The audit trail is the product, not the friction.
Finance does not buy answers from AI. It buys answers it can defend in front of a model-risk committee. Acuity Knowledge Partners has spent twenty years reading 10-Ks, building covenant trackers, and writing sector notes for sell-side desks and buy-side PMs in Bangalore, Gurgaon, Colombo, and Beijing. The interesting thing about plugging language models into that workflow is not the model. It is what survives once the regulator, the compliance lead, and the bank's annual vendor audit have all had a look.
In sixty seconds.
Equity research, credit underwriting, and PE diligence are made of the same atomic unit: an analyst reading something long, writing something short, and citing what they read. Acuity has spent two decades industrialising that loop for sell-side desks, asset managers, and lenders who could not staff it internally at the price they wanted to pay.
AI looks, at first glance, like the moment the offshore arbitrage gets eaten. It is the opposite. The model is the cheapest component in the stack. The expensive thing is the citation back to a paragraph in a 10-K, the redaction policy that keeps a draft research note out of training data, and the SR 11-7 review that proves the model behaves the same way next quarter.
Twenty years of labelled draft notes, tagged transcripts, and compliance-approved templates are the moat. The model is rented. The audit trail is owned. That is the bet Acuity is running, and it is the bet the analyst-replacement startup pitch consistently misses.
Where to land in the conversation.
Each chapter opens the YouTube video at that timestamp in a new tab.
Five ideas to carry into your own work.
Mental models lifted from the conversation that travel beyond financial research. Each one is the kind of thing you can quote in a Tuesday strategy review without sounding like you read it on a slide.
AI as reading-list compression
An equity analyst spends maybe 70% of the week reading and 30% forming a view. A junior analyst spends 90/10. The pitch isn't "replace the analyst"; it is "halve the reading week so the judgement week doubles." Same headcount, twice the coverage, same liability surface.
The audit trail as a product surface
In finance, an answer with no citation is unusable. The most expensive piece of engineering in the stack is not the LLM — it is the chain from a sentence in a draft note back to a paragraph in a filing, a line in a transcript, or a row in the covenants table. Treat citation as a UI, not a footnote.
Offshore research as the model factory
Two decades of analyst-graded drafts, redlines, sector notes, covenants extractions, and earnings-call summaries — every one of them paired with the senior's correction — is a proprietary supervised dataset that no open-web crawl can buy. The KPO didn't realise it was building a labelling factory until the labels became the asset.
Compliance-as-moat
Passing a J.P. Morgan vendor audit, then a Barclays one, then a State Street one is not friction — it is the moat. The fortieth review is cheaper than the first. A startup with a better RAG pipeline and no audit history cannot bid for the same desk's research budget. The Indian KPOs spent fifteen years building that posture before anyone called it AI.
The model is the swappable dependency
Acuity's stack is intentionally model-agnostic: Claude this quarter, an open weights model next, Gemini for transcripts, an in-house extractor for tables. The LLM is the cheapest tier. Tearing it out and putting in another one is a fortnight of regression tests, not a rebuild. The deep work sits in the retrieval index, the prompts, the redaction layer, and the eval set.
Fifteen things to walk away with.
Each one carries the timestamps where the moment lives, and a transferable note for work that isn't financial research.
Where the analyst's week actually goes.
A bulge-bracket sell-side analyst is paid for the call — the upgrade, the downgrade, the price target. Anyone honest about the week will tell you that the call takes a Tuesday afternoon. The rest is reading: 10-Ks, 10-Qs, conference-call transcripts, broker notes, sell-side competition, channel checks, sector studies. The Acuity argument is that AI is at its best precisely where the analyst is most bored — turning a 240-page annual report into the eight numbers and three risks the model needs to defend the call.
The framing matters because it predicts who keeps the seat. If the model only compresses inputs, the analyst's seat survives — and probably broadens, because the same person can now cover five names instead of three. If the model produces the call, that seat is gone. Acuity is selling the first story; they are explicit it isn't the second.
Sell-side and buy-side want different things from the same paragraph.
A sell-side analyst at a broker is writing the note that will be distributed to institutional clients. Citation density, sector benchmarks, and a defensible target price matter most. A buy-side PM at an asset manager is writing nothing — they are reading the sell-side and adding their own view in a private memo. The same earnings transcript pulled from the same filing gets summarised differently for each. The model has to know who is asking.
This sounds like a small product decision and it isn't. The retrieval index, the redaction policy, the citation format, the entitlements model — all of them split on this fork. A startup that pretends one product serves both audiences usually means they have not yet sold to either.
The 10-K is harder than it looks.
SEC 10-Ks are nominally structured: Item 1 Business, Item 1A Risk Factors, Item 7 MD&A, Item 8 Financials, footnotes. In practice every issuer formats them differently, footnotes reference items five sections away, and the interesting language hides in the boilerplate. Acuity's parsing pipeline has been tuned issuer by issuer for years — the tagged corpus is, in effect, a private benchmark that an open-source RAG over EDGAR cannot match in two months.
Earnings-call transcripts are worse. Speaker attribution drifts. Analysts get cut off mid-question. CFOs hedge in clauses three deep. The labelled correction set built across thousands of calls is the difference between a model that pulls the right number and a model that pulls the prior-year comparable and gets the call wrong.
RAG is the architecture. Fine-tuning is the temptation.
For a research workflow, retrieval-augmented generation is structurally the right call. The corpus changes weekly — new filings, fresh transcripts, broker updates. Fine-tuning a model on Monday is stale by Friday. Retrieval over a maintained, versioned, citable index lets the model stay small and the index do the work. Acuity's posture is unambiguous on this; the engineering pours into the index, the chunking strategy, the rerankers, and the prompt caching that keeps repeated context cheap.
The deeper reason is auditability. A fine-tuned model is a black box to a compliance lead. A retrieval pipeline can show which paragraph it pulled and why it ranked it first. The first thing a model-risk reviewer asks is "show me what it read." Fine-tuning cannot answer that question. Retrieval can.
SR 11-7 is the document the AI vendor never quotes.
The Federal Reserve's SR 11-7 — supervisory guidance on model risk management — is the spine of every American bank's internal MRM function. It defines what a "model" is (broad), demands independent validation, ongoing monitoring, conceptual soundness checks, and outcome analysis. Any LLM that informs a credit decision, a trading signal, or a research conclusion now lives under that framework. Basel's equivalent at international banks adds further capital-charge implications.
The vendor pitch deck that doesn't mention SR 11-7 is talking to the wrong person. The buyer at a Tier 1 bank is not the head of equity research — it is the model-risk committee, which has the right to kill the pilot after the procurement team has already signed. Acuity learned to write the validation memo before the bank's MRM team had to ask for it.
The redaction problem is bigger than the model problem.
A draft equity-research note contains the analyst's call before it goes out to clients — material non-public to anyone outside the firm, market-moving if it leaks. A buy-side memo contains the fund's positioning, the kind of thing the SEC would treat as confidential supervisory information. Pushing either through a third-party model API without a redaction layer is a career-ending compliance failure waiting for a quiet Wednesday.
Acuity's redaction posture is concrete: tickers and company names get tokenised at the API boundary, the model sees pseudonyms, the citations re-hydrate on the way back. Self-hosted inference for sensitive workflows. Prompt logs retained, customer-segregated, never used to train the vendor's next checkpoint. The contractual paragraph protecting that last point is more important than the model card.
Every bank runs its own audit. The ISO certification is the ticket, not the seat.
ISO 27001, SOC 2 Type II, sometimes ISO 42001 for AI specifically — Acuity carries all of them. Every Tier 1 bank still sends its own questionnaire, then its own on-site review, then a quarterly attestation, then an annual penetration test. The certifications get the conversation started; the bank's MRM team and infosec team decide whether it ends with a contract.
The structural lesson for anyone selling AI into financial services in 2026: the audit isn't a phase, it is a permanent surface area of the company. Build the engineering muscle that makes the fortieth audit cheaper than the first, the same way Acuity built it across fifteen years of vendor reviews for HSBC, Goldman, Citi, and the rest.
The cohort of sell-side jobs that gets eaten — and the one that doesn't.
Honest about the human cost: the junior associate whose week is spreadsheets, comparable-company tables, and pulling numbers from filings has a smaller seat in 2030. The senior analyst whose week is client calls, sector thesis, and the upgrade-downgrade has a larger one. The trade isn't headcount-flat; it is headcount-shaped-differently, and the firms that pretend otherwise are not levelling with their juniors.
Acuity is, paradoxically, hiring more — but the hire profile has shifted. They want analysts who can write a prompt, design an eval, and red-team a model's output as fluently as they once tagged a covenant. The CFA is still useful. The ability to spot a hallucinated EBITDA reconciliation is newly compensable.
Twenty years of labelled draft notes is proprietary supervised data.
Every research note Acuity has ever drafted exists in two forms: the analyst's first pass and the senior's redline. Every covenant extraction has a quality-assurance review. Every ESG materiality flag was reviewed by a sector lead. That is, in practice, a labelled supervised dataset across millions of examples — graded by professionals in the exact judgement task the model is being asked to learn.
An open-web crawl cannot buy this. A new entrant cannot synthesise it in eighteen months of paid labelling. The dataset isn't published, isn't licensed, and the contracts under which it was generated almost certainly prohibit selling it. The asymmetry is durable in a way the model itself is not.
Asset managers don't buy insight. They buy analyst-hours.
The implicit pricing model at most buy-side firms is time, not idea. A PM running a long-short book wants more names covered, faster turnaround, deeper comparables, fresher channel work. The intellectual centre of the fund is the PM's view; the marginal analyst-hour is what feeds that view. Acuity sells the analyst-hour, packaged as a research note or a screen or a covenant tracker, at a unit cost the in-house function can't match.
AI does not change the demand side. It changes the supply side: the same analyst can cover more names because the reading week shrinks. The buy-side spend per name goes down; the number of names covered per fund goes up; the absolute spend stays roughly flat or grows. That is the bull case for the offshore research firm in an AI world.
Prompt caching is the line-item that matters at scale.
A research workflow re-uses the same instructions, the same disclaimers, the same chain-of-thought scaffolding for every sector note. With prompt caching across the inference layer, the static prefix becomes ~free after the first call; only the per-document delta pays full token rate. At Acuity volumes — thousands of notes a day across hundreds of clients — that is the difference between a per-note unit cost that makes sense and one that does not.
The corollary is architectural: structure the prompts so the static portion is large and the per-call delta is small. Front-load the persona, the policy, the citation rules, the redaction prompt. Vary only the document at the end. This is the kind of detail nobody writes about in a model launch but that decides whether the line of business is gross-margin positive.
Hyperscalers do not eat this directly.
The reflexive worry: Microsoft, Google, AWS each have a "research copilot" SKU. Why does Acuity exist between the bank and Azure OpenAI Service? Because the bank cannot ship its draft research notes to a hyperscaler API without a redaction layer, a per-tenant index, a model-risk pack, an entitlements model, a vendor MSA the procurement team has already signed, and a human who answers when something goes wrong at 11pm. Acuity is the integrator that ships all six.
The hyperscalers compete to be the substrate; Acuity competes to be the customer-facing wrapper that includes humans in the loop on the days the model misfires. That is a stable equilibrium for at least the medium term — long enough to matter for the next financing round.
The KYC-for-AI-vendors framework is real, and it is new.
Banks have spent two decades building KYC processes for their customers. The mirror process for AI vendors is being built right now: which models the vendor uses, where they are hosted, who has access to logs, whether prompts leave the jurisdiction, whether outputs are reused in training, how often the model is changed under the customer, how the customer is notified. The questionnaires run twenty pages. Few startups have answers to all of them.
Acuity is at the front of this because they were already answering most of the questions for their non-AI services. The pivot was incremental: add the model-card disclosures, the eval-set evidence, the change-log of which checkpoint was in production on which day. The startup with no compliance history has to build the whole posture before the first contract.
ISO 20022 is the finance world's quiet FHIR moment.
Outside cross-border payments, ISO 20022 doesn't get much oxygen. It should. As the global messaging standard for financial transactions and reporting, it is becoming the structured backbone the unstructured research layer increasingly leans on. When a covenant tracker can pull both a free-text disclosure and a structured ISO 20022 payment message about the same counterparty, the model has stereo data — and far less opportunity to hallucinate.
The parallel to healthcare's FHIR is exact. FHIR didn't replace the clinical note; it gave it a structured neighbour. ISO 20022 plays the same role for unstructured finance documents. The firms that ingest both surfaces and reconcile them will pull ahead of the firms that ingest only the free text.
Acuity's twenty-year arc is the answer to the "is this defensible" question.
Founded in the early 2000s as a captive offshore analytics arm, spun out, rebranded, acquired, and acquired again — Acuity has lived through the original KPO boom, the global financial crisis that hollowed out sell-side desks, the post-2010 squeeze on research budgets under MiFID II, and now the AI re-platforming. Each cycle left the firm structurally stronger because each cycle pushed more analyst work offshore, not less. AI continues that arc; it does not break it.
The strategic claim is not that Acuity has a better model. It is that the firm has survived four cycles of "this is finally the end of the offshore research arbitrage" and each time the arbitrage got broader, not narrower. The fifth cycle has the same shape. The asset is the durability, not the product.
Lines worth keeping near your desk.
The jargon, unpacked.
Some of these will be obvious; some won't. Skim, mark the unfamiliar, come back later.
Check what you actually retained.
Try to answer before you click. The point is to notice where the conversation is fuzzy in your memory, then return to the transcript.
Five questions worth sitting with.
No correct answers. Type into the boxes — your responses are saved locally to this browser.
Acuity argues the audit trail is the product. In your domain, what is the analogue — the artefact the buyer cannot use without, even though no one prices it?
The reading week shrinks; the judgement week stays. Where does that trade play out in your own role, and which slice of your week is the model coming for?
The KPO turned out to have been building a labelled dataset for twenty years without calling it that. What has your organisation accidentally accumulated that becomes an asset only in retrospect?
"The model is the cheapest component." If you accept that, where should engineering investment shift — and what currently-prestigious work loses status?
Compliance-as-moat depends on the regulator staying meaningfully strict. What is your bet on whether financial regulators in 2030 are tighter or looser than today, and how does the moat re-price?
Where to push back.
The strongest version of each disagreement, written to be persuasive — not to win.
"This commoditises the analyst."
The counter: at the firm-wide level, the math doesn't necessarily hold. If every sell-side desk can cover five names where they previously covered three, the marginal call gets worse — the senior is spread thinner, the time per call drops, the differentiation between desks collapses. The endgame isn't a richer judgement week; it is a flat-rate research utility where calls are commoditised even though the analyst is theoretically more productive. Acuity's pitch is true for individual firms and false at aggregate equilibrium.
"Hyperscalers eat this directly."
The steelman for the hyperscaler: Microsoft has been writing the model-risk pack for two years inside their financial-services vertical. Azure OpenAI Service ships with logging, regional residency, content filtering, and a procurement contract that the bank's vendor team already accepts. The 11pm human is a managed-services line item, not a defensible product. Once the hyperscaler ships the integration partner programme, the Acuity wrapper becomes a thin margin sitting on someone else's substrate. Treat the equilibrium as medium-term, not permanent.
"Compliance is friction, not moat."
The counter: regulators iterate. SR 11-7 dates from 2011 — fourteen years old — and the Fed has been signalling an updated framework specific to generative AI. When the framework changes, the old compliance posture is partly invalidated and the entire industry resets to the new baseline. A startup that built to the new framework natively can sometimes leapfrog the incumbent that built to the old one. Friction is moat only between resets; the moat depreciates faster than people quoting the line acknowledge.
"Twenty years of labelled data is the moat."
The push: a lot of that data is contractually owned by the client banks, not by Acuity. The labelled corpus may be a moat in spirit and not in title — if a Goldman or HSBC asserts ownership over draft notes produced under an MSA, the asset Acuity carries on its balance sheet shrinks. Worse, the next generation of foundation models trained on EDGAR plus public broker reports plus synthetic data may close most of the quality gap without ever needing access to the private corpus. The moat is real, but the durability is partly legal-fiction.
Three angles on Monday morning.
If you don't work in financial research, here's what to take.
If you're a finance leader
- Map your analyst week into reading-hours and judgement-hours. The first number is the budget AI is coming for.
- Ask your AI vendor for the SR 11-7 validation memo before the demo, not after. If they can't write one, the procurement cycle will end the same way regardless of how good the demo was.
- Write the redaction policy as a one-pager. Tickers, fund positions, draft calls — what leaves the perimeter and in what form. The policy is more durable than the model selection.
- Track research-units-per-name, not headcount-per-desk. AI grows the first; the second is the wrong metric to defend in front of the CFO.
- Treat the audit trail as a UI surface. Citations are read by clients, by compliance, and by the regulator. They are not a footnote.
If you're an operator
- If you sell into regulated industries, write the customer's vendor-onboarding questionnaire into your engineering roadmap before product roadmap. The questions you can't answer are tomorrow's lost deals.
- Build the "fortieth audit cheaper than the first" muscle: runbooks, evidence packs, attestation templates. The asset is the runbook, not the certificate.
- Structure prompts as static prefix plus variable suffix. The cache hit rate is a margin line, not an engineering curiosity.
- Identify the labelled data you've been generating for years without calling it that. Senior redlines, QA corrections, customer-support transcripts. The label is the asset.
- Position so the hyperscaler is your substrate, not your competitor. Sell what they need to push through to reach the buyer.
If you're an engineer
- RAG, not fine-tune, when the corpus changes weekly and the customer needs citations. Defensibility beats benchmark scores in regulated workflows.
- Build the redaction layer at the API boundary. Pseudonymise on the way out, re-hydrate citations on the way in. Treat third-party model APIs as untrusted by default.
- Front-load prompt prefixes so cache hits dominate. Re-architecting prompts for cache is the cheapest performance win you will get this quarter.
- Version every retrieval index. Compliance will ask which version was queried on a specific date. The answer "we don't know" is the wrong answer.
- Maintain the eval set as a first-class artefact. When the model changes, the eval set is what decides whether the new one ships. The eval set is the firm's IP; the model is rented.
Twenty-five years, briefly.
The arc behind Acuity's positioning, lined up against the broader financial-research outsourcing wave.
The whole conversation, segmented.
Click a timestamp to open YouTube at that moment. This transcript scaffold is empty in the current draft — see the note below.