$4.1B
realized value on the AI-driven platform portfolio
AI that shows up in the P&L.
The board is asking about AI. The LPs are asking. And your operations still run on manual work. We close that gap: agents that read, one record underneath, a person at the gate, and the senior owner to drive it, without the $500K hire.
■ The problem we solve
The AI conversation is loud. The production systems are rare.
Six sample integrations, one per desk, drawn the way we build them for private capital and insurance. Pick a desk. Left to right: what comes in, the agents and checks, the person at the gate, and the line it moves in the P&L. Patterns we build to, not case studies.
Private capital
Insurance
Business problem
Packages land as PDFs, analysts re-key them, and the breach that matters is found at quarter-end. Firm-wide exposure to one borrower is a two-day fire drill.
What comes in
- Borrower package, 84 pp
- Compliance certificate
- Amendment No. 3
- Administrator NAV file
Agent pattern: Orchestrator and workers, then deterministic checks
1
What leaves
- Covenant exceptions, ranked, in-month
- Exposure to any borrower, all vehicles, in seconds
- Month-end that opens with the exceptions found
Where it shows up in the P&LAnalyst hours off re-keying (a quarter of the year per $150–250K analyst). Breaches in-month, not at quarter-end. The LP exposure question in seconds.
Human in the loop
The portfolio manager decides every waiver. No covenant status changes without that sign-off.
Data foundations underneath
entity model: firm → fund → vehicle → borrower → facilitycovenant definitions per facilitydocument store with page referencesadministrator ↔ ledger mapping
Governance
evals before productionaction log, every steprecorded human sign-offexaminer-readable memo
Business problem
Borrowing bases are sampled, not verified line by line. Fabricated invoices and double-pledged receivables live in the gap. First Brands was the reminder.
What comes in
- Borrowing base certificate
- AR aging, 380 debtors
- Debtor concentration report
- Field exam notes
Agent pattern: Parallel extraction, deterministic scoring, LLM as judge on the flags
1
What leaves
- Advance approved, exceptions listed line by line
- Duplicate or cross-certificate pledges escalated with evidence
- Certificate reconciled to prior month, every line
Where it shows up in the P&LOne caught double-pledge pays for the build several times over. Line-level verification holds the advance rate the market just cut to 80–85%.
Human in the loop
The credit officer approves the advance and every escalation. The judge recommends; a person releases money.
Data foundations underneath
debtor master with resolved namesinvoice ledger with certificate historyineligibility and aging rules, encodedfield exam findings, structured
Governance
evals before productionaction log, every steprecorded human sign-offexaminer-readable memo
Business problem
Seven-figure commitments priced on a partner’s read of a banker’s box. Duration is the silent IRR killer, and nobody has measured their own calibration.
What comes in
- Case file, 1,900 pp
- Medical records, 620 pp
- Docket, 142 entries
- Damages report
Agent pattern: Evaluator and optimizer: draft, critique, revise, beside a calibrated model
1
What leaves
- IC memo with a source on every claim
- Duration with an 80% band, never a point prediction
- Calibration study: assumed vs actual, book-wide
Where it shows up in the P&LDuration is the IRR. A 0.4-year miss on a 2.6-year asset is real basis points across a book. Intake to IC in days.
Human in the loop
The partner signs the IC memo. The system produces the memo and the band; the investment decision stays human.
Data foundations underneath
case master: counsel, jurisdiction, comparablesmilestone history per caseconcluded-case table: assumed vs actualdocument store with page references
Governance
evals before productionaction log, every steprecorded human sign-offexaminer-readable memo
Business problem
Cycle time is set by how fast one underwriter can read physician statements. Every extra day is an applicant who went somewhere else, on acquisition cost already spent.
What comes in
- Attending physician statement, 412 pp
- Application, 48 pp
- Labs and pharmacy history
- MVR
Agent pattern: Router first, then specialists in parallel, then a reconciler
1
What leaves
- Instant decisions on clean cases, to the face the rules allow
- Findings for the underwriter, organized and sourced
- Documentation the committee and the state bulletin can read
Where it shows up in the P&LPlacement rate is premium you already paid to acquire. Every point of instant-decision rate and every day of cycle time is conversion.
Human in the loop
The underwriter decides every routed finding. Instant offers issue only where the rules, not a model, say the case is clean.
Data foundations underneath
applicant → policy → evidence → treaty modeleligibility rules, versionedevidence store with page referencesexperience study tables by cohort
Governance
evals before productionaction log, every steprecorded human sign-offexaminer-readable memo
Business problem
Hundreds of adjusters, each holding hundreds of files of 500 to 50,000 pages. The adjuster reads the top of the pile. The reserve is set on assumptions nobody has checked against the claims data.
What comes in
- Claim file, LTD, 2,300 pp
- IME report
- Employer statement
- Treating notes, prior claims
Agent pattern: Grounded retrieval and summary, with a judge for contradictions
1
What leaves
- The adjuster starts at the decision, not page one
- Instant case answers, with the page
- A reserve assumption checked against your own claims
Where it shows up in the P&LAdjuster hours first. Then the real number: a duration assumption a few percent light on a cohort is a reserve figure.
Human in the loop
The adjuster decides the claim. The appointed actuary owns the reserve review.
Data foundations underneath
claimant → claim → policy → benefit → reserve modelfile store with page referencescohort tables for the reserve reviewIME and treating-note structure
Governance
evals before productionaction log, every steprecorded human sign-offexaminer-readable memo
Business problem
Bordereaux arrive as Excel from one cedant and text files from another; one policy sits across three quota-share sessions and is reconciled by hand. On a PRT bid, a slow cleanse is a padded quote.
What comes in
- Bordereau, cedant B, 9,140 rows
- Movement file
- Scheme data file, 740 participants
- Treaty terms, 50/20/10
Agent pattern: Chained validators, then a long-running workflow agent
1
What leaves
- Movement file approved, with the two items a person actually looked at
- Priceable scheme data on the bid clock, exceptions listed
- Regulatory returns assembled from governed data
Where it shows up in the P&LReconciliation headcount first. Then unpadded quotes on the deal clock, which is share on a $48.8B market. Then returns without the fire drill.
Human in the loop
The analyst approves the movement file; the actuary prices with the exceptions listed. Nothing outside tolerance passes without a person.
Data foundations underneath
cedant → treaty → session → policy → life modelformat mappings per cedanttolerance rules, versionedmovement history
Governance
evals before productionaction log, every steprecorded human sign-offexaminer-readable memo
agentdeterministic stepjudge / evaluatorperson at the gate
■ How we solve it
Everything above is built from four parts.
Agents that read, one record underneath, a person at the gate, and a senior owner who runs it. Take one, or take the sequence.
01Document-to-Decisionthe agents that read
02Data Foundationsthe one record underneath
03Fractional AI Leadershipthe senior owner of all of it
04Your First AI Hirethe team that runs it after us
Not sure where to start? Book the briefing.Ninety minutes with your leadership team. We work only in private capital and insurance, so we can tell you what firms like yours actually run in production, what’s vendor theater, and what your competitors are doing. $2,500, credited toward a diagnostic if you go further. The entry offers →
Track record / prior executive roles
$754M
in reserve accuracy improvements at a global insurer
99.9%
digitization accuracy across 50M+ pages of regulated records
34
data & AI systems shipped to production over six years
■ Why “Operating Alpha”
The name is the thesis.
Alpha comes from operations too: the speed a submission is quoted, the days a covenant check takes, the accuracy a reserve is held to. AI is the biggest operational lever in a generation for the firms that get it into production. Most do not. That is the business we are in.
Financial services experience
Richardson
BMO
Manulife
John Hancock
Preston Capital
Encore Life Settlements
Altriarch
Richardson
BMO
Manulife
John Hancock
Preston Capital
Encore Life Settlements
Altriarch
■ Who you’re hiring
Twenty years inside regulated finance, building systems that made it to production.
Operating Alpha is founded and personally led by Khurram Tehseen, who has started five data and AI departments from scratch. Zero-to-one builds are the specialty. The three most recent seats:
- Chief Data Officer, Altriarch Asset Management
- Managing Director, Data & AI, Preston Capital
- Director of Advanced Analytics, Manulife Financial
The platforms those teams built carried $4.1B in realized value, with 34 systems shipped to production.