Follow research from expert sources to AI-assisted decisions. Explore independent corroboration, vendor economics, permissions and the evidence behind an answer.
Financial AI makes evidence procurement part of production
The 8 October financial-services dossier expands this article from sourcing conversations to producing financial work. A research assistant retrieves and summarises; a workflow product also assembles the model, applies institutional conventions, generates editable materials and carries the evidence into review. This distinction matters because a fast answer is not necessarily an accepted deliverable. A credit memo with an untraceable cash-flow figure creates work for the reviewer rather than removing it.
The working architecture has five responsibilities. An entitlement layer establishes which internal and licensed sources the user may access. Retrieval preserves document, date, page and unit. Analysis separates reported facts from assumptions and calculations. Output production creates material that another professional can inspect and change. An accountable reviewer decides whether the evidence supports the conclusion. A model provider, data provider and workflow vendor can each own different parts of that chain; one interface does not remove the underlying contracts.
Current primary releases support this shift. Anthropic announced ten finance agent templates on 5 May 2026, including pitchbook, KYC and month-end workflows, and described connectors and specialist subagents as components. OpenAI’s 10 September 2026 financial-services release combines financial data with research, modeling and client-material workflows. These are vendor descriptions of offered functionality, not independent measurements of bank productivity. The procurement test remains an accepted output on the institution’s task set, with appropriate permissions and review.
Translate time saved into value that can be realised
The source dossier’s productivity case distinguishes assembly from judgment. Suppose a task takes eight active hours, 60% is assembly, automation saves 60% of assembly time, and extra checking costs 6% of the original task. The net saving is 0.60 × 0.60 − 0.06 = 30%, leaving 5.6 hours. Retained judgment, corrections and approvals explain why a demonstration that writes a first draft in seconds does not eliminate an eight-hour work process.
The following is the dossier’s hypothetical 1,000-user case, recalculated for this edition. It assumes 46 workweeks and a loaded hourly value of USD 100. Annual license cost is USD 3 million and first-year integration, security and training add USD 1 million. These are planning inputs, not any institution’s spending or a vendor quote.
| Scenario | Active users | Hours saved per active user/week | Gross capacity value | Realisation factor | Realised value | Value less USD 4m first-year cost |
|---|---|---|---|---|---|---|
| Conservative | 400 | 1.5 | USD 2.760m | 25% | USD 0.690m | −USD 3.310m |
| Base | 600 | 3.0 | USD 8.280m | 40% | USD 3.312m | −USD 0.688m |
| Aggressive | 800 | 5.0 | USD 18.400m | 60% | USD 11.040m | USD 7.040m |
At the base adoption and realisation assumptions, break-even requires 4,000,000 ÷ (1,000 × 0.60 × 46 × 100 × 0.40), or about 3.62 hours saved per active user per week. Capacity value is not automatically cash. The business must show avoided overtime, hiring or outsourcing, or separately substantiated value from additional useful output. Counting both saved hours and avoided headcount for the same work would double count the benefit. Ongoing inference, data licenses and support also belong in the cost model.
Scientific discovery is an evidence lesson, not a sales forecast
The September Buy Side Lens fluid-singularity note supplies a useful verification case. OpenAI’s 8 September 2026 announcement reports a Navier–Stokes result for a specifically forced system, with a written proof and Lean formalisation. That does not establish that every fluid problem is solved, nor that an unrelated financial agent is reliable. The mathematical question, assumptions and proof object must travel together. This article has not independently reconstructed the proof.
There is a material update after the original note: on 6 October OpenAI released additional mathematical results with revision and citation protocols and formalised proofs for many results, while describing consultation with an independent advisory group. The important business inference is that scalable discovery creates demand for stronger verification, provenance and revision management. It does not directly establish customer revenue or return on capital.
A research buyer should apply the same hierarchy to financial AI: an attractive answer, a traceable answer, a checked calculation, an approved decision and an economic outcome are different milestones. Test false corroboration when several outputs inherit the same source. Keep the original evidence and make corrections propagate to every report that relied on it. Human expert interviews remain valuable when they resolve a specific gap that licensed data and document reasoning cannot resolve; their role becomes more focused as preparation improves.
Figure 01 · Conceptual framework
From a research question to an accountable AI output
The research and evidence workflow
Expert research is most valuable when it tests a proposition that published information cannot resolve. A well-chosen interview can reveal how a buying decision works, why a workflow fails or which assumption in a market model is unrealistic. Repeated calls that restate the same view are expensive duplication, even when every conversation sounds informed.
This report develops the distinct expert-network research section of a historical dossier prepared on 4 September 2026. All network, client, employer and expert identities are removed. Market-size and price ranges remain reported historical estimates from that dossier. The analysis concerns research procurement and the working desk, not the disclosure of private engagement details.
Begin with the decision the research must change
A useful brief names the decision, the existing evidence and the unresolved question. “Understand the market” is too broad to determine whom to interview or when enough work has been done. “Test whether customers can replace the workflow within one budget cycle” is more specific and can be challenged with evidence.
The historical dossier identified several jobs for expert research: testing an investment thesis, accelerating learning, identifying overlooked questions, validating vendor claims, estimating poorly documented markets, understanding willingness to pay, following implementation and studying failures. Those jobs require different experts and different interview designs.
Match the kind of direct knowledge to the research question
| Aspect | Retained observation |
|---|---|
| Buyer | A buyer who has selected software can explain procurement constraints that a vendor salesperson may not see. |
| Implementer | An implementer can explain integration effort that a senior sponsor only observes at a high level. |
| Former operator | A former operator may offer comparative history while lacking current visibility. |
Expertise should be matched to the question rather than ranked solely by seniority.
Calls, surveys and transcripts are different products
Live interviews allow follow-up and clarification. Surveys create structured comparisons across respondents. Transcript libraries make previously collected conversations searchable. Custom research can combine interviews with desk analysis, while advisory engagements supply continued access to a particular specialist.
Choose the research product that matches the unresolved evidence gap
| Research product | Best use | Main limitation |
|---|---|---|
| Live expert interview | Resolve ambiguity and probe operating experience | Small sample and interviewer effects |
| Structured survey | Compare responses to consistent questions | Sampling and question-design bias |
| Transcript library | Learn the vocabulary and map recurring issues | Old context and limited ability to clarify |
| Custom recruitment | Find rare experience outside standard panels | Time, cost and uncertain conversion |
| Analyst synthesis | Integrate several forms of evidence | Quality depends on traceability and judgement |
| Ongoing advisory | Repeated work on a complex problem | Dependency and scope control |
A transcript can answer a background question cheaply but may not settle an issue whose context differs. A survey can count opinions but cannot make a biased respondent pool representative. The product should follow the evidence gap.
Pricing is a package, not an honorarium
An expert call, a searchable transcript library and a financial AI workspace buy different things. Compare the cost of a completed research decision, including internal preparation, licensed information, review and follow-up. Published or privately mentioned hourly fees cannot establish an equivalent price unless the entitlement, service scope and useful output are matched.
The earlier dossier’s unsourced market-size and commercial price bands have been removed from the current edition. They remain in the original source for audit, but they are not a sound October 2026 market benchmark. A buyer can make the decision with a transparent utilisation case and actual quotations instead.
Figure 02 · Qualitative comparison
Three research products, different evidence limits
| Approach or concept | What it contributes | What still matters |
|---|---|---|
| Expert call | Provides direct contextual practitioner knowledge | One participant's experience is not the whole market |
| Transcript or survey | Provides reusable or structured observations | Recruitment and question design shape the evidence |
| Financial AI workflow | Organises and synthesises information at scale | Reliability depends on permissions, grounding and review |
Subscription utilisation changes the effective price
Consider an explicitly illustrative $60,000 annual package intended to support 60 useful interviews. Full use would imply $1,000 per interview before internal research time. If only 40 are completed and useful, the effective cost rises to $1,500. If ten of those largely repeat information already known, the cost per distinct useful contribution rises further.
This is why procurement should measure completed useful work, not simply bookings or credits consumed. A subscription can create an incentive to schedule unnecessary calls near expiry. A pay-per-call model avoids that particular problem but may cost more for a genuinely high-volume team.
The commercial comparison should include cancellation terms, minimum durations, premium-expert surcharges, geographic differences, recording and transcript rights, and the ability to move credits across teams. The nominal hourly rate is only one part of the bill.
Recruit for direct knowledge and a clear boundary
Screening should establish the expert’s actual relationship to the issue: user, evaluator, budget holder, implementer, adviser or former employee. It should identify the period, geography and scope of that experience. A person can have deep knowledge of one product module without knowing enterprise-wide spending or every regional deployment.
The expert should not be prompted to adopt a more senior role or broader experience than they can support. The buyer’s evidence quality depends on accurate self-description. A confident answer about an unfamiliar area is less useful than a precise explanation of what the respondent knows and what they do not.
Client and third-party confidentiality remain separate from anonymisation. Removing a company name does not convert restricted information into publishable research. The interview brief should seek general processes, authorised experience and appropriately sourced market observations, while excluding live confidential transactions, protected contract terms and information the respondent cannot share.
Design interviews to falsify the thesis
A productive interview begins with the current proposition and deliberately seeks conditions under which it fails. If the thesis is that a vendor is hard to replace, ask about the last successful replacement, the module most easily removed and the circumstances that would justify migration. If the thesis is pricing power, ask which customers resisted increases and how the package changed.
Questions should separate observation from inference. “What did you directly see?” differs from “What do you think the market will do?” Both can be useful, but they should not enter a model with the same evidential weight.
The interviewer should ask for denominators and time periods whenever numbers are offered. Spending can refer to one team or the whole group; market share can refer to revenue, customers or installed assets; growth can be organic or acquisition-led. Many apparent disagreements disappear once these definitions are aligned.
Avoid false corroboration
Several experts can repeat the same public report or vendor presentation. That is not independent confirmation. Similarly, a group drawn from the same professional network can share the same assumptions and blind spots.
The research log should identify the basis of each claim: direct operating experience, public data, second-hand information, estimate or forecast. It should also record overlap between respondents. Independent corroboration requires different underlying evidence, not merely a larger number of voices.
Disagreement is often valuable. A specialist buyer and a price-sensitive customer may reach different conclusions because their requirements differ. The synthesis should explain those conditions rather than average their ratings into a misleading universal score.
Figure 03 · Conceptual framework
Has apparent corroboration come from independent evidence?
How the parts connect
- Repeated research claimPath AIndependent sources support it
- Repeated research claimPath BSeveral outputs repeat one origin
- Independent sources support itReviewDecision-ready conclusion
- Several outputs repeat one originReviewDecision-ready conclusion
AI should organise evidence without erasing context
AI is useful for transcript search, topic tagging, comparison, first-pass coding and identifying repeated claims. It can help a researcher locate all statements about switching, pricing or implementation across a large body of material. That can reduce duplication and improve the design of the next interview.
Its outputs need to remain linked to the original passage and context. A summary can easily omit a qualifier such as “in one region,” “three years ago” or “a rough estimate.” Removing that phrase can turn a bounded observation into a false enterprise-wide fact.
A controlled workflow distinguishes source statements from editorial conclusions. It preserves dates, speaker scope, uncertainty and disagreement. AI-generated synthesis should not manufacture a consensus or imply that a respondent endorsed the final investment argument.
The industry outlook in the original dossier
The historical source expected routine information retrieval and transcript summarisation to face greater automation, while senior judgement and custom recruitment in genuinely narrow domains remained more durable. That was its original forward view. It did not claim that all live research would disappear or that every provider would benefit equally.
The economic mechanism is that searchable existing material can answer more background questions before a call is booked. The remaining interviews can become more targeted and demanding. Providers that deliver rare, well-matched experience and a reliable process may retain value even if basic access becomes less differentiated.
This also changes the buyer’s workflow: read first, identify the unresolved question, then interview. The value of the call rises when the researcher arrives with a model to test rather than asking the expert to provide a generic introduction.
Know when to stop
A procurement review should consider failed matches as well as completed calls. Relevant measures include time to present qualified candidates, the proportion accepted, cancellation and rescheduling rates, the usefulness of completed interviews and the number of research questions actually resolved. A provider with a low quoted rate can be expensive if the internal team spends substantial time screening poor candidates.
The research record should also preserve the distinction between a prepared discussion and a completed interview. A questionnaire, briefing or scheduled call is evidence of work undertaken, but it is not proof that a conversation happened or that a respondent supplied the answers. That distinction matters when a large archive is merged: otherwise preparation text can be mistaken for independent corroboration.
A well-maintained evidence register makes this visible. It records the research product, completion status, date, scope and actual contribution. The resulting archive becomes easier to search and less likely to repeat the same claim under several apparently different reports.
Research should stop when additional interviews are unlikely to change the decision or when the remaining uncertainty cannot be resolved by this method. A question with a published statistical answer does not need an expensive call. A question requiring confidential transaction data should not be pursued through increasingly aggressive screening. A statistically representative estimate needs a suitable sample design, not a handful of conveniently available experts.
After a batch, the researcher should write an interim view: what is supported, what is contradicted, what remains unknown and what evidence would change the conclusion. The next batch should address those gaps. This prevents accumulation of conversations without accumulation of knowledge.
The central procurement principle is to pay for missing evidence and useful judgement. Match the expert to the question, preserve the scope of every claim and use AI to organise the record without smoothing away uncertainty. A smaller body of well-designed research can be more comprehensive than a much larger archive of repeated interviews.
Primary sources checked for this edition
- Anthropic — Ten financial-workflow agent templates announced. Source date: 2026-05-05. https://www.anthropic.com/news/finance-agents
- OpenAI — Financial services research, models and client material release. Source date: 2026-09-10. https://openai.com/index/introducing-chatgpt-financial-services/
- OpenAI — Forced Navier–Stokes result and formalisation are authors’ reported result. Source date: 2026-09-08; updated 2026-09-10. https://openai.com/index/navier-stokes-solution/
- OpenAI — New mathematics release and revision/citation protocols. Source date: 2026-10-06. https://openai.com/index/sharing-ai-progress-in-mathematics/
