Money is energy.
Now AI moves it.
I am lead author of the Agentic Finance Report (2026), co-published with TensorX, AMINA Bank, the Solana Foundation, APEX:E3 and the Cardano Foundation. This page is my own summary in my own voice, not a copy of the report: the definition, the five preconditions, and the chapters I wrote myself. For the report itself, read it at agenticfinancereport.com.
inside a mandate.
I define agentic finance, in the report, as capital managed continuously rather than periodically, by software that both decides and acts, on-chain, inside mandates that people write and can withdraw, on rails that settle at machine speed.
The distinction that matters is between an assistant and an agent. An assistant proposes, and a person still executes. An agent acts directly, inside limits fixed in advance. Move that boundary out of a system prompt and into infrastructure, custody, a policy engine, a signing key that structurally cannot do more than it is allowed to, and the boundary holds even when the model is wrong, compromised or simply confused. Leave it in a prompt, and it was never a boundary.
Three conditions hold at once, for the first time: models good enough to be bounded and tested, continuous low-cost settlement, and assets in machine-addressable form. That is why I treat this as infrastructure already in production, not a forecast.
An assistant proposes. An agent acts. A boundary that exists only in a prompt is not a boundary.
— chapter 02, Agentic Finance Reporthave to be true.
The report sets these out as the preconditions for letting software hold, move and settle capital. I use the same five when I assess whether an institution is actually ready, not just interested.
- The asset is held by a regulated institution.Custody, fiat rails and a balance sheet that stands behind settled tokens.
- The rail settles finally and cheaply.Settlement inside the decision cycle, at a cost that makes the smallest useful payment economical.
- The reasoning runs where the institution can see and trust it.Inference inside the jurisdiction, retaining nothing, pinned and replayable.
- The decision is attributable to the agent that made it.Orchestration that turns many narrow agents into one decision a committee can take apart.
- The agent is identifiable as acting for a legal entity.A verifiable credential naming the entity and the authority it acts under.
that are mine.
The report has a foreword, eight chapters and two guest contributions. Each chapter carries its author's name, because the point is that the people who run each layer wrote their own. Three are mine: the executive summary, chapter 02 and chapter 08.
My starting claim is that AI, blockchain settlement and tokenised assets have merged into one system. A person in the loop is no longer a physical necessity, but accountability does not relax: enforcement moves from review after the fact to a boundary set before the agent acts, namely the mandate, the identity it carries and the rails it may use. Three conditions hold at once, which is why this is a live deployment question rather than a forecast. Most AI programmes fail on the organisation, not the model, so the loop I set out is a workflow redesign with governance built in.
This is where I set out the framework the rest of the report uses. It draws the line between an assistant that proposes and an agent that acts, then walks through the five agent classes a treasury will meet: treasury, yield, compliance, collateral and payments. Compliance gates every other agent’s action and is never optional, rather than sitting as one class among the five. Every agent runs the same loop: signal, reason, decide, comply, execute, report, with compliance checked before execution and every step reported immutably afterwards. On top of that sit three governance tiers: operational actions an agent runs fully autonomously, tactical decisions it proposes for a person to confirm, and strategic choices, mandate design, the risk framework, governance itself, that stay exclusively human. The chapter's underlying argument is that most AI programmes fail on the organisation rather than the model, and that a mandate is a map of the jagged frontier drawn for one institution, not a generic policy document.
The closing chapter states what the report has established, what is already in production and what is still forward-looking, and sets out six actions across three horizons. The first needs no technology at all: write the mandate and set the tiers. It also names the six risks a committee will raise, regulation still moving, authorisation limits that must live in infrastructure rather than in a policy document, prompt injection upstream of the signature, correlated agent behaviour, keys and custody and issuer risk, and chain and model concentration, with mitigations for each. The line I keep coming back to is the closing test: build it so that you can answer for it. What no supplier can deliver is the mandate itself.
| # | Chapter | Author |
|---|---|---|
| FW | Build It So That You Can Answer For It | Tim Grant, TensorX |
| 01 | Executive Summary | Marcus Maute |
| 02 | What Is Agentic Finance? | Marcus Maute |
| 03 | The Regulated Banking Foundation | AMINA Bank |
| 04 | Why Solana Is the Execution Layer | Solana Foundation |
| 05 | Sovereign AI Infrastructure | TensorX |
| 06 | ALICE: The Orchestration Layer | APEX:E3 |
| 07 | Verifiable Identity, the Trust Layer | Cardano Foundation |
| Guest | The Attack That Never Breaks a Rule | Blindsight |
| Guest | Owning the Frontier Before It Is Priced | CV VC |
| 08 | Conclusions & Recommendations | Marcus Maute |
Maute, M. (ed.), with TensorX, AMINA Bank, Solana Foundation, APEX:E3 and Cardano Foundation; guest contributions by Blindsight and CV VC (2026). Agentic Finance Report. Industry report, edition 1.0. agenticfinancereport.com
For the report itself, its full chapters and its co-authors, go to agenticfinancereport.com. For who I am and what else I work on, see the about page. For occasional writing on energy, money and organisations, see the blog.