80% of Financial Firms Now Run AI. Only 40% See Higher Profit — Is the AI Reading Your Filings Actually Improving Your Portfolio?

Answer this before you read on. Over the past year you used AI to read financial statements, summarize news, digest annual reports. By how many percentage points did that raise your return?

No answer means you are standing exactly where every global financial institution is standing. A report has now put numbers on the spot.

The Cambridge Centre for Alternative Finance (CCAF), working with the World Economic Forum, published its global report on AI in financial services in late April 2026. It does not tell you "AI is here" for the ten thousandth time. It sets two figures against each other: the industry has adopted AI past the halfway mark, and most of it cannot prove the money bought anything.

I work in model risk at a Thai bank. The job is asking the annoying question — does this model create value, or does it just look good? The distance between using AI and getting paid for it is the distance retail investors are standing in without noticing.

1. The sample is large enough to carry the conclusion

This is not a small poll. The report draws on "628 respondent organisations" across "151 jurisdictions", and it covers every seat at the table — which is what makes it possible to see who leads and who trails.

Adoption is genuinely high. The report states that "More than 80% of financial services firms are adopting AI to some level" — more than 80% of financial firms run AI at some level. The sharper number is agentic AI, systems that execute rather than advise. It is no longer a future item: "52% are already experimenting with agentic AI" — over half have started.

Read that as: AI in finance has stopped being a question of whether. It is now a question of what comes back. That question is far harder, and most of the industry is still dodging it.

2. The gap: adoption is heavy, value is unproven

Adoption clears 80% and the profit line does not follow. "Only 40% of respondents report increased profitability from AI, while 43% report no change" — 40% report higher profit from AI, 43% report that nothing moved.

Read the ordering again. More firms see no change than see a result.

Part of why the value cannot be proven is that nobody knows how to measure it. The report finds "76% of respondents in large financial institutions finding it hard to measure the value" — 76% of large financial institutions admit that measuring what AI adoption is worth is hard.

This is the same failure I have hit in model validation for years, and it always has the same shape:

The "it feels better" loop — what runs inside 76% of large institutions

  New technology lands
        │
        ▼
  Deploy it now ─────────► ❌ No baseline recorded for what "before" looked like
        │
        ▼
  Output looks good (faster! more summaries!)
        │
        ▼
  Someone asks "was it worth it?"
        │
        ▼
  ❌ Nothing to compare against → the only answer left is "it feels better"

The model is not the problem. Nobody drew the starting line before switching it on. When the board or the examiner asks for a number, all that is left is a feeling.

3. Where this becomes your portfolio's problem

The instinct is to file this under large-institution problems. Wrong read. The mechanism that breaks is identical. Only the arena and the size of the cheque change.

Institutions pay millions in licence fees and cannot locate the ROI. Retail investors pay in time and belief — AI summarizes the financials, summarizes the news, summarizes the annual report, every day. Ask directly whether it improved returns and you give the same answer the 76% give: you can't say.

That last row carries the whole argument, and it needs unpacking.

4. Why general-purpose AI never hands you an edge

Edge in equities comes from one thing — knowing something the market does not know yet, or knowing it before the price absorbs it.

So trace the answer back. Ask a general chatbot whether a stock has good fundamentals and it assembles the reply out of whatever has been written about most on the internet. What comes back is the average of the opinions everybody has already read. It summarizes consensus brilliantly. Consensus is already in the price.

Generic black-box AI
  Question → assembled from whatever gets written about most → "a tidy consensus"
                                                              └─ already in the price = zero edge
                                                              └─ and no way to know which page the number came from

Specialist AI built to produce edge
  Question → search the documents that exist and nobody opens (every 56-1 / MD&A in the market)
           → pull the verbatim sentence the company wrote about itself
           → checkpoint: is this sentence actually on that page
                ├─ pass → show it with a link to the SEC original
                └─ fail → suppress it (no smooth guessing, no tidying up)
           → you read the original and decide yourself

The difference is not which model is smarter. It is which pile it reads: the one everyone is fighting over, or the one nobody has time to open — a 56-1 runs past two hundred pages per company per year, multiplied by every company on the exchange. That is the back pile, and that is where the edge is still sitting.

5. Three risks the report flags, translated into portfolio terms

The report is not optimistic. It names risks any data or risk practitioner knows cold, and all three have a retail-sized version.

Data quality first — "Data availability and quality remain the leading pain point hindering AI adoption, cited by 66% of AI vendors": 66% of the people selling the technology say the available data is not ready. Security second — "48% of respondents flagging adversarial AI as a top concern". Third, the one most readers skip, concentration: the report finds "OpenAI the most-used foundation model provider across all groups (76% of industry and 48% of regulators)".

The last row gets the least airtime and bites hardest. If your tool is the market's tool, however smart it is, you have bought a faster average. Not an edge.

6. Regulators are behind — which makes you your own governance function

The report also hammers the gap between the players and the referees. Of the 130 regulators surveyed, "48% of the 130 regulatory authorities surveyed reporting they are 'still in the exploring' stage" — nearly half of regulatory authorities are still at the exploring stage while industry has already run ahead.

In Thailand, the Bank of Thailand has issued AI risk management guidelines for financial institutions — but that governs financial institutions. It does not govern the chatbot you open at 2am to ask whether a stock is worth buying.

Read that as: while the rules lag, nobody is checking whether the AI you use is making things up. You are the user and the examiner at once. That sounds heavy. It reduces to a single question — where did this number come from, and can I open the source document?

7. A model risk test: which AI is worth keeping

After finishing the report I cut it down to three tests for whether an AI tool deserves continued use. They are the tests I apply when validating credit models at the bank:

1. Measurable from the very first use — not "check back in six months to see whether it paid off"
2. Solves a named problem — one task, with a clock on it, not "helps with everything"
3. Fails loudly — a point where the system will say "I can't prove this" instead of guessing smoothly

That is why I did not build another conversational chatbot. I built Boom Leverage Terminal as a specialist tool, against one brief from day one:

Cut the time to find the risks buried inside a 56-1 from roughly three hours of reading per company down to seconds

I use that as the example of AI with clear day-one ROI not because the number is pretty, but because you can verify it yourself on the spot — time yourself hunting the risk section through one 56-1, then time yourself searching and reading the actual passage the system pulls. That difference is ROI you measured. You do not have to take my word for it.

If you want to see how meaning-level search differs from hitting Ctrl+F, I broke it out in a piece on semantic search over MD&A. If the 56-1 document itself is unfamiliar, start with how to read a One Report / 56-1.

8. Three pieces of homework for investors already running AI

• Set the measure before you start, not after you finish — if 76% of large institutions still cannot measure AI's value, anyone who clocks and logs from day one lands in the small group with an immediate advantage. Start simple: which task does AI help with where you save minutes you can actually count
• Separate convenience from advantage — AI that makes the news easier to read is convenience. Useful, and everyone has it. AI that walks you to a line the company wrote about itself that nobody has picked up yet is advantage. Do not pay the price of the second to get the first.
• Never accept a number you cannot click through to the source — this one rule prevents more damage than any other. My own system burned me: it handed me a beautiful table containing exactly one real number, the rest "filled in to complete the grid" (full case in the piece on the fake-number checkpoint)

Bottom line

The CCAF 2026 report does not say AI failed. It says the financial industry is fast to adopt, slow to benefit — 80% using it, only 40% seeing higher profit, 76% still unable to measure the value.

The same gap has two faces. The risk face is what happens when AI agents start moving real money. The measurable face is what I built when I put together a text-driven early warning system inside a bank — the problem was never the model. It was setting the measure and proving the value in a form somebody else can audit.

For retail, the lesson collapses to one sentence: don't use AI to feel like you did the homework. Use it to see what others haven't seen. If it cannot do the second, all it does is make you average faster.

Try it: Boom Leverage Terminal searches every MD&A/56-1 in the market by meaning and returns the real passage with a link to the SEC original — start free, 10 credits/day, no card required. Take a stock you hold right now and search one phrase — "ความเสี่ยงสภาพคล่อง" [the cash is tight enough that they had to put it in writing] or "การพึ่งพาลูกค้ารายใหญ่" [the revenue line hangs on a handful of buyers, any one of whom can walk] — and time how many seconds it takes to hit a line you have never seen. · Team and institutional (seats · Excel export · API): the Enterprise page or contact@boomleverage.com

This content is for education and sourced analysis. It is not investment advice.

Sources

• Report finds uneven AI adoption in financial services — Cambridge Judge Business School
• Report Finds AI Integration Gap and Risks in Financial Services — Traders Magazine
• Cambridge report finds AI integration gap and risks in financial services — Cambridge Network