Illustrative AI workflows across the investment lifecycle: finding and evaluating ideas, underwriting and owning companies, and protecting the portfolio. Synthetic data; human review built into every step.
The companies (all fictional): nine appear in more than one workflow
Every company, ticker, and person on this page is invented. Names and tickers were checked against companies listed on US exchanges.
Calvorn Industrial Services
Industrial maintenance, private
Bought by the fund at 10.0x EBITDA; its deal team asks the Analyst Agent’s question.
One fictional theme, Grid modernization, taken from its first mentions in earnings calls to a rules-based index and its quarterly review. Two of its companies, Halden and Ostrander, agree to merge in Research.
Discover
Define
Universe
Score
Construct
Monitor
1Discover
Theme discovery
A theme builds up in earnings calls for several quarters before anyone packages it.
Transcript
Inputs. Eight quarters of earnings transcripts, filings, and news (a synthetic corpus).
AI step. Clusters recurring concepts into candidate themes, counts mentions per 100 transcripts each quarter, and places each theme on a lifecycle (emerging, accelerating, mature, fading) from two numbers: its latest level and its change over two years.Evidence: Top passages per theme, linked to transcript and page; Lifecycle rule: level and two-year change
Human checkpoint. Research lead: Selects Grid modernization to formalize and folds Grid-scale storage into it as a sub-theme.
Output. A theme map with attention trends: eight candidate themes, two of them accelerating.
What the reviewer checks. The research lead chooses from eight themes, each with its mention count, two-year trend, and source passages.
2Define
Theme definition and taxonomy
Before anyone screens for a theme, someone has to write down what counts.
Transcript
Inputs. The selected theme, Grid modernization, and the research lead’s note.
AI step. Expands the theme into four sub-themes with a concept set for each, separates enablers (companies selling into the theme) from adopters (companies spending on it), and drafts inclusion and exclusion rules.Evidence: Concept sources: passages from Discover; Enabler test: revenue from theme products or services
Human checkpoint. Analyst: Prunes two concepts that pull in unrelated companies (smart home devices, residential solar installers) and approves the taxonomy.
Output. A theme definition document: sub-themes, concepts, the enabler and adopter split, and inclusion and exclusion rules.
What the reviewer checks. Universe, scoring, and index rules all read from this document, so a change to the definition is made in one place.
3Universe
Universe construction
A seed basket of known names misses companies one step removed from the theme.
Transcript
Inputs. A seed basket of five existing holdings and a broad equity universe.
AI step. Expands from the seed to companies linked directly or indirectly, and shows the passage behind each addition. It proposes sixteen names.Evidence: Each addition: matching passage and source
Human checkpoint. Analyst: Removes three false positives: a company name match, warehouse storage racks, and a word coincidence in a transcript.
Output. A candidate universe of eighteen names, each with the filing or transcript behind it.
What the reviewer checks. The analyst reviews sixteen proposed names with their evidence on one screen and removes three.
4Score
Revenue exposure scoring
Two companies can mention the grid equally often while one earns most of its revenue there and the other earns a fifth.
Transcript
Inputs. Segment disclosures for each company in the universe.
AI step. Maps each business segment to a sub-theme and computes a revenue exposure score: theme-linked revenue divided by total revenue, floored at 0% and capped at 100%. Assigns a tier from 0 to 4.Evidence: Filing segment note per company; Tier bands: 10%, 25%, 50%, 75%
Human checkpoint. Analyst: Reviews names within three points of a tier boundary (Corrigan Switchgear and Marlow Relay) and the segment the model could not map, a licensing line at Tarnhill Controls, which stays unscored until mapped.
Output. A scored table that shows the segments behind each exposure figure.
What the reviewer checks. Brightwater Industrial talks about the grid on its calls but earns 18% of revenue there, so it scores tier 1 and stays out of the index.
5Construct
Index construction
An index committee has to be able to recompute each weight from the written rules.
Transcript
Inputs. The scored universe from Score.
AI step. The rules engine (not the model) applies eligibility (tier 2 or higher, market cap of at least $1.0B, average daily value traded of at least $5M, enablers only), weights by market cap times exposure, caps any single name at 10%, and keeps up to 12 constituents. The same inputs always produce the same index.Evidence: Methodology: eligibility, weighting, cap; 13 eligible names, 12 constituents
Human checkpoint. Index committee: Approves the constituent list, the weights, and the methodology summary.
Output. A constituent list with weights summing to 100%, and a methodology summary.
What the reviewer checks. Five names hit the 10% cap; the excess is spread across the other seven in proportion to their scores.
6Monitor
Drift and reconstitution
Companies move into and out of a theme between reviews.
Transcript
Inputs. New quarterly filings for the constituents and the watch list.
AI step. Rescores every name on the new segment data and flags exposure changes. Tarrant Engineering Group fell from 35% to 11% after selling its grid engineering unit, below the tier 2 floor. Solway Semiconductor rose from 22% to 30%.Evidence: TREG 10-Q: discontinued operations note; SLWY 10-Q: segment revenue table
Human checkpoint. Index committee: Reviews the proposed adds and drops at the scheduled quarterly reconstitution.
Output. A reconstitution memo with a diff view: SLWY added, TREG dropped, weights recomputed.
What the reviewer checks. Membership changes on a fixed calendar, so a single filing flags a name for review without triggering a trade on its own.
Find and evaluate
Research
One fictional merger, followed by a team of research agents: Halden Grid Systems agreeing to buy Ostrander Metering for $48.00 a share in cash. The question is whether it closes. Both companies sit in the Thematic grid index.
Evaluate
Scope
Retrieve
Synthesize
Monitor
Improve
1Evaluate
Choosing a model for each task
Model releases arrive every few months, and the best model for reading filings is rarely the best one for arithmetic.
Transcript
Inputs. Three candidate models and an internal benchmark of research tasks with known answers.
AI step. Runs the three candidates on the benchmark and scores each on six task types: passage retrieval, numerical reasoning, multi-step planning, citation accuracy, robustness to noisy inputs, and scenario analysis. Proposes a routing table: for each task type, the cheapest model scoring within two points of the best.Evidence: Benchmark run: graded answers per task type; Cost per task: model C at 6% of model B
Human checkpoint. Applied AI lead: Approves the routing table and asks for a larger sample on citation accuracy, where the cheaper model was chosen at two points below the top score.
Output. A routing table: two task types go to model A, two to model B, and two to model C.
What the reviewer checks. The routing table averages 83.8 across the task types, against 81.2 for the best single model, at 56% of its cost.
2Scope
Turning a question into a research plan
A portfolio manager asks whether a deal will close. The answer depends on four separate questions that need different sources.
Transcript
Inputs. The portfolio manager’s question and the announced terms of the deal.
AI step. Breaks the question into four risk dimensions (regulatory, financing, shareholder vote, and other conditions), writes the sub-questions for each, and assigns one agent per dimension with the source types it may search.Evidence: Team playbook: merger review checklist; Deal announcement: consideration and conditions
Human checkpoint. Portfolio manager: Approves the plan and adds one sub-question: which holders have signed agreements to vote for the deal.
Output. A research plan: four dimensions, seven sub-questions, and the sources each agent may use.
What the reviewer checks. The plan is agreed before any search runs, so the portfolio manager can see what will and will not be looked at.
3Retrieve
Sources, within access limits
The relevant facts sit in a merger agreement, a proxy statement, call transcripts, and broker notes, and access differs by team.
Transcript
Inputs. Filings, transcripts, broker research, and news, filtered by the team’s access rights.
AI step. Each agent searches its permitted sources and returns passages with document, section, and page. Access rules sit in the search index itself, document by document, so excluded material never reaches an agent. Text inside documents is treated as content to cite, never as instructions: one broker note carries hidden text telling the model to rate the deal certain to close, and it is flagged.Evidence: Merger agreement §7.1: conditions to closing; Access log: excluded documents; Injection check: one document flagged
Human checkpoint. Analyst: Spot-checks three citations against the source documents, confirms the excluded document was outside the team’s access, and reads the flagged broker note.
Output. A source set of ten passages across four dimensions, each with its document and section; two document excluded by access rules and one flagged for embedded instructions.
What the reviewer checks. Each claim in the next step points to a passage the analyst can open, and the hidden instruction changed no estimate.
4Synthesize
A close probability against the market
The gap between the offer price and today’s share price tells you how likely the market thinks the deal is to close. The team needs its own estimate to compare against it.
Transcript
Inputs. The cited source set, the offer price, the current share price, and the team’s estimate of where the stock would trade if the deal fell through.
AI step. Each dimension agent estimates the chance its conditions are met, with the passages behind it. The combined estimate is their product: 84%. The market-implied probability from the spread is 70%.Evidence: Merger agreement §6.3 and §7.1; Market-implied: (price − failure price) ÷ (offer − failure price)
Human checkpoint. Portfolio manager: Reviews each dimension’s estimate and its sources, and decides whether the gap to the market justifies a position.
Output. A deal memo: close probability 84% against 70% implied by the market, a gross spread of 8.8%, and the evidence for each dimension.
What the reviewer checks. The team estimate sits 14 points above the market, and the memo shows which dimension carries the difference.
5Monitor
Re-estimating as filings arrive
Deal odds change with each filing and press release, and a spreadsheet updated by hand falls behind.
Transcript
Inputs. New filings and press releases for both companies, checked as they arrive.
AI step. Compares the new filings with the last quarter’s. Of 1,240 changed sentences, a scoring rubric keeps three: a second request from the antitrust regulator (a demand for more information that lengthens the review), a deleted sentence that had said no extended review was expected, and a later expected closing date. Lowers the regulatory estimate from 90% to 74%, recomputes the close probability to 69%, and alerts because the change exceeds five points.Evidence: Ostrander Metering 8-K: second request received; Ostrander Metering 10-Q: risk factors, sentence removed; Price feed: $42.30 after the filing
Human checkpoint. Portfolio manager: Reads the alert and the filing, and decides whether to change the position.
Output. An updated memo: close probability 69% against a market-implied 56% after the price moved to $42.30.
What the reviewer checks. The analyst reads three flagged changes out of 1,240, and the estimate moved 15 points on the day the filings arrived.
6Improve
Test cases and outcomes
When an agent misses something, the lesson is lost unless it becomes a test the next version has to pass.
Transcript
Inputs. A correction from the portfolio manager and the benchmark suite.
AI step. Writes the correction up as a new benchmark case and confirms the current agent fails it. Updates the merger-review instructions, then reruns the full suite. Separately, scores the agents’ past estimates against how those deals ended.Evidence: Benchmark run before and after; Calibration: estimates against outcomes, 30 deals
Human checkpoint. Applied AI lead: Reviews the instruction change, the before-and-after results, and the calibration table, and approves the release.
Output. A suite of 49 cases, 46 passing (up from 44), with no case that passed before now failing, and a calibration table of 30 past deals.
What the reviewer checks. Time saved is easy to count. The calibration table counts results: deals the agents put at 91% on average closed 92% of the time.
Underwrite and own
Private Equity
One fictional buyout, followed from the CIM to the second year of ownership: Calvorn Industrial Services, a B2B maintenance services company bought at 10.0x EBITDA. Calvorn comes back in the Analyst Agent, Scenarios, and Early Warning.
Set Up
Build
Forecast
Return
Stress-Test
Monitor
1Set Up
From CIM to model shell
The CIM carries three years of history across prose, charts, and tables. The fund template wants those figures as inputs on one tab, each with a page reference.
Transcript
Inputs. The Calvorn Industrial Services CIM (fictional, 84 pages) and the fund’s model template.
AI step. Extracts three years of history (technicians, utilization, bill rate, cost per technician, SG&A) and the balance sheet items the model needs, each figure linked to its page. Each figure is read twice, in two separate passes, and scored for confidence: 18 of 21 agree above 95%, and three are routed for review. Lays out the shell in the fund’s tab order and applies spreadsheet conventions: blue inputs, black formulas, no numbers typed inside formulas.Evidence: CIM p. 14: FY25 revenue $181.4M; CIM p. 17: technicians and utilization; CIM p. 21: cost per technician; CIM p. 38: PP&E $38.0M
Human checkpoint. Associate: Checks the three low-confidence figures against their pages first, then spot-checks the rest. One correction is typed over the extraction and logged with the page reference.
Output. An assumptions tab with a source page next to each input, and an empty model shell ready for the build.
What the reviewer checks. A misread figure fixed here is one correction. Found at committee, it means rerunning the five stages built on it.
2Build
The integrated three statements
A buyout model is three statements and six schedules that have to agree with each other in each year of the hold.
Transcript
Inputs. The approved assumptions tab from Set Up.
AI step. Builds the revenue, cost, working capital, capex, tax, and debt schedules, then links them into the three statements. A linkage map records which schedule feeds which line.Evidence: Linkage map: schedules to statements; Balance check row: assets less liabilities and equity
Human checkpoint. VP: Reviews the linkage map and the balance check row. The check reads zero from close through FY30 before sign-off.
Output. An integrated three-statement model that balances from close through FY30, when total assets reach $358.0M.
What the reviewer checks. Links are reviewed as they are built, so a break shows up on the day it was made.
3Forecast
Driver-based projections
The investment committee will ask why revenue compounds at 7.6% a year. The answer has to sit next to the number.
Transcript
Inputs. Historical drivers for FY23 to FY25 and the management plan in the CIM.
AI step. Proposes a five-year path for each operating driver with a one-line rationale, and flags where the path departs from the historical trend: technician growth of 4.0% a year against a three-year trend of 4.8%, bill rate growth of 3.5% against 4.0%.Evidence: CIM p. 17: technician count by year; CIM p. 52: management utilization target; Model: FY23 to FY25 driver history
Human checkpoint. Deal team: Accepts, adjusts, or overrides each driver. Overrides are logged with who changed what and why. Here the team caps utilization at 76.5%, the level the CIM describes as the practical ceiling.
Output. Projections the investment committee memo can cite: FY30 revenue of $261.9M and EBITDA of $52.4M (20.0% margin), with a rationale for each driver.
What the reviewer checks. When committee asks about a driver, the reason and the name of the person who approved it are already in the table.
4Return
Sources, uses, and returns
Committee wants the handful of assumptions the return depends on, in order.
Transcript
Inputs. The approved forecast and the deal structure: 10.0x LTM EBITDA at entry, a senior term loan at 5.0x.
AI step. Builds sources and uses, the term loan schedule (1% annual amortization plus a 75% cash sweep), and exit values from 8.0x to 12.0x. Calculates IRR two ways, from the cash flows and from MOIC over the hold, and ranks the assumptions that move IRR most: bill rate growth, exit multiple, entry multiple.Evidence: Term sheet: 5.0x senior term loan; Model: FY30 EBITDA $52.4M; IRR check: cash flows and MOIC agree
Human checkpoint. Deal lead: Selects the base case: exit at 10.0x, the same as entry, so no part of the return depends on multiple expansion.
Output. A returns summary: 21.6% IRR and 2.66x MOIC on $172.0M of equity over five years, with a value creation bridge from EBITDA growth, multiple change, debt paydown, and fees.
What the reviewer checks. Moving bill rate growth one point a year up or down swings the IRR across an 11.5-point range; the exit multiple, next on the list, swings it 5.6 points.
5Stress-Test
Integrity, sensitivities, downside
Integrity checks are usually run by hand, by whoever built the model, in whatever order they remember.
Transcript
Inputs. The base case from Return.
AI step. Runs the integrity checks (balance, cash tie-out, sources and uses, sign conventions, circular interest on the average debt balance, hardcodes in formula cells), two sensitivity grids, and a downside case with covenant headroom by year.Evidence: Credit agreement: net debt to EBITDA test by year; Downside: IRR 6.7%, tightest headroom 20% in FY27
Human checkpoint. VP: Signs off on the integrity checklist. An open item holds the model back from committee.
Output. An integrity report with pass or fail for each check, the sensitivity grids, and the downside case.
What the reviewer checks. The same seven checks run on each model, and the report records which passed.
6Monitor
Actuals against the underwriting case
After closing, quarterly packages arrive in different shapes and the underwriting case goes unopened.
Transcript
Inputs. The closing model and six quarters of company reporting in four formats.
AI step. Extracts revenue and EBITDA from each package, maps them to the underwriting case, and tracks variance and covenant headroom each quarter. Flags quarters where EBITDA misses the case by more than 5%: Q3 FY26, Q4 FY26, Q1 FY27.Evidence: Board pack Q3 FY26, p. 6: EBITDA; Compliance certificate Q4 FY26: net debt; Closing model: quarterly case
Human checkpoint. Operating partner: Reviews the flagged quarters only, and asks management how much of the Q4 FY26 miss came from lower utilization.
Output. A monitoring view: actual against the underwriting case by quarter, with covenant headroom.
What the reviewer checks. The operating partner reads three flagged quarters out of six, each measured against the case the deal was bought on.
Underwrite and own
Analyst Agent
CADET (Clarify, Acquire, Derive, Examine, Teach): a chat agent and a coding agent that answer one quantitative question from a written plan. The question comes from the Calvorn deal team, whose forecast assumes technicians get busier: when a maintenance company bills more of its technicians’ hours, how much does its profit margin rise? One of the listed peers in its data, Morlan, reappears in Surveillance.
Clarify
Acquire
Derive
Examine
Teach
1Clarify
The plan is the analysis
An investor asks a one-line question. A useful answer needs the data, the method, and the output agreed before any code is written.
Transcript
Inputs. The investor’s question and the firm’s written guides for this type of analysis.
AI step. The chat agent asks two clarifying questions: which margin, and whether to compare each company with its own past quarters or the companies with one another. It then writes a plan of six tasks. Each task names one table, its columns, and the tasks it depends on.Evidence: Analysis guide: compare each company with its own history; Plan: table columns and dependencies
Human checkpoint. Investor: Answers both: EBITDA margin, and each company against its own history, so differences in size and business mix between companies do not distort the answer. Approves the plan.
Output. An approved plan: six tasks, each with a named output table and its inputs.
What the reviewer checks. The code, the checks, and the report are all built from this plan, so a wrong assumption gets fixed here, before any code runs.
2Acquire
Finding the right data series
A search for utilization returns several series with similar names but different frequencies, currencies, and units.
Transcript
Inputs. The time-series catalog and the plan’s data requirements.
AI step. Searches the catalog and inspects each candidate the way an analyst would: frequency, currency, unit, and whether the values look plausible. Accepts two series and rejects three, with a reason for each.Evidence: Catalog entry per series: frequency, unit, currency; Value check against the plan’s expected range
Human checkpoint. Investor: Confirms the two accepted series and agrees with the three rejections.
Output. Two confirmed series covering six companies over twelve quarters.
What the reviewer checks. Three of five candidates were rejected on frequency, unit, or currency, and each rejection is recorded with its reason.
3Derive
Code written from the plan, in parallel
Code written one step at a time is slow, and two runs of the same request can produce different numbers.
Transcript
Inputs. The approved plan and the two confirmed series.
AI step. The coding agent writes one function per task, all at once, because the plan already fixes each table’s inputs and columns. The tasks run in four waves, each task starting as soon as the tables it needs exist, under a hard cost cap for the session. A second, independent generation of the same plan produces identical outputs.Evidence: Task order: four waves; Output check: run 1 and run 2 match to twelve decimal places; Session cost: $6.80 of a $25.00 cap
Human checkpoint. Investor: Works from the tables and the chart. The code is stored with the report for audit.
Output. Tables for each task and a line fitted through 72 company-quarters, each company measured against its own average.
What the reviewer checks. Two separate code generations gave the same slope to twelve decimal places, so the figure in the report does not depend on which run produced it.
4Examine
Checks before the report
Code that runs can still give an answer that makes no sense.
Transcript
Inputs. Each task’s output table and its specification from the plan.
AI step. A fixed set of code checks compares each table with its specification: row count, columns, missing values, and value range. The agent then reviews the result. Measured within each company, one point of utilization is worth 0.62 points of margin. One line drawn across all six companies together gives 0.92, because the busier companies also earn higher margins for other reasons, such as size and contract mix. Last, it refits on FY23 and FY24 only and tests the fit on FY25, which it has not seen.Evidence: Specification checks: seven of seven passed; FY25 test: the fit explains 95% of the variation it had not seen; Within-company and all-company figures compared
Human checkpoint. Investor: Reads the report and the caveats. Saving the fitted series to the shared database takes a Publish on the final table; the agent cannot write there on its own.
Output. A report: when a company bills one point more of its technicians’ hours (say 74% to 75%), its EBITDA margin is about 0.62 points higher. For the Calvorn plan, that puts the utilization assumption in margin terms. Chart and data attached.
What the reviewer checks. The all-company figure of 0.92 would have overstated the effect by 49%. The self-review caught it before the investor saw the report.
5Teach
One correction, shared with all users
If an investor corrects the agent in one conversation, the next investor should not have to make the same correction.
Transcript
Inputs. The investor’s steer and the agent’s benchmark suite.
AI step. Writes a benchmark that fails on the current agent, updates the analysis guide, and reruns the suite: the new case passes and the other 60 still do. The change is proposed for approval.Evidence: New benchmark: fails before, passes after; Full suite: 61 of 61 passing
Human checkpoint. Applied AI team: Reviews the proposed change to the analysis guide and merges it.
Output. An updated analysis guide and a suite of 61 benchmarks, all passing.
What the reviewer checks. The next investor who asks a question of this kind sees each company’s figure without asking for it.
Protect the portfolio
Scenarios
One fictional fund of six holdings, Calvorn among them, put through a scenario a risk lead wrote in one sentence: rates up, oil down, inflation up, growth down. The question is which companies fail a loan test, and whether the fund has the cash to help.
Frame
Map
Run
Roll Up
Respond
1Frame
A sentence becomes a scenario
A risk lead can describe a scenario in one sentence. Turning it into numbers that every holding’s model can use usually takes a week of meetings.
Transcript
Inputs. The risk lead’s scenario in plain English, the factor history, and the list of holdings.
AI step. Turns the sentence into four shocks: policy rate up 1.5 points, oil down 25%, inflation up 2 points, and GDP growth down 1.5 points. Checks each against the largest one-year move in the factor history, then checks the combination. Inflation rising while oil falls is unusual, because oil usually feeds inflation, so it asks the risk lead what drives it.Evidence: Factor history: largest one-year moves; Consistency rule: oil and inflation usually move together; Plausibility checks: four of five pass; one waits for a driver
Human checkpoint. Risk committee: The risk lead answers: wages. The committee records wage-led inflation as the driver, so labor costs carry the inflation shock, and approves the scenario before anything runs.
Output. A scenario card: four shocks, the stated driver, the checks each passed, and who approved it.
What the reviewer checks. Every holding is tested against the same approved numbers, so differences in the results come from the companies, not from how each team read the sentence.
2Map
Shocks to each company’s drivers
The same shock hits a trucking company and a software company differently. Each holding needs its own sensitivities.
Transcript
Inputs. The approved scenario, each holding’s latest board pack, and its credit agreement.
AI step. Drafts each holding’s sensitivities from its board pack: how revenue moves with growth and with oil, how much of its cost is fuel and wages, how much of a cost change it passes to customers, and how much of its debt floats. Each figure links to the page it came from.Evidence: Board pack Q2, p. 7: Pellmoor cost breakdown; Credit agreements: floating share and hedges; Sensitivity table, six holdings
Human checkpoint. Deal teams: Each deal team confirms its own line. Pellmoor’s team changes fuel pass-through from 20% to 85%: its customer contracts carry fuel surcharges, so most of any fuel saving goes back to customers.
Output. A sensitivity table for six holdings, each line confirmed by its deal team.
What the reviewer checks. With the drafted figure, Pellmoor gained $16.7M of EBITDA from cheaper fuel. With the contract figure it loses $2.7M. The deal team knew the contracts; the board pack did not say.
3Run
Every holding, the same shock
A covenant that fails in a scenario is the first thing a lender will ask about, and each credit agreement has its own test.
Transcript
Inputs. The approved scenario and sensitivities, and each holding’s covenant test.
AI step. Runs the scenario through each holding: EBITDA after the shock, a year of extra interest on unhedged floating debt, net debt to EBITDA against the covenant, and equity value. Then runs 2,000 variations, each shock scaled up or down at random, to see how often each covenant fails.Evidence: Central case: one covenant fails; Variations: 2,000 runs, same random seed every time; Covenant tests, six credit agreements
Human checkpoint. Deal teams: Kilnworth Building Products fails its test in the central case, at 6.18x against a maximum of 6.00x; its deal team gets the result with the covenant page. The teams at Stavely, Selbrook, and Culverin get theirs too: each passes, with less than 10% room.
Output. A results table: EBITDA, interest, the covenant test, and equity value for each holding, with the share of variations in which each covenant fails.
What the reviewer checks. Kilnworth fails in the central case and in 69% of the variations. Stavely passes in the central case but fails in 26% of them, which a single run would not show.
4Roll Up
The fund view
The committee asks two questions: how much value is at risk, and whether the fund has the cash to support its companies.
Transcript
Inputs. The holding results and the fund’s own capital position.
AI step. Adds up the holdings. Equity value falls 32% in the central case, and more than 42% in one variation in twenty. Extra interest across the fund is $16.3M a year. The fund may need $10.3M to cure Kilnworth’s covenant, plus the extra interest if the companies cannot carry it, against $220M available after reserves.Evidence: Holding results, six companies; Fund capital: uncalled commitments and credit line; Reserves for follow-on investments
Human checkpoint. CFO: Checks the capital figures against the latest capital account statement and signs off on the fund view.
Output. A one-page fund view: value at risk, covenant fails, cash the fund may need, and cash available.
What the reviewer checks. Available cash covers the possible need of $26.7M about eight times, so the discussion moves to what to do about Kilnworth, not whether the fund can cope.
5Respond
Options, not orders
A covenant that fails in a scenario is a question for the investment committee, with several answers at different prices.
Transcript
Inputs. The results for Kilnworth, its credit agreement, and the fund view.
AI step. Lists three responses for Kilnworth and tests each against the covenant: an equity cure of $10.3M; an early covenant amendment at about $3.5M in fees and a year of higher margin; and swapping its unhedged debt to fixed for $1.4M. The swap alone does not bring it back under the test.Evidence: Credit agreement: equity cure and amendment terms; Each option re-tested against the covenant; Scenario card and sensitivity table
Human checkpoint. Investment committee: Opens amendment talks with the lenders now, while the company still passes its test, and holds the cure as the fallback. The workflow places no trade and moves no money.
Output. A decision record: the option chosen, its cost, the fallback, and the date the scenario will be scored against what happened.
What the reviewer checks. Asking lenders for room before a breach usually costs less than asking after one, and the record shows why the committee chose to ask early.
Protect the portfolio
Early Warning
The same six holdings as Scenarios, watched between quarterly packages. Most are private, so the signals come from outside: hiring, shipments, customers, and news about the companies they buy from and sell to.
Source
Resolve
Baseline
Test
Alert
1Source
Approve the data first
Outside data can carry information a firm is not allowed to trade on, and examiners ask how each data set was checked.
Transcript
Inputs. Six data sets proposed by portfolio operations, each with the vendor’s answers to a due diligence questionnaire.
AI step. Reads each questionnaire and drafts a recommendation with fixed rules: a source that may carry inside information is rejected, collection that breaks a website’s terms is rejected, an undocumented method is held, and personal data is approved only in aggregate.Evidence: Vendor questionnaires, six data sets; Firm policy: expert calls and alternative data; Rules: approve, approve with conditions, hold, reject
Human checkpoint. Compliance: Reviews each draft and decides: approves three, rejects two, and holds the web traffic panel until the vendor documents how it collects the data.
Output. A source register: each data set, the decision, the conditions, and who signed.
What the reviewer checks. The channel-check notes were rejected before anyone used them: one caller works at a listed supplier, which could put inside information in the file.
2Resolve
The right company, once
One story can run in forty outlets, and a company with a similar name can land in the wrong file.
Transcript
Inputs. A week of news, releases, and call transcripts that mention the holdings, or the companies they buy from and sell to.
AI step. Groups the 74 items into seven stories, matches each to a holding, and rejects two about companies with similar names. Two stories reach a holding through a customer: a homebuilder cutting its plans matters to Kilnworth, and a drilling budget cut matters to Stavely.Evidence: Legal-name check: two lookalikes rejected; Customer lists: two indirect links; Duplicates: 74 items to seven stories
Human checkpoint. Analyst: Checks the two rejected lookalikes and the two customer links against the legal names and customer lists, and confirms each.
Output. A list of five stories by holding, each with its sources.
What the reviewer checks. The Kilnworth hiring pause counts once, not 38 times, so a widely copied story does not drown out a quiet one.
3Baseline
Against its own normal
A company that always posts few jobs is not a warning. A company that suddenly posts fewer than usual might be.
Transcript
Inputs. Weekly counts for each approved signal: twelve weeks of history and the latest four.
AI step. Compares each signal’s last four weeks with its own twelve-week history, measured in standard deviations. Flags three of eight signals at two or more: Kilnworth job postings (−5.8) and import shipments (−3.3), and Selbrook patient complaints (+3.1).Evidence: Twelve weeks of history per signal; Flag rule: two standard deviations from its own average; Three of eight signals flagged
Human checkpoint. Portfolio operations: Reviews the three flags before any reach a deal team, and asks which of these signals have warned in the past.
Output. Three flagged signals, each shown against its own history.
What the reviewer checks. Pellmoor posts more jobs than any other holding and is not flagged. Kilnworth posts fewer than it usually does, and is.
4Test
Does it warn early?
A signal that fires often and is usually wrong teaches deal teams to ignore alerts.
Transcript
Inputs. Records from past holdings: when each signal fired, and whether an EBITDA miss of more than 5% followed within two quarters.
AI step. Scores each signal type on that history: how often a firing was followed by a miss, how many misses it caught, and how many days of warning it gave. Hiring slowdowns were right 75% of the time, with about 57 days of warning. Shipment drops: 67%, 62 days. Complaint spikes: 29%, 20 days.Evidence: Backtest: 48 past company-periods; Promotion rule: right at least half the time, catches 40% of misses, 30 days of warning; False alarms counted per signal
Human checkpoint. Head of portfolio operations: Promotes hiring slowdowns and shipment drops to alerts. Holds complaint spikes back: wrong more often than right, so Selbrook’s flag is logged, not sent.
Output. Two promoted signals with their track records; one held back, to be retested next quarter.
What the reviewer checks. Only signals with a track record reach a deal team, and the record travels with every alert.
5Alert
To the owner, with evidence
An alert helps only if it reaches the person who owns the position, with enough evidence to act on.
Transcript
Inputs. The promoted flags, the story list, and last quarter’s alert outcomes.
AI step. Combines the two Kilnworth flags and the customer story into one alert for its deal team, with the signal history and sources. From last quarter’s outcomes, finds that two of three dismissed alerts were December hiring pauses, and proposes a seasonal adjustment.Evidence: Kilnworth: hiring −5.8, shipments −3.3; Customer story: homebuilder cuts guidance on new starts; Last quarter: three dismissed, two seasonal
Human checkpoint. Kilnworth deal team: Escalates: asks the CEO for the order book ahead of the quarterly package. The seasonal adjustment goes to portfolio operations for review before it takes effect.
Output. An alert log: each alert, the owner’s decision and reason, and the threshold change awaiting review.
What the reviewer checks. Kilnworth is also the holding that failed its covenant in the rate scenario. On the record so far, a hiring slowdown has come about 57 days before the numbers show it.
Protect the portfolio
Surveillance
A private equity firm with a credit fund. Deal teams receive confidential information and staff trade in personal accounts. The workflow keeps the restricted and watch lists current, screens every trade, and leaves a record an examiner can follow. People appear by staff number only. The industrials team’s bid target, Morlan, is one of the listed peers in the Analyst Agent.
List
Screen
Connect
Triage
Prove
1List
Keep the lists current
Deal teams receive confidential information every week, and each time the restricted or watch list has to change the same day.
Transcript
Inputs. This month’s deal pipeline events and the current lists.
AI step. Drafts a list change for each event, with its reason and date: an NDA puts a company on the watch list, and a bid or a board seat puts it on the restricted list. For a bid on a private company, it finds the listed parent and puts the parent on the watch list.Evidence: Pipeline: four events; Company hierarchy: Pryor Medical Devices is owned by Pryor Holdings; List rules: event type to list
Human checkpoint. Compliance: Approves the four changes and adds one the pipeline missed: the credit fund joined a revolving loan to Brannock Chemicals, which came in by email. Brannock goes on the restricted list.
Output. Updated lists: three restricted names and two on the watch list, each tied to an event, a date, and an approver.
What the reviewer checks. Each entry records why it was added, so it can come off the list when the reason ends. Examiners check both.
2Screen
Every trade, against the lists and the calendar
A trade can break no rule on the day it is made and still look wrong a week later, when the company announces a deal.
Transcript
Inputs. A month of personal trades and pre-clearance requests, the lists, and public announcements.
AI step. The rules engine (not the model) refuses pre-clearance for restricted names. The screen then flags trades in watch-list names and trades made within 15 days before an announcement in the same company: two blocked, four alerts, six clear.Evidence: Pre-clearance: two requests refused; Timing window: 15 days before an announcement; Watch list: Tolland Aggregates and Pryor Holdings
Human checkpoint. Compliance analyst: Confirms both refusals came before any order was placed, and opens the four alerts.
Output. A screening log: every trade, the rule it met, and the result.
What the reviewer checks. A request to buy Morlan Facility Works from a member of its own deal team never became a trade: pre-clearance refused it automatically.
3Connect
Follow the relationships
An account that trades before an announcement is not suspicious on its own. What matters is whether it connects to someone who knew.
Transcript
Inputs. The alerts, and a map of people, household accounts, deal teams, and meetings.
AI step. For each alert, finds the shortest chain from the trading account to the deal behind the security. Two alerts connect: a household account to a deal-team member (three steps), and a credit analyst through a shared meeting (four steps). The other two connect to no deal.Evidence: Chain: H-114a → E-114 → Industrials deal team → Tolland NDA (P-1); Chain: E-207 → Meeting M-3 → E-114 → Industrials deal team → Tolland NDA (P-1); Rosters and meeting records
Human checkpoint. Investigator: Checks each link against its record: the household account statement, the deal-team roster, and the meeting invitation.
Output. Each alert shows its chain of evidence, or states that none was found.
What the reviewer checks. A risk score cannot be checked. A chain of records can, one record at a time.
4Triage
Context, then a person
Rule-based alerts are mostly false alarms, and investigators spend their day clearing them.
Transcript
Inputs. The alerts with their chains and trade details.
AI step. Ranks the queue: two high (a chain to the deal and a trade shortly before the announcement), one medium, and one low. Writes a summary of each that cites the trade, the chain, and the dates, and proposes a closing reason for the lower two.Evidence: Priority rule: a chain plus timing is high; Summaries cite trade, chain, and dates; Proposed reasons for the lower two
Human checkpoint. Investigator: Escalates the two high alerts to the chief compliance officer and closes the other two, rewriting one proposed reason to add the trade and bid dates. The model closed none.
Output. A decision log: two escalated and two closed, each with a written reason and the investigator who decided.
What the reviewer checks. The investigator’s time went to the two alerts that needed judgment, and every closed alert carries a reason a person wrote or approved.
5Prove
Ready for the examiner
An examiner asks for the process, not the number of alerts: what was covered, what was missed, and who decided.
Transcript
Inputs. The lists, the screening and decision logs, and the broker feeds.
AI step. Runs the coverage checks: every list entry has an event and a security identifier, every access person’s accounts are in the feed, every alert has a reason, and the change to the timing window was validated before use. five of six pass. One fails: a new household account for E-322 is missing from the feed.Evidence: Coverage checks: five of six pass; Broker feeds: accounts per person; Change log: timing window from 10 to 15 days
Human checkpoint. Chief compliance officer: Signs the monthly report and assigns the missing account to an owner, with a date to close it.
Output. A monthly report: coverage checks, list changes, alerts, decisions, and the open item with its owner.
What the reviewer checks. The firm’s own check found the gap and gave it an owner and a date, before an examiner could find it.
How these are built
Synthetic data
Every company, ticker, document, and figure is fictional. The Calvorn deal is one data file: three historical years, the deal terms, and the exit assumptions. The three statements, sources and uses, debt schedule, IRR, and MOIC are calculated from it in code, and tests confirm that the balance sheet balances in each period, sources equal uses, and cash ties across statements. The deal probabilities, the analyst agent’s regression, the fund scenario, the signal backtests, and the trade screen are computed the same way, and tested on every change.
Named reviewers
Each workflow names the role that reviews the AI step and what that person approves, adjusts, or overrides. Cobalt marks content the model produced; ink marks content a reviewer approved. Overrides are logged with who changed what.
Sources
Extracted figures carry a citation back to their source: a page in the CIM, a segment note in a filing, a passage in a transcript. Index weights come from fixed rules that run without the model.
Illustrative demonstrations using synthetic data. All companies, tickers, and people are fictional; names and tickers were checked against companies listed on US exchanges. Not investment advice. Not a representation of any employer’s or client’s systems.