3264.ai
Book an audit

Essays

Field notes for the executives, operators and engineers building AI systems at banks, fund managers and asset managers.

Featured essays

No. 11Strategy7 min read

Why document extraction pilots stall at 80% accuracy

The demo reads the clean page. Production reads the scanned schedule, the amended definition and the table that breaks across two pages. The last fifth of the accuracy is where the work, and the risk, sits.Read the essay
No. 03Strategy8 min read

A procurement checklist for AI vendors, from the build side

The model is the one component every vendor can buy. The harness around it decides whether a system survives an audit, a model retirement and a bad quarter. These are the questions we would ask if we were sitting on your side of the table.No. 12Operations7 min read

Instrumenting a workflow before you automate it

The best dataset for automating a back-office decision is the record of people making it. Most firms throw that record away. Capture it first, and the automation that follows will be cheaper, better tested and easier to justify.

All essays

15 essays
  1. No. 01Engineering8 min read

    An index is not understanding

    Structured maps of a credit agreement help an AI system find the right clause faster. They do not make its reading right. Build them as tools the system keeps using, not documents it reads once.
  2. No. 02Governance8 min read

    Twelve rules for AI agents that touch money

    In lending and fund operations, a wrong answer becomes a wire, a missed breach or a sentence in an investor letter. These are the design rules we apply whenever a model sits anywhere near that path.
  3. No. 03Strategy8 min read

    A procurement checklist for AI vendors, from the build side

    The model is the one component every vendor can buy. The harness around it decides whether a system survives an audit, a model retirement and a bad quarter. These are the questions we would ask if we were sitting on your side of the table.
  4. No. 04Product7 min read

    Chat is an input, not an interface

    Language is the fastest way to say what you want. It is a poor way to show a controller what a system did. Operations software needs both halves, and they call for different designs.
  5. No. 05Engineering7 min read

    Draft, check, repeat: review loops that converge

    Pairing a model that drafts with a model that critiques can lift the quality of AI output. Without a defined bar, checks that run in code and a rule for stopping, the loop mostly burns tokens.
  6. No. 06Strategy7 min read

    When the people who run the process can change the software

    Operators can now describe a tool and have a coding agent build it in an afternoon. The speed is real. In a regulated firm it needs a workshop, a gate and a record, or the afternoon's work becomes next year's audit finding.
  7. No. 07Engineering7 min read

    Define done before you build

    Prompts drift and model runs vary. An ordered list of events, each with a proof that can be checked against the real system, gives an AI workflow a finish line that holds across rewrites, model changes and restarts.
  8. No. 08Engineering7 min read

    A workshop, not the vault

    AI systems stall when they cannot provision, test or verify their own work. Handing them production access is the wrong fix. Give them a complete environment of their own, and let only reviewed changes cross into yours.
  9. No. 09Engineering7 min read

    Flexible where it's safe, fixed where it matters

    Composable tools let an AI system choose its own path through an investigation. A money path should have no path to choose. Designing the tool layer for credit and fund workflows starts with knowing which is which.
  10. No. 10Engineering7 min read

    What coding agents teach us about back-office automation

    Coding agents became genuinely useful before most enterprise agents did. The reasons are structural: a tight loop, tools that fail loudly, and an environment full of checks. Operations teams can borrow all three.
  11. No. 11Strategy7 min read

    Why document extraction pilots stall at 80% accuracy

    The demo reads the clean page. Production reads the scanned schedule, the amended definition and the table that breaks across two pages. The last fifth of the accuracy is where the work, and the risk, sits.
  12. No. 12Operations7 min read

    Instrumenting a workflow before you automate it

    The best dataset for automating a back-office decision is the record of people making it. Most firms throw that record away. Capture it first, and the automation that follows will be cheaper, better tested and easier to justify.
  13. No. 13Engineering7 min read

    Generated interfaces need schemas

    Models can now compose screens, reports and dashboards on request. In a regulated firm, the safe version is a model that fills a typed layout from an approved set of components, bound to reconciled data it never types itself.
  14. No. 14Engineering7 min read

    Many lanes, one production

    AI features change prompts, data, tools and model versions at the same time. Shared staging turns every change into a queue. Give each engineer a complete lane of their own, and make the gate into production the only thing they share.
  15. No. 15Strategy7 min read

    When reading is nearly free

    The cost of having a model read a document keeps falling. In fund operations, the constraint is moving from how much can be read to how much can be verified and signed. Plan for that shift now.

Get new essays by email

One email when we publish. Nothing else.

Start

Talk to the team behind these essays.
A two-week audit starts with one process.

We price the build against the baseline the audit measures, and you leave with a ranked use-case ledger either way.

Book a two-week audit