Report · PCT-066 · Sourced card
AI exposure in layers
A credit team wants its AI exposure measured in layers (compute, data-centre infrastructure, power and fibre enablers, and borrowers AI could displace) before its investors ask for a limit.
Today vs automated
- Export the bookAnalyst · manual
- Hand-tag AI namesAnalyst · manual
- Find shared tenantsHead of research / IR · manual
- Email / meeting hand-offAnalyst → Head of research / IR
- Quarter-end, or a new filing is publishedTrigger: starts on its own
- Tag four layersAutomated · Data store
- Find shared counterpartiesAutomated · LLM
- Measure % of NAVAI draft, human review · Reporting template
- Where to set a limit, if anyApprove / override · Head of research / IR
AutomatedAI draft, human reviewHuman onlyTools: Filing feeds, Data store, LLM, Reporting template
Estimated valueestimate
20–180 analyst hours a year
$3k–$45k of time, vs an assumed $20k–$100k tool cost
Net: −$97k to $25k a year net of tool cost; at the low end of the assumptions it does not cover the tool cost.
Assumptions
- 8–16 analyst hours per review today; 1–3 hours with an agent, including review
- 4–12 reviews a year for one team
- $150–250 per hour loaded cost
- Tool or build cost $20,000–100,000 a year (both)
Estimate, based on the assumptions shown; your numbers will differ. How we estimate
What else it can do for the team
- Find shared tenants and suppliers across loans
- Measure exposure as % of NAV by layer
- Draft the risk committee note
- Set and track concentration limits
- Re-run after each new deal
How you get thereFour steps from prompting to an agentic workflow; most teams start at step 2.
Needs: an approved AI tool, a confidentiality policy.
Risk: unchecked answers, data pasted outside policy.
Needs: field definitions, a template, document parsing.
Risk: breaks silently when formats change.
Typical starting point
Needs: inbox/doc-store access, system write-back, logging.
Risk: acting on a misread; needs review gates.
The automated lane above
Needs: shared data model, permissions, audit trail, exception queue.
Risk: errors compound across steps; needs monitoring.
Solution optionsBuy (6 vendors listed), build or hybrid, with the trade-offs of each.
AlphaSense, Octus, The LoanEdge, Bloomberg Terminal, Moody's, Standd
Fastest start; you adapt to the product's data model and pricing.
An LLM API, a document parser, a vector or document store, your data warehouse, and a review queue.
Fits your definitions exactly; you own maintenance and controls.
Vendor extraction feeding your own checks, templates and systems.
Common middle path; two things to integrate and govern.
Listing is not endorsement.
What changes for the team8–16 h → 1–3 h per review (estimate); coverage goes from the names someone had time to pull to the full set, refreshed when sources change.
How it works
Full title: AI exposure isn't one number: tag direct compute, infrastructure, enablers and AI-substitution risk separately
“Carlyle: private credit may fund ~$1tn of AI compute. Know which layer of your book that money touches.”
Scenario
Public: a credit team wants its AI exposure measured in layers (compute, data-centre infrastructure, power and fibre enablers, and borrowers AI could displace) before its investors ask for a limit.
What the human still decides
Where to set a limit, if any. Carlyle reports some investors asking for about 8–10% on syndicated loans; that is one ask, not a benchmark.
Why it matters
Carlyle's 1 Oct 2026 paper says private credit may finance roughly $1 trillion of AI compute and that failing to set concentration limits 'may prove to be the biggest mistake of all'.
What the record shows
Checked 2026-10-16.
Carlyle estimates private credit could finance roughly $1 trillion of AI compute, more than half the industry's AUM, and Carlyle's credit head says some investors want 8–10% limits on syndicated loans. For a lender's own book the question is the same at every layer: how much exposure sits in compute and power (layers 1–3) versus software that AI may displace (layer 4), and which counterparties are shared across positions. Public lender books are one outside reference for the software layer; see the BDC section.
Try it: prompt and data
Tag each position in {portfolio} as L1 direct compute, L2 data-centre infrastructure, L3 power/fibre enabler, L4 AI-substitution risk, or none. Quote the sentence that justifies each tag; no quote, no tag. Report exposure by layer, drawn and undrawn.Data
Public inputs only (SEC/EDGAR filings, published reports, public news). Every figure is cited.
Who uses it, vendors and build
- Segments
- Direct lending
- Who uses it
- Portfolio manager, Risk
- Build
- Vendor + in-house
- In-house
- Classification over credit documents plus the SOI. Estimate: 1–2 weeks.
- Last verified
- 2026-10-15
Vendors that claim this capability
- Standd: credit-stress-tracking
- The LoanEdge: credit-stress-tracking
- AlphaSense: credit-research
- Bloomberg Terminal: credit-research
- Moody's: credit-research
- 1 more vendors claim related capabilities. See all vendors for this stage →
Tags reflect each vendor's own public claims at the linked page (checked 2026-10-15). Not tested, ranked or endorsed.
Sources and related
Next step
Related use cases
Get new walk-throughs
Not investment advice. Names marked SAMPLE are fictional. Found an error? See the corrections log.