited transparency on credit quality, growing retail exposure, interconnections with banks. FSB language tends to seed national regulator phrasing over 6 to 12 months. Worth watching whether APRA or ASIC pick it up. (FSB)
From my week
I ran a structured test of Anthropic's new finance agents against AU mid-market private credit inputs. Two days in, the pattern is clear.
- Day 1: Pitch Agent built a 30-slide AU IPO roadshow deck in 16 minutes. AU framing was total: BBSW, AFSL, AMIT, ASIC, real comparables, real sources. When challenged on a chart, the agent admitted half the numbers were fabricated. Its own line: "Listing them in a verification flag list after the fact doesn't fix it. Those numbers are now sitting in a polished deck that looks authoritative." The polish travelled. The flag list stayed in chat.
- Day 2: Market Researcher produced two 15-page analyst notes in 13 minutes. No fabrication. The difference wasn't the model. It was two collaborative-layer choices: a markdown output format that carries
[unverified]markers natively (a PowerPoint slide cannot), and one prompt-level instruction, "hard gaps only, no estimates." Same agent. Different scaffolding. Different output. - One AU GP we are scoping with named the same dynamic this week. He said the hardest part of his AI build is the roll out to the org. The intelligence layer is the easy part.
Levercon is the AI team for investment funds. We embed alongside your team to automate workflows, build custom tools, and compound operational gains. Ultimately we help Australian based investment funds leverage AI.
Book a strategy call.
See where AI fits in your fund.
We'll show you where AI lifts efficiency and performance across your organisation.