Reliable, Smart Workflows for Finance, Wealth and Insurance

Karthic Rao, Perpetual15:21 · Oct 2026 · 3,037 views
Thumbnail for Reliable, Smart Workflows for Finance, Wealth and Insurance Watch on YouTube
TL;DR
  1. 1

    Regulated workflows should use deterministic control flow, evidence handling, and side-effect execution, with models limited to explicit judgment points.

  2. 2

    A planner turns a request into a typed semantic graph, while a deterministic compiler validates it and returns a valid plan, a clarification request, or a refusal.

  3. 3

    In an illustrative 20-decision run, 19 decisions use a smaller model and one uses a frontier model, creating a path toward privately hosted inference.

Summary

Karthic Rao argues that regulated workflows need both runtime judgment and deterministic execution. He uses an IRS W-9 request as an example: identifying the correct Harper household, gathering missing data, waiting for a client, collecting a signature, and preserving evidence can take days. A fixed flowchart cannot handle the ambiguity, while an improvising agent cannot satisfy compliance review. Perpetual's design assigns control flow, evidence, and external actions to a deterministic system. A model translates the request into a typed semantic graph, and a compiler checks it against the firm's action contracts. Plans can be approved, pinned, and executed by a state machine. Models make small typed judgments, such as whether a reply is complete or a document matches. External actions have their own approval gates, and uncertain outcomes go to manual review instead of automatic retries. Rao presents smaller models and private hosting as consequences of this architecture, while the training loop remains future work.

Key ideas
00:00

Regulated workflows need a small amount of intelligence inside a large deterministic process

Rao says finance and wealth workflows are complex, involve sensitive data, and carry serious consequences when policies are violated. A process cannot be scripted for every path, but most steps can follow a known state machine. Models are needed at a few judgment points, such as choosing a branch or interpreting a reply. The rest of the system should control execution reliably. He rejects the goal of using more tokens everywhere. His design uses intelligence minimally and on purpose, while leaving control flow, evidence, and side effects to deterministic software.

02:19

A simple W-9 request becomes a multiday workflow with ambiguity and evidence requirements

The request to fill out an IRS W-9 for Mr. Harper first requires finding the form and identifying the correct household when several Harper records exist. The form then requires decisions about taxpayer identity, W-9 versus W-8, tax classification, the correct name, SSN or EIN, exemption codes, and the signer. Missing values lead to a client email and a wait. A partial reply causes another follow-up. After completion, the form goes for e-signature, is filed, and the evidence about how each value was obtained is retained. The one-sentence request crosses systems, waits on people, and loops.

05:44

The reliability boundary separates deterministic execution from model judgment

Rao contrasts two failures. A fixed flowchart breaks when the request contains ambiguity, while an agent that improvises its way through the task cannot get compliance approval. His answer is to place a boundary between deterministic and judgment-based work. Determinism governs what runs next, what evidence exists, and actions that change the world, such as writing a record or sending an email. Intelligence interprets ambiguous instructions and client responses. A judgment can open a smaller plan, but that plan must be compiled, gated, and run deterministically again.

07:45

A planner and compiler turn an instruction into an approved executable plan

The planner is a model call with a narrow translation role. It emits a typed semantic graph containing intents, entities, facts, artifacts, events, constraints, and ambiguities, with each node grounded in the request. A compiler then processes the graph through deterministic stages without calling a model. It returns one of three outcomes: valid, needs clarification, or refused. The compiled plan is immutable and pinned by a hash. The compiler checks the plan against the firm's vocabulary and 42 typed action contracts covering forms, documents, email, e-signature, calendar, and human steps. Unsupported or malformed actions fail closed.

09:56

External actions require approval for the exact payload

A compiled W-9 plan can include preparing the form, rendering a PDF, drafting an email, waiting for a reply, checking the returned form, and filing it. Drafting may use a model, but sending the email is a separate action gate. A person must approve the exact email, and the approval is tied to a hash of its payload. The hash is checked again immediately before sending. The same pattern applies to signature requests and other external actions. Waiting and follow-ups are explicit plan nodes, with deadlines handled by rules rather than by a model.

10:55

Partial replies and follow-ups create new plans without changing the approved parent

When a client supplies only some requested values, a reply judge marks the response as partial. That judgment is provisional, so the step remains blocked until a person confirms it. Once confirmed, the follow-up runs as a new child plan. The original approved parent plan is not modified, and the workflow allows at most five rounds. This lets the system respond to new information without giving a model permission to rewrite an approved process. The grammar can also add required steps, such as rendering a PDF, attaching it, or judging a reply, even when the planner did not request them.

11:45

A state machine handles typed judgments and uncertain side effects

After approval, a state machine selects each step and records each result. Judgment steps use fixed-schema model calls for tasks such as composing an email, extracting fields, checking for missing values, judging a reply, or verifying a document. The returned answer is typed, and the state machine decides what happens next. Side effects have their own lifecycle, including preparation, dispatching, success, and an unknown outcome that goes to manual review. The system does not retry automatically when the external world gives an unclear result. A chat message can answer a question or approve a step, but it cannot alter an approved plan.

12:52

Typed logs support smaller models and a path to private inference

Rao says the remaining model work consists of small typed judgments, so many cases do not need frontier intelligence. In an illustrative run of 20 decisions, 19 use the flash tier and one uses a frontier model. He presents this as a design target rather than a measured fleet result. Each judgment is typed and logged, and human confirmations or rejections create evaluation data. The training loop is still forthcoming. Because every model call is checked against a schema, moving a call to a privately hosted model can be a configuration change rather than a rewrite. The stated goal is to make about 95% local inference viable.

"Reliability and intelligence are not opposites. It is up to you where you draw the boundary between them and that is the point of this talk deciding where that boundary lies."00:18
Who should watch
  • You are building AI workflows for finance, wealth management, insurance, or another regulated field and need a way to handle ambiguity without giving an agent unrestricted control.
  • Your current automation relies on fixed flowcharts that fail when records are incomplete, clients respond partially, or work must wait across multiple systems.
  • You want model calls that can move toward private hosting and smaller models without redesigning the workflow engine.