How to Get AI to Use Your Firm's Definitions

To get AI to use your firm’s definitions, store each one once, outside the prompt, as a governed formula with an owner, a version and an effective date, and have code compute every figure from it. That’s how we built Kepler for investment firms: your firm’s terms, from adjusted EBITDA with your add-backs to net IRR with or without the subscription line, carry one definition in your firm’s ontology, applied the same way in every answer, with the history of every change. A definition typed into a prompt or a custom instruction gets applied some of the time and silently dropped the rest.

Why the same word means different things

Investment vocabulary looks standardized, and most of the important terms have several defensible definitions. An AI that doesn’t know which one your firm uses will pick one, usually the most common in its training data, and won’t say so.

Term	Where the definitions split	What goes wrong when AI picks one
Adjusted EBITDA	Which add-backs count: stock compensation, restructuring, acquisition costs, projected savings	Comps mix each filer’s version with yours, and margins stop being comparable
Net IRR	With or without the subscription line; fee and carry treatment	Two funds look further apart, or closer, than they are
Exposure	Gross or net; look-through or direct; by issuer or ultimate parent; delta-adjusted	A risk figure that understates concentration
ARR	Contracted or live; usage included or not; currency basis	Growth rates that compare different things
Leverage ratio	The credit agreement’s Consolidated EBITDA and its caps; net of cash or gross	Covenant headroom computed on the wrong denominator
Fiscal quarter	Fiscal or calendar; 52- or 53-week years	A retailer’s Q3 covers a different window from yours
Peer set	“Our semis comps,” “core holdings”	The model assembles a peer set of its own

Three of them need a closer look.

Adjusted EBITDA

Companies define their own adjusted EBITDA, and the rules only constrain how they present it. Regulation G and Item 10(e) of Regulation S-K require a reconciliation to the most directly comparable GAAP measure, and the SEC staff’s non-GAAP guidance (Question 100.01) warns that a performance measure excluding normal, recurring cash operating expenses can be misleading. That reconciliation table in the earnings release is the raw material: each add-back listed with its amount. A firm’s house definition decides which of those add-backs it accepts. One filer adds back stock compensation and restructuring; another also adds back acquisition-related costs; your house view may reject restructuring that has recurred three years running. A comp sheet labeled “EV/Adjusted EBITDA” means nothing until it says whose adjustments it used. Our guide to building trading comps with AI walks through a comp sheet on both bases.

Net IRR and the subscription line

A subscription line, or capital call facility, lets a fund pay for investments by borrowing and call LP capital later. LP money is at work for less time, so IRR usually rises, while the multiple on invested capital dips slightly by the interest cost. ILPA’s Performance Template, released in January 2025, reports performance both with and without the impact of fund-level subscription facilities, and ILPA recommends it for funds that start operating on or after January 1, 2026. An allocator comparing managers has to put every fund on one basis, and many existing funds’ quarterly reports predate the template. Our guide on how LPs use AI on fund documents covers the reading side of that work.

The leverage ratio on the credit agreement’s terms

A credit agreement defines Consolidated EBITDA in its own words: a list of permitted add-backs, often a cap on projected cost savings and synergies, and rules for acquisitions made during the test period. The covenant’s leverage ratio is computed on that figure, sometimes net of cash up to a limit. The borrower’s reported adjusted EBITDA can differ from it by a wide margin, in either direction. Covenant headroom has to be computed on each agreement’s own definition, one agreement at a time, which is why reading the credit agreement comes before any calculation.

Why prompts and custom instructions don’t hold

The first thing most teams try is telling the AI: a paragraph in the prompt, a custom instruction, a document uploaded to a project. It works in a demo and drifts in daily use, for six reasons.

1. Instructions compete with everything else in context. A definition at the top of a conversation is one paragraph among many. As filings and tool results fill the context, instructions lose force; we describe watching that drift in our post The Context Wall.
2. The model still does the arithmetic. Even when it picks the right add-backs, it sums them in the same pass that writes the sentence, with the same error rate as any other generated number.
3. The inputs aren’t resolved. “Add back restructuring” needs the right restructuring figure for the right period, and that figure can appear on the income statement, in a footnote and in MD&A with three different scopes.
4. There’s no version. Change the definition in March and nobody can reproduce last quarter’s memo under the old one.
5. Every analyst has their own prompt. Five people paste five slightly different definitions, and the firm’s numbers drift apart without anyone deciding they should.
6. Prose is ambiguous. “Exclude one-time items” leaves open which items count and who decides.

Custom instructions and project files are good at style: the memo format, the tone, what to lead with. Definitions that produce numbers need a home the model can’t overrule.

The semantic layer, extended to documents

Business intelligence teams solved a version of this years ago with the semantic layer. The dbt Semantic Layer, for example, lets data teams define a metric once in the modeling layer, and in dbt’s words, “If a metric definition changes in dbt, it’s refreshed everywhere it’s invoked.” Every dashboard computes revenue the same way because no dashboard computes it on its own.

That works for metrics over warehouse tables. At an investment firm, many definitions reference facts that live in documents: the add-backs in an earnings release reconciliation, Consolidated EBITDA in a credit agreement, subscription-line draws in a capital account statement. The semantic layer has to reach those documents, which means resolving entities, extracting facts with their sources and periods, and writing definitions as code over those facts. That combination is an ontology, the core of Kepler, covered in our guide to what an ontology is for investment firms.

How to write a definition AI can apply

A definition an AI system can apply reliably has the same fields as a well-run data contract. This is the shape definitions take inside Kepler:

name:      Adjusted EBITDA (house)
aliases:   Adj. EBITDA, house EBITDA
owner:     Head of research
version:   3, effective 2026-01-01
formula:   operating income + D&A + stock-based compensation
           + acquisition-related costs
           + restructuring, unless recurring in each of the 3 prior years
inputs:    each filer's reported line items, as filed
scope:     public comps and IC memos; never covenant tests
tests:     FY2025 values for three companies, checked by hand

The formula names concepts, never numbers, so it applies to any company and period. The scope line stops a house definition from leaking into places that need another one, such as a covenant test. The tests give anyone who changes the definition a way to see what moved.

Governing definitions

Once definitions live outside the prompt, they need the same governance as any other model input:

• An owner for each definition, and a named approver for changes.
• Versions with effective dates, so an old answer reproduces under the definition it used.
• Scope: which desks, funds and deals use which definition. A covenant definition belongs to one agreement; a house definition belongs to the research team.
• Side by side when they differ: the company’s reported figure, the lender’s covenant figure and the house figure, each labeled.
• Corrections captured as rules, so a fix made once applies everywhere the same case appears.
• Access and audit: who can change a definition, with every change logged.

Kepler applies all six to every definition in a firm’s ontology.

Four tests for whether a tool uses your definitions

1. Ask for a metric by your house name, and check the formula the tool shows.
2. Change one add-back, and check that every output using it updates.
3. Ask for last quarter’s figure under the previous version.
4. Have two analysts ask the same question in their own words, and compare the numbers.

On Kepler, every figure opens onto its formula and inputs, so the first test takes one click.

These sit alongside the checks in our guide to checking AI-generated numbers, and they belong in any AI vendor evaluation.

How Kepler does this

Your definitions live in your firm’s ontology. Our engineers build a living model of your firm from your own documents and systems in the first week, and your house terms go into it as definitions: adjusted EBITDA with your add-backs, net IRR with or without the subscription line, look-through exposure by vehicle, each portfolio company’s leverage ratio on its own credit agreement. Each one carries an owner, a version and an effective date, and every change is kept, so last quarter’s memo reproduces under last quarter’s definition.

Code applies them. Kepler’s deterministic engine computes each figure from cited facts using the definition’s formula, shows the formula and every input, and returns the same number every time. The model reads the question and writes the answer, and it never does the arithmetic.

The inputs are already extracted. Extraction models we train read every page at ingest, including the reconciliation tables in earnings releases, the definitions sections of credit agreements and the capital account statements in fund reports, and keep each fact with its period, scale and source. Public filers’ statements assemble from their own XBRL, as filed or as restated.

Different definitions, side by side. When the company’s adjusted EBITDA, the lender’s covenant EBITDA and your house figure differ, Kepler shows all three, labeled, each traced to its source line.

Corrections become rules. When an analyst corrects a mapping, the fix becomes a rule applied everywhere the same case appears, and your playbooks (the comp sheet, the IC memo, the LP report) run on your definitions every time.

Everywhere, under your rules. The same definitions apply in the Kepler app, Excel, PowerPoint, Claude, ChatGPT and our API. Changes to a definition pass the same access check and audit log as everything else on Kepler. Our engineers built this layer before at Palantir.

Questions firms ask in the first week:

• “Trading comps for US cybersecurity on each filer’s own adjusted EBITDA and on ours, side by side.”
• “Net IRR for every fund we hold, with and without the subscription line, as of June 30.”
• “Leverage ratio and covenant headroom for each portfolio company on its own credit agreement’s definition.”

Try Kepler at app.kepler.ai, see how definitions fit the wider system in what enterprise intelligence is, or book a demo and bring your house definition of adjusted EBITDA.

Frequently asked questions

How do you get AI to use your firm's definitions?

Store each definition once, outside the prompt, as a formula with an owner, a version and an effective date, and have code compute every figure from it. That's how Kepler works: your firm's terms live in your firm's ontology, so adjusted EBITDA, net IRR or exposure means your definition in every answer, with the history of each change. Prompts and custom instructions get applied some of the time and dropped the rest.

What is a semantic layer in finance?

A semantic layer maps business terms to precise data definitions, so revenue or net exposure is defined once and computed the same way by every tool. In BI it sits over warehouse tables. At an investment firm many definitions draw on facts inside documents, such as a credit agreement's EBITDA, so the semantic layer has to reach documents too, which is the job of an ontology. Kepler builds one for each firm.

Why does adjusted EBITDA differ between sources?

Because each adjustment is a choice. A company's reported adjusted EBITDA reflects its own add-backs, a lender's covenant EBITDA follows the credit agreement's definition and its caps, and a firm's house figure applies its own view of what recurs. SEC rules require companies to reconcile non-GAAP measures to the most comparable GAAP measure, which is the place to start when comparing them.

What is net IRR with and without a subscription line?

A subscription line lets a fund borrow before calling capital, which shortens the time LP money is at work and usually lifts early IRR. Showing IRR both ways isolates that effect. ILPA's Performance Template, released in January 2025, reports gross and net figures both with and without the impact of fund-level subscription facilities.

Can custom instructions in ChatGPT or Claude hold a firm's definitions?

They can describe a definition, and the model may follow it, yet nothing enforces it: instructions compete with everything else in context, the model still does the arithmetic, and there's no version history. Custom instructions work well for style and format. For definitions that drive numbers, store the formula outside the model and compute it in code.