1 00:00:00,761 --> 00:00:03,401 Hello, I'm Karthic, CTO at Perpetual. 2 00:00:03,681 --> 00:00:10,701 In this presentation, I'll talk about building workflows that are both reliable and intelligent for finance, wealth and insurance. 3 00:00:10,941 --> 00:00:18,281 In regulated domains, workflows must be reliable enough to meet regulatory standards while still being intelligent enough to handle the real world. 4 00:00:18,401 --> 00:00:23,721 In fields such as finance and wealth management, this presents both a technical and a compliance issue. 5 00:00:23,981 --> 00:00:26,241 Reliability and intelligence are not opposites. 6 00:00:26,241 --> 00:00:29,101 It is up to you where you draw the boundary between them. 7 00:00:29,281 --> 00:00:32,441 And that is the point of this talk, deciding where that boundary lies. 8 00:00:32,861 --> 00:00:37,741 For us, it's not merely a thought experiment, but a question that surfaces serious consequences. 9 00:00:38,201 --> 00:00:40,221 Perpetual is a wealth management roll-up. 10 00:00:40,361 --> 00:00:43,201 We acquire running traditional wealth management firms. 11 00:00:43,701 --> 00:00:47,481 And we maximize their throughput and revenue with the help of AI. 12 00:00:48,301 --> 00:00:50,481 Wealth management is a regulated industry. 13 00:00:50,661 --> 00:00:53,521 And that is where we build with AI in production. 14 00:00:53,521 --> 00:00:57,321 Everything in this talk is grounded in building for those constraints. 15 00:00:58,461 --> 00:01:01,721 Workflows in finance and wealth management come with four facts, 16 00:01:01,901 --> 00:01:05,001 and each of them pushes the architecture in the same direction. 17 00:01:05,341 --> 00:01:09,581 First, they're complex and they need runtime strategies as the process unfolds. 18 00:01:09,941 --> 00:01:12,041 You can't script every path in advance. 19 00:01:12,561 --> 00:01:18,781 Second, the data is sensitive, so you need a path to move these workflows entirely onto privately hosted models. 20 00:01:18,781 --> 00:01:23,821 Third, in a highly regulated industry, a policy violation isn't a bug report. 21 00:01:24,001 --> 00:01:25,501 It has severe consequences. 22 00:01:25,961 --> 00:01:30,181 So, fourth, execution has to be highly reliable and deterministic. 23 00:01:30,341 --> 00:01:34,321 Inject intelligence only where runtime judgment is genuinely required 24 00:01:34,321 --> 00:01:37,461 and minimize reliance on LLMs everywhere else. 25 00:01:39,361 --> 00:01:42,061 AI is spreading through regulated industries, 26 00:01:42,261 --> 00:01:44,561 but there's still no consensus on an architecture 27 00:01:44,561 --> 00:01:48,521 that gives you high reliability with the least use of LLM intelligence. 28 00:01:48,781 --> 00:01:53,641 Meanwhile, most of what developers hear from the big labs is how to burn more tokens. 29 00:01:53,921 --> 00:01:57,461 But in a regulated industry, token maxing is the wrong goal. 30 00:01:57,641 --> 00:01:59,061 These tasks are close to main. 31 00:01:59,361 --> 00:02:02,981 You can lay them out in advance and build them into a massive state machine. 32 00:02:03,321 --> 00:02:06,681 And that machine needs only a sprinkle of intelligence to navigate it. 33 00:02:06,981 --> 00:02:09,101 Most steps just follow the next edge. 34 00:02:09,501 --> 00:02:12,561 Only at a few judgment points does the model pick a branch. 35 00:02:13,121 --> 00:02:17,501 So the intelligent way to build is to use intelligence minimally and on purpose. 36 00:02:20,241 --> 00:02:23,541 Let's make this concrete with a real wealth management workflow. 37 00:02:24,181 --> 00:02:27,941 A wealth advisor turns to their assistant and says one sentence. 38 00:02:28,281 --> 00:02:33,201 Fill out the IRS W-9 for Mr. Harper to a person that's a simple instruction. 39 00:02:33,781 --> 00:02:37,301 Watch what it actually triggers when a human being executes it. 40 00:02:37,601 --> 00:02:40,941 The task has complexities that go beyond what meets the eye. 41 00:02:41,261 --> 00:02:46,021 First, the assistant finds form W-9 from the firm's records or from the IRS website. 42 00:02:46,021 --> 00:02:49,701 Then they look for Mr Harper in the system of record and find 43 00:02:49,701 --> 00:02:53,341 three Harper households. So they call the advisor. Which Harper 44 00:02:53,341 --> 00:02:56,861 do you mean? Only once that ambiguity is resolved do they move 45 00:02:56,861 --> 00:03:00,901 on to the form. A lookup, an ambiguity and a human in the loop 46 00:03:00,901 --> 00:03:03,001 all in the first minute of a one line task. 47 00:03:09,121 --> 00:03:13,801 Filling a form sounds like data entry. It isn't. Every form is a 48 00:03:13,801 --> 00:03:21,281 small program. A W9 first decides who the taxpayer is, the person or the owner if it's a disregarded 49 00:03:21,281 --> 00:03:27,781 entity. A foreign person stops right here because they need a W8 instead. Then how they're taxed? 50 00:03:28,241 --> 00:03:35,541 Exactly one box on line 3A and an LLC adds its letter. That decides which name goes on line 1 51 00:03:35,541 --> 00:03:43,341 and which on line 2. Then whose number? An SSN or an EIN. Never both. Taken from a verified document. 52 00:03:43,341 --> 00:03:48,841 never from memory. Exemption codes apply only to some entities and never to an S-corporation. 53 00:03:49,361 --> 00:03:54,581 And finally, who signs under penalties of perjury? Each answer changes the next question 54 00:03:54,581 --> 00:04:00,941 and the people involved only ever see three kinds of gates, a question, a document to collect, or 55 00:04:00,941 --> 00:04:10,541 a signature. Now they start filling the form and realize the system of record is missing three of 56 00:04:10,541 --> 00:04:16,181 the values it needs. They check the documents on file and the system of record again still 57 00:04:16,181 --> 00:04:20,981 missing so they start an email thread with the client asking for exactly those three 58 00:04:20,981 --> 00:04:25,361 values. The task now depends on someone outside the firm on their timeline. 59 00:04:30,941 --> 00:04:35,201 now the workflow waits. The request is with the client and all three values are still 60 00:04:35,201 --> 00:04:40,101 missing. If the client doesn't reply within two days, the assistant follows up and brings 61 00:04:40,101 --> 00:04:44,841 it to the advisor at the daily stand up. If the client replies with only two of the three values, 62 00:04:44,961 --> 00:04:49,821 they follow up again for the missing one. When all three arrive, they complete the form and write 63 00:04:49,821 --> 00:04:54,521 the new information back to the system of record, including the trail of how they obtained those 64 00:04:54,521 --> 00:04:59,621 values. In this industry, how you know a value matters as much as the value itself. 65 00:05:08,101 --> 00:05:10,761 Once the form is filled, it goes out for e-signature. 66 00:05:11,541 --> 00:05:15,381 If the client doesn't sign within a day or two, the assistant nudges them again. 67 00:05:16,181 --> 00:05:21,141 Once it's signed, they file it for the client, send the acknowledgement and update the advisor 68 00:05:21,141 --> 00:05:22,581 at the next day's stand-up. 69 00:05:22,801 --> 00:05:27,901 One sentence from the advisor became a process that runs for days, crosses several systems, 70 00:05:27,901 --> 00:05:31,421 weights on two people and loops back on itself more than once. 71 00:05:45,421 --> 00:05:49,661 Look at what's inside that one sentence. Ambiguity to resolve 72 00:05:51,221 --> 00:05:57,481 data together. Approvals, weights measured in days, follow-up loops and an 73 00:05:57,481 --> 00:05:58,641 evidence trail at the end. 74 00:06:04,641 --> 00:06:09,521 Workflow automation tools exist, but they rarely handle this real-world complexity and 75 00:06:09,521 --> 00:06:09,941 ambiguity. 76 00:06:10,321 --> 00:06:15,201 A fixed flowchart breaks at the first witch harbour and the lack of a native system of 77 00:06:15,201 --> 00:06:20,501 record with an integration and an agent that improvises its way through is exactly what 78 00:06:20,501 --> 00:06:22,201 a compliance team can't sign off on. 79 00:06:22,361 --> 00:06:25,341 One breaks on ambiguity, the other can't be signed off. 80 00:06:25,341 --> 00:06:31,401 neither handles what's inside that sentence. So what do we do about it? The fashionable framing 81 00:06:31,401 --> 00:06:38,221 is more agency and better models. Given agent the whole task and wait for the next model to 82 00:06:38,221 --> 00:06:44,581 make it reliable. The reflexive answer, especially inside regulated firms, is the opposite. Less 83 00:06:44,581 --> 00:06:51,201 agency, more rules, lock it down and hard code every path. Both are the wrong axis. 84 00:06:51,201 --> 00:06:57,021 In regulated verticals, reliability comes from where you place the boundary between the two, 85 00:06:57,461 --> 00:07:02,261 between the parts that must be deterministic and the parts that genuinely need judgment. 86 00:07:02,641 --> 00:07:03,361 Here's the principle. 87 00:07:04,301 --> 00:07:07,941 Determinism governs control flow, evidence, and side effects. 88 00:07:08,201 --> 00:07:13,421 What runs next, what we know, and how we know it, and anything that changes the world, 89 00:07:13,581 --> 00:07:17,121 like a record written, an email sent, or a form filed. 90 00:07:22,761 --> 00:07:50,261 Intelligence provides interpretation and judgment. Reading an ambiguous instruction, deciding which Harper, understanding a client's reply. The model answers a typed question and the deterministic system decides what to do with the answer and the interplay is recursive. A judgment can open a new smaller plan that's compiled, gated and deterministic again. You'll see that in a moment with follow ups. The boundary between the two is the design. 91 00:07:50,261 --> 00:07:52,761 Here's how that works in the workflow engine we've built. 92 00:08:01,701 --> 00:08:06,021 The request goes to a planner. That is a model call and its 93 00:08:06,021 --> 00:08:10,701 only job is translation. It emits a typed semantic graph of 94 00:08:10,701 --> 00:08:14,181 intents, entities, facts, artifacts, events, and 95 00:08:14,181 --> 00:08:17,721 constraints, each grounded in the words of the request with any 96 00:08:17,721 --> 00:08:21,801 any ambiguities it found. Every node points back to the words 97 00:08:21,801 --> 00:08:25,501 it came from. That graph goes to a compiler that never calls a 98 00:08:25,501 --> 00:08:28,921 model. In 11 deterministic stages, it turns the graph into 99 00:08:28,921 --> 00:08:33,041 an execution tag or it fails closed. There are only three 100 00:08:33,041 --> 00:08:36,301 verdicts. Valid needs clarification or refused. 101 00:08:39,501 --> 00:08:43,861 Which Harper comes back as a question, not a guess. The plan is 102 00:08:43,861 --> 00:08:46,741 is compiled, immutable and pinned by its hash. 103 00:08:56,981 --> 00:09:01,561 The model translates. The plan is compiled. The compiler checks 104 00:09:01,561 --> 00:09:05,221 the plan against the grammar. The vocabulary of what the firm 105 00:09:05,221 --> 00:09:09,821 is allowed to do. 42 typed action contracts across forms, 106 00:09:09,821 --> 00:09:13,841 documents, email, e-signature, calendar and human steps. 107 00:09:14,501 --> 00:09:19,241 Each action the planner asks for has to match a contract. Here 108 00:09:19,241 --> 00:09:23,041 form fill finds its forms contract and locks in. If the 109 00:09:23,041 --> 00:09:26,641 planner asks for an action that isn't in the grammar or wires 110 00:09:26,641 --> 00:09:28,781 one up wrong, the compiler refuses. 111 00:09:29,401 --> 00:09:33,441 There are about 40 ways to fail closed from a cycle in the graph 112 00:09:33,441 --> 00:09:34,901 to an unsupported action 113 00:09:35,141 --> 00:09:57,001 And the same contract digest is pinned in the engine, the backend and the UI. So all three agree on what an action means. That is what a closed domain buys you. The model can only ask for things the firm already knows how to do safely. Here is a compiled plan for a W9 chase. Send the client the form, wait for it to come back, check it and file it. 114 00:09:57,001 --> 00:10:00,861 Before anything runs, a person approves the plan itself. 115 00:10:01,101 --> 00:10:06,301 Find the form, prepare it for the client to fill, render the PDF and draft the email. 116 00:10:06,941 --> 00:10:09,241 Drafting is a model call, so it is violet. 117 00:10:13,661 --> 00:10:19,281 Sending is an action gate. Nothing leaves the firm until a person approves that exact email. 118 00:10:19,641 --> 00:10:23,121 Then the plan waits for a reply and a judge steps in to read it. 119 00:10:24,681 --> 00:10:31,701 The returned form is verified. If it matches, it's archived behind another gate. If it doesn't, 120 00:10:31,901 --> 00:10:40,141 the plan follows up. Now look at the dashed steps. Rendering the PDF, attaching it, judging the reply. 121 00:10:40,441 --> 00:10:46,421 The planner never asked for them. The grammar added them because the contracts require them. 122 00:10:46,881 --> 00:10:50,961 In the story, waiting, approvals, and follow-ups were most of the work. 123 00:10:50,961 --> 00:10:58,681 Here there are nodes in the plan, not prompts. A reply wait defaults to 72 hours and an explicit 124 00:10:58,681 --> 00:11:04,881 deadline can be anything from a minute to 30 days. When it passes, a rule decides what happens, 125 00:11:05,161 --> 00:11:11,661 not a model. One approval covers the plan and every external action like sending an email or 126 00:11:11,661 --> 00:11:17,941 a signature request has its own gate. The approval is tied to a hash of the exact payload and it's 127 00:11:17,941 --> 00:11:22,881 checked again right before sending. When a client replies with only part of what we asked, 128 00:11:23,161 --> 00:11:29,801 the reply judge marks it partial. That judgment is provisional. The step stays blocked until a 129 00:11:29,801 --> 00:11:35,601 person confirms it. A person confirms it and the step unblocks. Then the follow-up runs as a new 130 00:11:35,601 --> 00:11:40,901 child plan. The approved parent is never changed and there are at most five rounds. 131 00:11:46,641 --> 00:11:53,141 Once a plan is approved, a state machine runs it. It picks every step and records every result. 132 00:11:53,361 --> 00:11:59,601 When a step needs judgment, it makes one fixed schema model call. Compose an email, judge or 133 00:11:59,601 --> 00:12:06,081 reply, spot missing values, extract fields, verify a document. The answer comes back typed, 134 00:12:06,081 --> 00:12:11,201 and the state machine decides what to do with it. Side effects get a life cycle of their own. 135 00:12:11,781 --> 00:12:17,541 Prepare, dispatching, succeeded, or when the world doesn't answer cleanly, outcome unknown, 136 00:12:17,781 --> 00:12:23,401 and manual review. There are no automatic retries. A chat message can approve a step 137 00:12:23,401 --> 00:12:29,901 or answer a question. It can't change an already approved plan. The model steers. It never drives. 138 00:12:36,081 --> 00:12:40,321 Now look at what is left for the model to do. Outside the planner, the judgments are 139 00:12:40,321 --> 00:12:46,801 small and typed. Is this reply complete? Does this document match? Which field is which? 140 00:12:47,001 --> 00:12:52,221 They don't need frontier intelligence. A fast small model, the flash tier, can handle 141 00:12:52,221 --> 00:12:58,881 them. In this illustrative run of 20 decisions, 19 go to the flash tier and one goes to the 142 00:12:58,881 --> 00:13:05,141 frontier. That is why we designed for about 95% of model decisions in a workflow never 143 00:13:05,141 --> 00:13:08,321 never needing a frontier model and the flash tier can run on 144 00:13:08,321 --> 00:13:12,661 hardware you control. Only the genuinely hard cases go to a 145 00:13:12,661 --> 00:13:13,281 frontier model. 146 00:13:15,421 --> 00:13:20,041 the 5% is not a permanent tax because every judgment is 147 00:13:20,041 --> 00:13:23,281 typed and logged and the provisional ones are confirmed 148 00:13:23,281 --> 00:13:26,961 or rejected by a person. Each frontier decision produces 149 00:13:26,961 --> 00:13:30,981 eval grade distribution real training data by construction. 150 00:13:30,981 --> 00:13:36,521 That data can post train the flash tier. Logging and evals 151 00:13:36,521 --> 00:13:40,281 exist today. The training loop is next. So the frontier share 152 00:13:40,281 --> 00:13:44,461 shrinks release over release until even the hard cases run 153 00:13:44,461 --> 00:13:44,881 locally. 154 00:13:51,841 --> 00:13:55,641 Remember the requirement we started with because data is 155 00:13:55,641 --> 00:13:59,361 sensitive. You need a path to move workflows entirely to 156 00:13:59,361 --> 00:14:04,001 privately hosted LLMs. So can this design really get there? Take 157 00:14:04,001 --> 00:14:08,081 one type call from the Harper Burbinine plan. Every model call 158 00:14:08,081 --> 00:14:12,461 is small, typed and checked against a schema. So pointing it 159 00:14:12,461 --> 00:14:16,061 at a privately hosted model is a configuration change, not a 160 00:14:16,061 --> 00:14:20,621 rewrite. One line changes and the same answer comes back. Every 161 00:14:20,621 --> 00:14:24,641 other call has that same small type shape. So the same change 162 00:14:24,641 --> 00:14:26,181 applies to almost all of them. 163 00:14:26,181 --> 00:14:34,221 That is how about 95% local inference becomes viable as an architectural consequence, not a policy promise. 164 00:14:34,601 --> 00:14:42,481 That is the path that sensitive data demands and compliance risk goes down as an engineering roadmap, release by release. 165 00:14:42,941 --> 00:14:48,861 So if you remember one thing, don't ask whether your system should have more agency or less. 166 00:14:49,201 --> 00:14:50,521 Ask where the boundary is. 167 00:14:50,881 --> 00:14:55,541 Let determinism govern control flow, evidence and side effects. 168 00:14:55,541 --> 00:14:58,841 Let intelligence handle interpretation and judgment. 169 00:14:59,281 --> 00:15:05,621 At typed points, you can inspect and use every judgment you log to need less of it next time. 170 00:15:05,881 --> 00:15:07,801 Ask where the boundary is. Thank you. 171 00:15:08,001 --> 00:15:10,521 I'm Karthic, CTO at Perpetual.