Introduction
Deal teams produce decisions through a chain of evidence, judgement, calculation, challenge and approval. A family-office investment team may screen an opportunity, test portfolio fit, interrogate a manager, reconcile a model, write an investment-committee memorandum and monitor the position. A fund manager may originate an asset, manage a data room, coordinate diligence, prepare a valuation case, draft committee materials, answer limited-partner questions and preserve a record of the decision. Each output inherits the quality and permissions of the material that preceded it.
Generative artificial intelligence can assist with several parts of that chain. It can retrieve approved material, extract terms, compare documents, draft summaries, structure questions, prepare first-pass narrative and route exceptions. Those capabilities do not establish a productivity premium by themselves. A faster draft that requires more review, loses source traceability, exposes confidential information or introduces a material error can reduce economic value. The relevant unit is a quality-adjusted, accepted deal-work packet: a bounded output that a named reviewer accepts after required evidence, calculations, conflicts and permissions have passed their controls.
The external evidence supports a measured approach. In a preregistered experiment involving 453 college-educated professionals completing writing tasks, Noy and Zhang reported a 40% reduction in completion time and an 18% increase in output quality for the group given ChatGPT [1]. In a field study of 5,179 customer-support agents, Brynjolfsson, Li and Raymond reported a 14% average increase in issues resolved per hour, with larger gains among novice and lower-skilled workers [2]. In an experiment involving 758 consultants, AI-assisted participants completed more tasks, worked faster and produced higher-rated work for tasks inside the tested capability frontier; performance fell for a task outside that frontier [3]. These studies examine writing, service and consulting tasks. They do not measure investment returns, deal completion, fiduciary outcomes or realised revenue for the two target ICPs in this paper.
Evidence published since those early studies adds important boundaries. A field experiment across 66 organisations and 7,137 knowledge workers found that active users of an integrated generative-AI tool spent two fewer hours on email per week, while the researchers did not detect changes in the overall quantity or composition of tasks from individual access alone [5]. A randomised study of experienced open-source developers working on their own repositories found that early-2025 tools increased completion time by 19% in that setting [6]. METR's 2026 follow-up reported indications of modest later speed-ups, while explicitly describing selection and measurement problems that prevented a reliable estimate [7]. Productivity is therefore a property of a task, workflow, population, tool configuration and measurement period. It is not a transferable percentage attached to a product.
Financial-services adoption is already material. The Bank of England and Financial Conduct Authority received 118 responses in their 2024 survey; 75% of respondents said they were using AI and another 10% planned to use it within three years [8]. Foundation models represented 17% of reported AI use cases, operations and IT represented the largest business area, and 46% of respondents described only a partial understanding of the AI technologies they used [8]. The same survey identified data-related issues across four of the five highest perceived current risks and reported cybersecurity as the highest perceived systemic risk [8]. These results describe the survey population and respondents' perceptions. They do not establish an outcome for a specific family office or fund manager.
This paper develops an operating and measurement system for Family-Office CIOs and Heads of Alternatives (A2) and GCC Fund Managers or GPs Raising Capital (B2). It answers five questions:
- Which deal tasks are suitable for a copilot, and which decisions remain outside autonomous scope?
- How should a firm establish a baseline and measure speed, quality, throughput, review burden and reallocation?
- What architecture preserves source lineage, numerical control, confidentiality, recordkeeping and named authority?
- How should the productivity result be translated into capacity value, cost, risk and potential revenue without unsupported attribution?
- What staged adoption path can produce credible evidence before scale?
The central proposition is testable: a deal-team copilot creates economic value only when it raises accepted, quality-adjusted output or releases capacity that the organisation can verify and redeploy. The proposed scorecard records the full conversion from model interaction to accepted output, realised capacity and approved economic attribution. At launch, Attributed revenue: USD 0. Attributed cost reduction: USD 0. Quantified loss reduction: USD 0. Those fields change only after a named authority accepts observed evidence and its attribution method.
Scope, Definitions And Evidence Boundaries
What this paper means by an AI copilot
An AI copilot is an assistive system operating inside a human-owned workflow. It can combine a language model with retrieval, document parsing, deterministic calculations, workflow tools and policy controls. It does not possess delegated investment authority under this framework. It may propose, organise or test work; the accountable professional owns the decision, the communication and the release.
The term deal team covers the professionals and approved specialists producing an investment, financing, fund-raising or transaction decision. It includes investment professionals, capital-formation professionals, finance, legal, tax, compliance, operations and external advisers where their work is part of the controlled process. The proposed measurements apply to work packets rather than job titles, because the same professional may perform high-volume drafting and high-stakes judgement within one day.
The productivity premium is the measured difference between a governed copilot workflow and an approved comparator after speed, quality, acceptance, review burden, failure cost and realised reallocation have been considered. It is expressed separately as time, capacity, quality, throughput and economics. Combining them prematurely can hide a trade-off.
| Term | Operational definition | Required evidence |
|---|---|---|
| Attempted packet | A bounded task submitted to the copilot workflow | Packet identifier, owner, task class, start time |
| Completed packet | An output returned with the required fields | Output, version, elapsed time, tool record |
| Accepted packet | A completed packet approved by the named reviewer | Review result, exceptions, approval time |
| First-pass acceptance | Acceptance without material correction | Reviewer classification and correction log |
| Material defect | An error that could alter a decision, disclosure, recipient action or compliance outcome | Defect record, severity, affected source or calculation |
| Gross time saved | Baseline median time less treatment median time for comparable accepted packets | Matched or randomised observation |
| Realised capacity | Verified time released and demonstrably redeployed or removed | Time record plus approved reallocation evidence |
| Productivity premium | Quality-adjusted accepted output per paid hour relative to the comparator | Baseline, treatment and quality denominator |
| Economic attribution | Approved connection between measured change and revenue, cost or avoided loss | Attribution method, finance approval, period and caveats |
Evidence hierarchy
This paper separates four kinds of evidence. Tier 1 is a randomised or preregistered study with observable task outcomes. Tier 2 is a large field study, official survey or peer-reviewed empirical paper. Tier 3 is an official regulatory, standards or supervisory publication. Tier 4 is a vendor study, working paper, management estimate or proposed Matchpoint framework. Each tier can inform a decision, but it cannot answer questions outside its design.
The evidence cut-off is 1 August 2026. Working papers are labelled. Vendor research is not treated as independent validation. Regulatory statements are reported for their stated jurisdiction and scope. The paper does not provide legal advice and does not determine whether a particular system, firm or use case is subject to a specific rule.
What the external studies do and do not show
| Source and setting | Reported result | Boundary for deal teams |
|---|---|---|
| Noy and Zhang; 453 professionals, incentivised writing tasks [1] | Average time decreased 40%; quality increased 18% | Useful benchmark for bounded professional drafting; not a deal, valuation or revenue study |
| Brynjolfsson, Li and Raymond; 5,179 support agents [2] | Productivity increased 14% on average; 34% for novice and lower-skilled workers | Shows heterogeneity and possible knowledge diffusion; work was customer support |
| Dell'Acqua et al.; 758 consultants [3] | More tasks, higher speed and higher quality inside the tested AI frontier; lower correctness outside it | Closest task structure to analytical deal work; capability boundary and task design remain decisive |
| Dell'Acqua et al.; 776 product professionals [4] | Individuals using AI matched the performance of teams without AI on the tested product-innovation challenge | Supports testing collaboration structures; not evidence that AI replaces transaction teams |
| Dillon et al.; 7,137 workers across 66 firms [5] | Active users spent two fewer hours on email per week; no detected change in total task quantity or composition | Time saved can remain unconverted unless workflows and reallocation change |
| METR; 16 experienced developers and 246 repository tasks [6] | AI use increased completion time by 19% in the tested early-2025 setting | Demonstrates that expertise, context and tool overhead can reverse the expected effect |
| METR 2026 update [7] | Later results suggested small speed-ups but were described as unreliable because of selection and measurement effects | Measurement design can become invalid when users select out or run parallel agents |
| GitHub; 95 developers on a controlled coding task [23] | Vendor study reported 55% faster completion for the Copilot group | Relevant as product-sponsored task evidence; not an independent finance benchmark |
The product-innovation teamwork experiment adds evidence on human-AI collaboration [4]. The GitHub result adds a vendor-conducted coding benchmark and is labelled accordingly [23]. The studies collectively support three design choices. First, results must be reported by task and experience cohort because averages conceal heterogeneity. Second, a credible pilot requires an outcome that includes quality and reviewer effort. Third, workflow redesign and reallocation must be measured after the individual task effect; an organisation cannot book a capacity benefit merely because a draft was generated faster.
Financial-services context
FINRA states that its rules and the federal securities laws continue to apply when member firms use generative AI [9]. Its 2020 report describes AI uses across customer communication, investment processes and operations, while highlighting governance, model risk, data, privacy, cybersecurity, outsourcing and supervisory considerations [10]. IOSCO's 2025 consultation describes capital-markets use of large language models in information and process management, internal knowledge search, document standardisation, meeting materials and code-related work; it also examines risks to investor protection, market integrity and financial stability [11]. ESMA's 2024 statement identifies potential benefits and risks when firms use AI in investment services and says management bodies remain responsible for decisions [12].
The SEC's 2024 settlements with two investment advisers concerned false or misleading statements about how they used AI [13]. Separate SEC recordkeeping actions against financial firms reinforce that electronic communications and required records remain subject to preservation obligations [21]. These publications do not prohibit controlled internal assistance. They establish a strong reason to connect every released claim, client communication and decision record to evidence, supervision and policy.
For UAE-based organisations, Federal Decree-Law No. 45 of 2021 provides the federal personal-data protection framework described by the UAE Government portal [16]. Other regimes may apply depending on establishment, data subjects, free-zone status, clients and processing. The EU AI Act applies according to its stated scope, including certain providers and deployers in or connected to the Union [17]. A firm should obtain qualified advice for its exact facts before relying on a jurisdictional conclusion.
The Deal-Team Frontier
The task, not the job, is the unit of adoption
A single investment professional may complete tasks with very different risk profiles. Extracting a maturity date from a signed agreement is bounded and verifiable. Reconciling EBITDA adjustments requires accounting context and deterministic calculations. Judging the credibility of management, assessing a conflict, setting a valuation range or approving a commitment involves material professional authority. A useful adoption plan decomposes the workflow into packets with explicit evidence, tests and owners.
Three dimensions determine initial suitability:
- Verifiability: Can the output be tested against authoritative sources or deterministic rules?
- Materiality: Could an error alter capital allocation, a client communication, a legal right or a regulatory outcome?
- Context stability: Is the task repeatable with stable definitions, or does it depend on tacit knowledge and a changing transaction narrative?
The resulting frontier is deliberately conservative. Green tasks can enter a controlled pilot after data and permission approval. Amber tasks require stronger review, deterministic checks and restricted release. Red tasks remain human decisions; the copilot may assemble evidence or questions without making the decision.
| Deal task | Initial zone | Copilot role | Human control |
|---|---|---|---|
| Data-room indexing and document classification | Green | Tag approved files, detect duplicates, identify missing classes | Data-room owner confirms population and permissions |
| Contract-term extraction | Green | Extract named terms with page and clause citations | Reviewer verifies against executed version |
| Meeting-note structuring | Green | Convert approved transcript into actions and open questions | Meeting owner edits, approves and controls distribution |
| DDQ and RFI first draft | Green | Retrieve approved prior answers and draft cited response | Function owner verifies currency, scope and recipient |
| Comparable-company evidence pack | Amber | Assemble dated sources and normalise fields | Analyst validates source date, units, adjustments and population |
| Financial-model commentary | Amber | Explain approved model outputs and surface inconsistencies | Model owner controls formulas, assumptions and conclusions |
| Investment-memo first draft | Amber | Structure evidence, alternatives, risks and unresolved questions | Deal lead owns thesis, challenge, recommendation and sign-off |
| LP or client communication | Amber | Draft from approved facts and disclosure library | Authorised person approves accuracy, balance and distribution |
| Valuation conclusion | Red | Assemble evidence and sensitivity questions | Qualified professional determines and approves value |
| Investment recommendation or commitment | Red | Prepare record and challenge prompts | Investment committee retains decision authority |
| Conflict waiver or legal interpretation | Red | Retrieve approved policy and route issue | Compliance or qualified counsel decides |
| Autonomous external negotiation or release | Red | No autonomous action under this framework | Named authorised professional communicates or executes |
Before-and-after workflow
The manual workflow often moves through email, shared drives, spreadsheets, chat, decks and memory. Provenance fragments across versions. Reviewers spend time rebuilding the population and locating support. A governed copilot can create a packet with a stable identity, source manifest, draft, exceptions and decision record. The objective is a shorter, more auditable route to acceptance.
| Stage | Common manual state | Governed copilot state | Acceptance evidence |
|---|---|---|---|
| Intake | Request arrives without a stable identifier | Packet receives deal, entity, task, owner, deadline and materiality fields | Complete intake record |
| Retrieval | Analyst searches several repositories | Permission-aware retrieval returns approved sources and exclusions | Source manifest and population check |
| Production | Facts, calculations and narrative are mixed | Extracted facts, deterministic calculations, model inferences and draft narrative remain separate | Field-level lineage and calculation outputs |
| Challenge | Review occurs through comments and memory | Contradictions, missing evidence and policy triggers are explicit exceptions | Exception register and resolution |
| Approval | Approval may be implicit in circulation | Named authority accepts, rejects or returns the packet | Signed decision state |
| Release | Output is copied to email, deck or portal | Approved version and recipients pass a release gate | Release log and retained record |
Failure modes that erase the premium
The most common measurement error is to count model speed while ignoring the downstream work it creates. Hallucinated citations, stale comparables, unit errors, missing pages, duplicated entities, uncontrolled assumptions and prose that overstates evidence all increase review time. Retrieval can fail silently when permissions, indexing or document versions are wrong. A fluent answer can also encourage automation bias, causing reviewers to inspect less carefully.
The University of Chicago page for a withdrawn working paper on large-language-model financial-statement analysis states that co-authors identified inconsistencies in underlying data and analyses and withdrew the paper while reviewing the findings [24]. That event does not establish a general conclusion about LLM financial analysis. It illustrates why reproducibility, source retention and independent checks matter when claims are financially consequential.
NIST's AI Risk Management Framework organises risk work around Govern, Map, Measure and Manage [14]. Its Generative AI Profile identifies risks that include confabulation, data privacy, information integrity, information security, intellectual property and value-chain or component integration, and proposes actions across the same functions [15]. The OECD AI Principles and NIST Cybersecurity Framework 2.0 provide additional cross-sector governance and cybersecurity reference points [20,22]. The framework in this paper translates those principles into the packet, metric and release controls used by deal teams.
Measuring The Productivity Premium
The measurement chain
The measurement chain has six gates:
- Eligible task: The task belongs to an approved use-case class and data boundary.
- Completed output: The system returns the required fields within the measurement window.
- Accepted output: A named reviewer accepts the output under a defined rubric.
- Quality adjustment: Material defects, rework, reviewer time and delayed exceptions are included.
- Realised capacity: The organisation verifies how released time was redeployed or removed.
- Economic attribution: Finance or management approves the link to capacity value, cost, revenue or avoided loss.
Each gate has a denominator. Reporting only successful attempts creates survivorship bias. The dashboard therefore includes all eligible packets, withdrawals, system failures, rejected outputs and manual fallbacks.
The primary productivity measure is:
Quality-adjusted accepted output per paid hour = accepted packets x quality weight / (producer time + reviewer time + remediation time).
The relative productivity premium is:
Premium = treatment quality-adjusted output per paid hour / comparator quality-adjusted output per paid hour - 1.
This definition prevents a faster producer from appearing productive when reviewer or remediation time rises. The quality weight should be fixed before the pilot and based on observable fields. A simple scheme can assign 1.00 to an accepted packet with no material defect, 0.75 to an accepted packet with material correction before use, and 0 to a rejected or post-release materially defective packet. A firm may choose another rubric and should retain it unchanged during a measurement period.
Baseline and treatment design
A credible pilot needs comparable tasks and enough observations to distinguish signal from transaction mix. Random assignment at packet level is strongest when contamination can be controlled. A matched crossover can be used where the same professional performs repeated work and withholding the tool is operationally acceptable. Team-level assignment may be necessary when collaboration makes packet-level contamination unavoidable.
| Design element | Minimum specification | Reason |
|---|---|---|
| Population | Named teams, roles, locations and experience bands | Heterogeneous effects are expected [2,3] |
| Task classes | Predefined packet taxonomy and materiality | Prevents easy tasks entering only the treatment group |
| Comparator | Current approved workflow, versioned before pilot | Establishes what the tool is compared with |
| Assignment | Random, matched or staggered with documented method | Reduces selection bias |
| Start and stop | System timestamps plus active-time convention | Makes time definitions reproducible |
| Quality rubric | Pre-registered fields and defect severity | Prevents standards changing after results appear |
| Reviewer allocation | Balanced or independently sampled reviewers | Avoids reviewer leniency driving the result |
| Contamination | Record outside AI use and cross-team sharing | Protects comparator integrity |
| Failure capture | Include time-outs, fallbacks and rejected packets | Prevents survivorship bias |
| Analysis | Report median, distribution, confidence interval and cohort | Avoids an average masking dispersion |
For a bounded first programme, the paper proposes a four-week baseline, a two-week shadow period and an eight-week controlled pilot. The duration is a proposed design, not a universal requirement. Baseline data should cover at least one complete cycle for the selected tasks. The shadow period runs the copilot without releasing its output, which helps identify missing sources and control failures before production. The controlled pilot releases only accepted packets within the approved scope.
Metric dictionary
| Metric | Numerator | Denominator | Decision use |
|---|---|---|---|
| Completion rate | Completed packets | Eligible attempted packets | Tool and workflow reliability |
| Acceptance rate | Accepted packets | Completed packets | Practical usefulness |
| First-pass acceptance | Accepted without material correction | Completed packets | Draft quality and review burden |
| Median cycle time | Median end-to-end elapsed time | Accepted packet | Service speed |
| Active producer time | Logged or sampled producer minutes | Accepted packet | Direct labour input |
| Reviewer time | Reviewer minutes | Accepted packet | Hidden transfer of effort |
| Material defect rate | Packets with material defect | Released packets | Decision and compliance risk |
| Citation validity | Claims supported by correct cited source | Claims requiring support | Evidence quality |
| Numeric consistency | Numeric fields passing deterministic checks | Numeric fields tested | Model and unit integrity |
| Exception latency | Time from exception creation to resolution | Resolved exception | Workflow friction |
| Realisation rate | Verified redeployed or removed hours | Gross measured hours released | Conversion to organisational value |
| Adoption rate | Active eligible users | Eligible users | Behavioural reach |
| Economic attribution rate | Approved attributed value | Modelled gross value | Discipline of value claims |
Speed and quality should be reported side by side. Throughput is useful when demand exists, yet more output can include unnecessary work. A firm should define the service-level or decision need before treating volume as valuable.
From time to capacity and economics
Gross time release for task class i is the observed difference between comparator and treatment producer time for accepted, comparable packets, multiplied by accepted treatment volume. Reviewer and remediation time are then deducted. Realised capacity applies a verified reallocation factor:
Realised capacity hours = net measured hours released x verified reallocation rate.
Capacity value can be modelled using an approved loaded hourly cost or an opportunity-value method. It is not cost reduction unless payroll, contractor spend or another cost line actually changes and finance approves the attribution. Revenue attribution requires a defined causal mechanism, such as additional qualified opportunities worked, faster accepted response, increased win rate or more fee-bearing capacity. The model should retain the revenue field at zero until observed conversion and attribution data exist.
Net measured economic value = approved capacity value + approved incremental contribution + approved avoided external cost + approved loss reduction - licences - integration - governance - training - incremental review - incident and remediation cost.
Every term should show its evidence state: observed, approved model, unverified assumption or zero pending evidence. Mixing these states in one headline ROI figure is prohibited by this framework.
Controlled Architecture And Tool Stack
Design principle
The architecture starts with transaction identity, permissions and approved sources. The model is one component inside the workflow. Retrieval, calculations, policy checks, review and release remain separately observable. A user should be able to reconstruct which source versions, prompts, tools and validations produced an accepted packet.
The proposed stack has eight layers:
- Identity and policy: user, role, deal, client, entity, jurisdiction, data classification, permitted purpose and retention policy.
- Source layer: approved virtual data room, CRM, portfolio system, research library, policies, executed agreements and financial models.
- Ingestion and lineage: file hash, version, page, table, date, source owner, permission and supersession state.
- Retrieval: permission-aware search that returns passages and exclusions, with a population or coverage test where completeness matters.
- Model gateway: approved model and configuration, prompt template, token and tool limits, data-use settings and version pinning where available.
- Deterministic services: calculations, unit checks, reconciliations, entity resolution, date logic and policy rules outside free-form generation.
- Evaluation and observability: citations, field-level checks, exceptions, latency, cost, model changes, reviewer decisions and incidents.
- Human authority and release: named reviewer, approval state, recipients, channel, retained record and rollback or withdrawal action.
| Architecture layer | Failure exposure | Required control evidence |
|---|---|---|
| Identity and policy | Wrong user, deal or purpose | Authentication, role, deal membership, purpose and classification |
| Source layer | Unapproved, stale or incomplete evidence | Source register, owner, version, inclusion and exclusion record |
| Ingestion and lineage | Lost page, table or version context | File hash, locator, extraction test and supersession rule |
| Retrieval | Missing or unauthorised context | Permission test, retrieval evaluation, population reconciliation |
| Model gateway | Model change, uncontrolled data use or tool access | Approved provider, terms, settings, version and allow-list |
| Deterministic services | Arithmetic, unit or entity error | Test cases, reconciliation, locked formula and exception result |
| Evaluation | Fluent error passes unseen | Rubric, sampled review, red-team case and drift threshold |
| Release | Wrong recipient or unapproved claim | Named authority, release gate, distribution record and retention |
Source-grounded drafting
Source-grounded drafting requires more than attaching citations. Each factual proposition should link to the exact authoritative source version and locator used. The system should distinguish a source fact from a calculation, a model inference, a professional judgement and an unverified management assumption. A citation is invalid when the source exists but does not support the claim.
For completeness-sensitive work, retrieval must be tested as a population problem. If a task asks for every change-of-control clause across 120 contracts, finding plausible examples is inadequate. The packet should reconcile the expected contract population, processed population, unreadable files, excluded versions and unresolved exceptions. If completeness cannot be established, the output states that boundary before any summary.
The same rule applies to comparable companies, portfolio exposures, DDQ responses and regulatory obligations. Search ranking is a discovery mechanism. It is not proof of completeness.
Numerical control
The language model may explain an approved calculation. Material arithmetic should be performed by a deterministic spreadsheet, code service or controlled financial model with versioned inputs. The packet records units, currencies, dates, signs, scaling, source cells and reconciliation. Any number copied into narrative should be checked against the approved output.
The minimum numerical gates are:
- required input population reconciled;
- currency and unit explicit;
- formula version identified;
- balance, ownership or cash-flow checks passed where applicable;
- sensitivity assumptions separated from observed inputs;
- narrative numbers reconciled to tables and model outputs;
- reviewer confirms material adjustments and judgement.
This division keeps the copilot useful for explanation and exception routing while preserving the calculation engine as the numerical authority.
Confidentiality, privacy and security
Deal information can include personal data, material non-public information, commercial secrets, bank details, beneficial ownership and privileged advice. Each use case should therefore have an approved data boundary before prompts or retrieval are enabled. Public consumer tools should not receive deal data unless the organisation has approved the exact service, contract, configuration, location, retention and use terms for that information.
NIST's GenAI Profile recommends governance and measurement across the AI lifecycle and discusses data privacy, information security, confabulation, information integrity, human-AI configuration and value-chain risks [15]. The IMF's 2023 note identifies privacy, bias, opacity, robustness, cybersecurity and potential systemic effects in finance [18]. Its 2026 work on AI and financial-sector cybersecurity emphasises third-party concentration, shared digital infrastructure and the potential for faster propagation of vulnerabilities [19]. These publications support controls for least privilege, data segregation, provider assessment, monitoring, recovery and exit.
Prompt injection and hostile document content require a specific boundary. Retrieved documents are untrusted data, not instructions. Tool calls should use typed allow-listed functions with scoped permissions. A document cannot authorise a payment, external message, model change, file deletion or recipient change. High-impact actions remain outside autonomous scope and require a separate human approval state.
Model and vendor change
A copilot can change without the workflow owner changing code. Provider model updates, retrieval changes, prompt edits, embedding changes and source-index rebuilds can affect results. The operating model therefore treats each significant component version as part of the packet evidence.
A release candidate should pass a fixed evaluation set containing ordinary cases, edge cases and known failures. Regression thresholds should cover citation validity, numeric consistency, material-defect rate, refusal or escalation, latency and cost. A model update enters production only after the owner accepts the evaluation and records any changed limitation. Provider concentration and exit plans should be reviewed in proportion to materiality.
Icp Playbooks
A2: Family-Office CIOs and Heads of Alternatives
The family-office use case is decision support across a private, often concentrated portfolio. The information set can span fund documents, direct investments, private banks, custodians, managers, operating businesses and family governance. Value comes from better evidence assembly, faster comparison and more consistent monitoring. The highest-risk failure is a fluent answer that obscures a missing source, stale valuation, conflict or concentration.
| A2 workflow | Bounded copilot use | Core measure | Authority retained by people |
|---|---|---|---|
| Opportunity triage | Build cited summary, mandate fit and open-question list | Time to accepted screen; source coverage | CIO or delegate decides whether to proceed |
| Manager due diligence | Reuse approved DDQ library, compare versions, surface inconsistencies | First-pass acceptance; unresolved exceptions | Investment team evaluates manager and recommendation |
| Investment committee memo | Structure thesis, alternatives, risks, portfolio fit and evidence | Reviewer time; citation validity; material defects | Deal lead and committee own recommendation and decision |
| Portfolio monitoring | Extract notices, covenants and reported changes; route exceptions | Exception latency; coverage; false-negative sampling | Asset owner determines action and valuation implications |
| Board or family reporting | Draft from approved portfolio and decision records | Preparation time; numeric reconciliation | Authorised family-office leadership approves distribution |
The family office should begin with repeatable packets where evidence is already controlled. A manager-update comparison or investment-memo source manifest is safer than autonomous portfolio advice. Portfolio construction, liquidity planning, related-party conflicts, tax, legal and succession conclusions require their relevant qualified authorities.
B2: GCC fund managers or GPs raising capital
The fund-manager use case combines investment activity with capital formation, operational due diligence and ongoing LP communication. The copilot can reduce duplication across DDQs, data rooms, quarterly letters, investment-committee materials and pipeline management. It can also increase risk if a stale answer, inconsistent metric or unsupported performance statement reaches an investor.
| B2 workflow | Bounded copilot use | Core measure | Authority retained by people |
|---|---|---|---|
| DDQ and RFP response | Retrieve approved answer blocks, flag expiry and draft cited response | Cycle time; reuse; owner corrections | Functional owners approve each response |
| Data-room readiness | Classify documents, reconcile required index and flag missing items | Population coverage; exception closure | CFO, COO, counsel or authorised owner approves room |
| Fund-raising materials | Check consistency across strategy, team, track record and terms | Numeric consistency; claim support | Authorised principals and counsel approve marketing content |
| LP meeting preparation | Assemble relationship history, open questions and approved updates | Preparation time; source freshness | Relationship owner determines messaging |
| Investment committee | Produce evidence manifest, first draft and challenge list | Accepted-packet throughput; reviewer time | Investment professionals and committee decide |
| Quarterly reporting | Draft narrative from approved performance and portfolio records | Reconciliation; review time; late corrections | CFO, fund administrator and authorised signatory approve |
FINRA's technology-neutral reminder and the SEC's AI-washing actions are US-specific publications, yet their control logic is useful more broadly: existing obligations and truthfulness remain relevant when AI is used [9,13]. A GCC manager should map the exact rules, fund documents, investor jurisdictions and marketing permissions with qualified advisers.
Shared service design
Both ICPs benefit from a shared packet schema:
- deal or fund identity;
- task class and materiality;
- requester, producer, reviewer and decision authority;
- approved source population and exclusions;
- facts, calculations, inferences and judgements in separate fields;
- open questions and contradictions;
- required legal, compliance, tax or specialist review;
- outcome, corrections and release state;
- timing, model, tool and cost telemetry;
- retained evidence and disposal rule.
The schema allows a small team to accumulate reusable knowledge without turning prior text into an ungoverned source. Reuse occurs through approved answer blocks, policies, templates and structured facts with owners and expiry dates.
Illustrative Economics
Scenario status
[Unverified illustrative management assumptions] The scenarios in this section are worked examples. They are not observed Matchpoint, family-office, fund-manager or market data. They do not represent a forecast, promise, benchmark or quoted vendor price. Each organisation should replace every input with approved observed evidence.
The illustrative team has 12 professionals, 220 working days per year and 8 paid hours per day, producing 21,120 annual paid hours. The loaded hourly value is assumed at AED 650. The model separates task eligibility, direct task-time reduction, active adoption, quality acceptance and verified reallocation. Annual licences, integration, governance, training and incremental review are grouped as programme cost. No revenue, cost reduction or loss reduction is attributed.
| Assumption | Conservative | Base | Upside | Evidence state |
|---|---|---|---|---|
| Annual paid hours | 21,120 | 21,120 | 21,120 | Unverified illustrative assumption |
| Share of time in eligible packets | 25% | 35% | 45% | Unverified illustrative assumption |
| Direct task-time reduction | 10% | 20% | 30% | Unverified illustrative assumption |
| Active adoption or coverage | 60% | 75% | 85% | Unverified illustrative assumption |
| Quality acceptance factor | 80% | 90% | 95% | Unverified illustrative assumption |
| Verified reallocation factor | 50% | 65% | 75% | Unverified illustrative assumption |
| Loaded value per realised hour | AED 650 | AED 650 | AED 650 | Unverified illustrative assumption |
| Annual programme cost | AED 250,000 | AED 310,000 | AED 390,000 | Unverified illustrative assumption |
Worked outputs
| Modelled output | Conservative | Base | Upside |
|---|---|---|---|
| Eligible annual hours | 5,280 | 7,392 | 9,504 |
| Direct gross hours released | 528 | 1,478 | 2,851 |
| Hours after adoption and quality | 253 | 998 | 2,302 |
| Realised capacity hours | 127 | 649 | 1,727 |
| Modelled realised capacity value | AED 82,368 | AED 421,621 | AED 1,122,388 |
| Less annual programme cost | AED 250,000 | AED 310,000 | AED 390,000 |
| Modelled net capacity value | AED (167,632) | AED 111,621 | AED 732,388 |
| Attributed revenue | USD 0 | USD 0 | USD 0 |
| Attributed cost reduction | USD 0 | USD 0 | USD 0 |
| Quantified loss reduction | USD 0 | USD 0 | USD 0 |
Rounding explains small differences between displayed intermediate values and calculated totals. The model shows why a percentage speed gain is insufficient. Eligibility, adoption, quality and reallocation compound. In the conservative case, the programme has negative modelled net capacity value. The base case becomes positive only after a 65% reallocation assumption. The upside case depends on broad eligibility and strong adoption that require evidence. No scenario should be presented externally as an achieved ROI.
Revenue channel
The Topic Tracker hook asks how the technology can multiply productivity and revenue. The proposed revenue logic is a measurement route:
Released qualified capacity -> additional or faster qualified work -> observable client or investment outcome -> approved contribution attribution.
For a fund manager, observable intermediate measures could include complete qualified LP follow-ups, approved DDQ response time, accepted data-room readiness and additional substantive meetings handled without a fall in quality. For a family office, useful measures could include additional screened opportunities, faster completion of accepted diligence packets and earlier closure of material exceptions. These are operational measures. Revenue or investment performance requires a further causal link and should remain unclaimed until it is demonstrated.
An attribution design can compare contemporaneous matched cohorts, staggered rollout or pre-specified changes in eligible throughput. It should control for market conditions, team seniority, transaction mix, fund-raising stage, seasonality and parallel process changes. Management approval of an estimate does not convert it into observed fact; the dashboard should retain its evidence label.
Risk, Governance And Acceptance
Risk register
| Risk | Leading indicator | Gate or response | Accountable owner |
|---|---|---|---|
| Unsupported factual claim | Citation-validity failure | Block release; correct source or remove claim | Packet reviewer |
| Numeric inconsistency | Reconciliation or unit failure | Re-run deterministic check; block narrative | Model owner |
| Missing document population | Expected and processed counts differ | State incompleteness; route exception | Data-room owner |
| Confidential-data exposure | Policy or data-loss-prevention alert | Stop processing; contain and investigate | Information-security owner |
| Prompt injection or tool abuse | Untrusted instruction or unexpected tool request | Isolate content; deny action; review incident | AI platform owner |
| Automation bias | Falling reviewer time with rising correction or incident rate | Increase independent sample; retrain or pause | Business owner |
| Model drift | Evaluation threshold breached after component change | Roll back or reapprove | Model owner |
| Stale reusable content | Owner or expiry missing | Remove from retrieval until renewed | Knowledge owner |
| Misleading AI or performance claim | Unsupported external statement | Block release; legal or compliance review | Communications owner |
| Vendor concentration or outage | Service-level or exit trigger | Use tested fallback; review continuity | Technology and procurement owners |
| Recordkeeping gap | Missing prompt, output, approval or distribution record | Block or reconstruct before release | Compliance or records owner |
| Unproved economic value | Modelled value presented as achieved | Correct label; keep attribution at zero | Finance or management authority |
Acceptance gates
A use case enters controlled production when its owner accepts all applicable gates:
- Purpose: the business purpose, user population and prohibited uses are explicit.
- Data: sources, permissions, personal data, confidentiality, locations and retention are approved.
- Evidence: retrieval and population tests meet the threshold for the task.
- Quality: citation, numeric and material-defect thresholds pass on the fixed evaluation set.
- Security: prompt-injection, data-exfiltration, access and tool-abuse tests pass.
- Human authority: reviewers, decision owners and escalation routes are named.
- Operations: latency, availability, cost, fallback, incident and rollback procedures are tested.
- Measurement: comparator, assignment, metrics and analysis are pre-specified.
- Release: recipient, channel, recordkeeping and disclosure requirements are enforced.
- Economics: assumptions and observed values remain separate; attribution authority is named.
The owner may accept a residual risk within delegated authority. Material residual risk outside that authority is escalated. An expired or failed gate returns the use case to shadow mode or manual fallback.
Governance cadence
Daily operations review failed packets, security alerts and blocked releases. Weekly product review examines exceptions, task-level performance, user feedback and evaluation drift. Monthly business review examines quality-adjusted productivity, realised capacity, cost and incidents by task and cohort. Quarterly governance review confirms use-case inventory, provider dependencies, legal and regulatory change, data boundaries, control effectiveness and continued authority.
The Bank/FCA survey reported that 84% of responding firms had an accountable person for their AI framework and that accountability was often distributed across several persons or bodies [8]. The proposed cadence reflects that operating reality while requiring a single named owner for each use case and packet.
Adoption Roadmap
Days 0 to 30: define and baseline
- appoint business, technology, data, security and compliance owners;
- select two or three repeatable green or amber packet classes;
- define prohibited data, actions and releases;
- version the current manual workflow and quality rubric;
- collect baseline producer, reviewer, quality and failure data;
- build the approved source and reusable-content register;
- create the fixed evaluation set and incident taxonomy.
The output is an approved pilot charter, baseline dataset and go or no-go decision for shadow operation.
Days 31 to 60: build and shadow
- connect only approved sources using least privilege;
- implement packet identity, lineage, deterministic checks and reviewer workflow;
- run retrieval, completeness, prompt-injection and numerical tests;
- operate in shadow mode with no external release;
- compare copilot packets with the manual output;
- repair source, workflow and evaluation failures;
- approve the controlled-pilot design and user training.
The output is an evaluation report, control record and explicit production-scope decision.
Days 61 to 90: controlled pilot
- assign eligible packets according to the pre-specified design;
- capture every attempt, failure, fallback, review and correction;
- hold daily incident and weekly quality reviews;
- keep external and investment authority with named professionals;
- measure task and cohort distributions, not only averages;
- publish an internal result with confidence intervals and limitations;
- accept, amend, pause or retire each use case.
The output is an observed task-level productivity and quality result. Economic attribution remains separate.
Months 4 to 12: scale evidence, not access alone
Scaling adds use cases only after the existing workflow remains within quality and security thresholds. Reusable knowledge receives owners and expiry dates. Model and retrieval changes pass regression tests. Realised-capacity evidence is reviewed with team planning, and finance approves any cost or revenue attribution. The firm maintains a manual fallback and tested exit path for material provider dependency.
| Phase | Decision | Required evidence | Stop condition |
|---|---|---|---|
| Define | Is the packet suitable and valuable? | Purpose, baseline, task frontier, data boundary | No owner, no comparator or prohibited data |
| Shadow | Can the system produce a reviewable packet? | Evaluation results, lineage, controls, fallback | Material defect or security threshold breached |
| Pilot | Does quality-adjusted productivity improve? | Assigned observations and full denominators | Reviewer burden or failure cost erases benefit |
| Scale | Can the benefit survive broader use? | Cohort results, realised capacity, operations | Drift, concentration, incident or control failure |
| Attribute | Did economics change for the stated reason? | Finance-approved causal and accounting evidence | Model estimate presented as observed value |
Limitations And Research Agenda
The external productivity studies use different tasks, populations, tools, time periods and outcomes. Their reported percentages are not directly comparable and should not be pooled into a deal-team forecast. Tool capabilities and user behaviour also change quickly. Model updates can invalidate a prior benchmark, while learning can alter both comparator and treatment performance.
Deal work has additional measurement problems. Transaction complexity varies; difficult cases may select into human-only treatment; team members share knowledge; quality can become observable only after closing; and reviewers may change effort when they know AI was used. Revenue and investment outcomes are affected by market conditions, client relationships, capital availability and judgement. A short pilot can measure packet quality and time more credibly than long-horizon investment performance.
The illustrative economics use unverified assumptions and exclude taxes, financing, depreciation, opportunity costs outside the stated capacity method and tail losses. A negative or positive modelled result is sensitive to the eligible-time, quality and reallocation assumptions. Actual decision-makers should perform sensitivity analysis with local data.
The framework also requires legal, regulatory, tax, data-protection, employment, intellectual-property and contractual analysis for the specific firm and use case. Official publications cited in this paper do not constitute permission, approval or a complete obligation map.
Future research should publish task-level experiments for investment and capital-formation workflows, including independent quality ratings, reviewer time, error severity and longitudinal reallocation. Studies should compare source-grounded copilots, deterministic tools and general chat interfaces; report negative results; preserve evaluation sets; and examine whether gains persist after task novelty and training effects fade.
Conclusion
An AI copilot can accelerate bounded deal work when the task is verifiable, the source set is approved, deterministic calculations remain authoritative and a named professional controls release. External studies show material gains in several knowledge-work settings and material losses in another. That dispersion makes measurement a governance requirement.
For family-office investment teams, the strongest starting points are evidence assembly, manager-diligence comparison, monitoring exceptions and investment-memo support. For GCC fund managers, the strongest starting points are DDQ reuse, data-room readiness, consistency checking, LP preparation and controlled committee drafting. Valuation, investment decisions, conflicts, legal conclusions and autonomous external actions remain with authorised professionals.
The proposed productivity premium is quality-adjusted accepted output per paid hour. Its measurement includes producer time, reviewer time, remediation, defects, failures and fallbacks. Released time becomes capacity value only after verified reallocation. Revenue, cost reduction and loss reduction become attributed only after an approved causal and accounting basis exists.
The practical sequence is clear: define the task frontier, establish the baseline, build a source-grounded controlled stack, run shadow evaluation, conduct a bounded pilot, publish the full result and scale only the use cases whose evidence survives. That process can turn a technology claim into an operating fact.
