How to use this blueprint
The blueprint has five parts, applied in order to one workflow at a time. It assumes a private capital firm, or an administrator serving one, with a system of record for fund and investor data — such as Private Capital Suite — and a team that remains accountable for every output that leaves the building.
Every automated workflow follows the same path, and the control points do not move. Approval is required whenever an output leaves the firm; internal work, such as a reconciliation, can close at review.
- 01AI workflowPrepareGather inputs from the systems of record and prepare the draft.
- 02AutomatedCheckRun the agreed checks; raise exceptions instead of guessing.
- 03A personReviewCheck against sources; correct, accept or return.
- 04A personApprove & releaseRequired for anything that reaches an investor, bank or counterparty.
All examples in this blueprint are illustrative. They are not drawn from any client engagement.
01Selecting workflows
Score each candidate workflow from 1 to 3 on five criteria. The scores decide whether AI prepares the work, assists a person who prepares it, or stays out of it.
| Criterion | 1 — weak | 2 | 3 — strong |
|---|---|---|---|
| Input consistency | Bespoke or unstructured each time | Mostly standard, with some variation | Structured data or standard documents |
| Review speed | Checking takes as long as preparing | Checking takes about half the time | Checking takes a fraction of the time |
| Error containment | An escaped error changes what an investor is owed, or a filing, with no independent check | An error could reach investors or counterparties, but an independent check stands before release | An error stays internal and is easily corrected |
| Volume and frequency | Rare or one-off | Monthly or quarterly, moderate volume | Frequent or high volume |
| Data sensitivity | Needs broad access to sensitive investor or legal data | Needs a few named sensitive fields | Operational data only |
- Automate preparation — a total of 12 or more, and no criterion scored 1. AI prepares; people review and approve.
- Assist only — a total of 8 to 11, or any criterion other than error containment scored 1. A person prepares the work; AI helps gather, check and draft.
- Do not automate — a total of 7 or less, or error containment scored 1. The work stays manual until its inputs or controls change.
Where the rules overlap, the more cautious outcome applies: do not automate outranks assist only, which outranks automate preparation.
| Workflow | Inputs | Review | Errors | Volume | Data | Total | Outcome |
|---|---|---|---|---|---|---|---|
| Cash reconciliation | 3 | 3 | 3 | 3 | 3 | 15 | Automate preparation |
| Capital call notices | 3 | 3 | 2 | 2 | 2 | 12 | Automate preparation |
| Quarterly letter, first draft | 2 | 2 | 2 | 2 | 3 | 11 | Assist only |
| Waterfall and carried interest | 2 | 1 | 1 | 1 | 2 | 7 | Do not automate |
02Data and source boundaries
For each workflow, write down the system of record for every piece of data it touches, and what the workflow may and may not do with it. Access is granted to the workflow, not to the model.
| Data | System of record | The workflow may | The workflow may not |
|---|---|---|---|
| Fund, entity and investor structures | Private Capital Suite | Read, to prepare notices and reports | Change them |
| Commitments, capital activity and balances | Private Capital Suite | Read, and compare with other sources | Post or adjust entries |
| Cash and bank activity | Bank feeds and statements | Read, to reconcile | Initiate or approve payments |
| Investor identity and KYC | Onboarding system or CRM | Read named fields, such as legal name and notice address | Read identity documents or tax identifiers |
| Side letters and fund documents | Document management | Read approved extracts the workflow requires | Interpret terms without legal review |
| Valuations and waterfall results | The firm’s valuation and accounting process | Read approved results | Generate or alter them |
Four principles apply to every workflow:
- Processing stays inside the firm’s environment — for example, within the firm’s own Snowflake™ account using Snowflake Cortex — so data remains under existing access controls.
- Least access. Each workflow runs under its own service identity, limited to the rows and columns it needs.
- No retention or training by vendors. Confirmed in the contract, not assumed.
- Records follow the firm’s policy. Outputs and audit records are retained under the existing records policy.
03Reviewer and exception design
Four roles, each with something it may not do. The separation is the control.
| Role | Who | Responsible for | May not |
|---|---|---|---|
| Preparer | The AI workflow | Gathering inputs, preparing the draft, running checks, raising exceptions | Release anything, or resolve its own exceptions |
| Reviewer | A named member of the team | Checking the draft against sources; correcting, accepting or returning it | Approve investor-affecting work they reviewed |
| Approver | A senior owner, such as the controller | Approving investor-affecting outputs before release | Act as reviewer on the same output |
| Workflow owner | The accountable manager | Design, thresholds, measures and changes | Change the workflow without a record |
| Exception | Example | Routed to |
|---|---|---|
| Missing or late input | A bank statement has not arrived | Run held; workflow owner notified |
| Tolerance breach | A reconciliation difference above the agreed threshold | Reviewer, with both values and their sources |
| Low confidence | An unreadable document or an ambiguous match | Reviewer — the workflow never guesses |
| Policy trigger | A new payee, an out-of-pattern amount, a changed investor address | Reviewer and approver |
| System or model error | A timeout, a failed check, an unexpected format | Run stopped with no partial output; owner notified |
And four rules that do not bend:
- Nothing reaches an investor, bank or counterparty without a person’s approval.
- Every exception has an owner and a target response time.
- Overrides are allowed, and recorded with a reason.
- A returned draft is corrected or re-run within the record — never edited silently outside it.
04The audit record
Every run leaves a record that someone outside the team — a CFO, an auditor, an investor’s operational due diligence team — can read without help. It shows what went in, what was checked, what was flagged, who looked at it, and what they decided. Two examples:
- Workflow
- Capital call notice preparation · version 1.4
- Fund
- Example Fund III, L.P.
- Run
- August 14, 2026 · 09:12 ET
- Inputs
- Private Capital Suite extract 58213 (commitments and prior calls); call approval memo of August 13; investor notice addresses (named fields only)
- Method
- Approved model under the firm’s AI policy; notice template version 3.2
- Checks
- Call amounts sum to the total call — passed. Each amount equals commitment × call percentage — passed. No investor exceeds unfunded commitment — passed. Due date matches the memo — passed.
- Exceptions
- One policy trigger: an investor’s notice address changed in the last 30 days
- Review
- Fund accountant · 11:40 — address confirmed with investor relations; one wording correction
- Approval
- Controller · 15:05
- Outcome
- Approved with edits; released through the investor portal on August 15
- Workflow
- Cash reconciliation, operating accounts · version 2.1
- Run
- September 2, 2026 · 07:30 ET
- Inputs
- Bank statement feed for August 31; Private Capital Suite cash balances, extract 59102
- Method
- Matching rules version 2.1 — no generative model used
- Checks
- 46 of 47 accounts reconciled within tolerance
- Exceptions
- One tolerance breach: account ending 4471, difference of $18,250.00 against a $1,000 threshold
- Review
- Senior fund accountant · 10:15 — timing difference: wire received August 31, booked September 1
- Approval
- Not required — internal reconciliation
- Outcome
- Closed; exception explained, no entry needed
The examples name roles; a real record names people. The second shows something worth stating plainly: not every step needs a language model. Rules that work should stay rules.
05Measuring the result
Measure the last three cycles before launch — or one quarter — and the same measures after. Review them quarterly with the workflow owner.
| Measure | Definition | Direction |
|---|---|---|
| Hours returned | Preparation time before, less review and exception time after | Up |
| Cycle time | From inputs available to output released | Down |
| First-pass acceptance | Share of outputs approved without edits | Up |
| Exception rate | Exceptions per run, by type | Stable or down; a sudden change is a signal |
| Review time | Minutes spent reviewing each output | Down |
| Escaped errors | Errors found after approval | Zero |
| Adoption | Share of eligible runs that use the workflow | Up |
A workflow is redesigned or retired when any of these holds:
- First-pass acceptance stays below the agreed level for two consecutive quarters.
- Review and exception time exceed the preparation time saved.
- An escaped error affects an investor — the workflow is paused and investigated first.
Putting it to work
Start with one workflow that scores well and has a clear owner. Write its boundary matrix, agree the roles and thresholds, define the record, and take the baseline before anything is automated. The first quarter of measures will show whether to widen the scope — and the record will show anyone who asks exactly how the work was done.
Our workflow review applies this blueprint to one of your workflows in about two weeks. For the thinking behind it, see Where AI belongs in fund operations — and where it doesn’t.
This blueprint is general information. Its examples are illustrative and not drawn from any client engagement. It is not legal, regulatory, tax or accounting advice; firms should apply it with their own advisers and policies.