What Is Enterprise Intelligence?

Enterprise intelligence is a firm’s documents, systems and institutional knowledge turned into one living model of the firm that AI can reason over, prove and act on, under the firm’s own rules. Kepler builds it for investment firms: our engineers build each firm’s ontology from its own data, every document is read in full, and code computes and cites every number, so answers come in the firm’s own terms, every figure traces to its line, and agents do recurring work on their own with every step on the record.

The idea matters now because models are no longer the scarce part. Frontier models get smarter and cheaper with every release. What they still lack is context (your entities, your definitions, your history) and trust (a way to prove an answer before someone stakes money on it). This guide uses Kepler as its worked example.

A working definition

Each phrase in the definition carries weight.

• One living model of the firm. Every company, fund, vehicle, manager, person, deal and document the firm works with resolves to a single record. The firm’s definitions are attached, and every relationship points to its source. “Living” means it updates as new documents arrive and remembers what was true on any date. The technical name for this model is an ontology; our guide on what an ontology is for investment firms covers it in depth.
• That AI can reason over. Questions run against that model. “Our exposure to this company” means the firm’s holdings, through the firm’s vehicles, on the firm’s definition of exposure.
• Prove. Every number traces to the page, line or cell it came from, and every calculation shows its formula and inputs. The same question asked twice returns the same number.
• Act on. Agents run on schedules and events (a new filing, a capital call notice, a guidance change) and deliver finished work, acting with the permissions of the person who set them up.
• Under the firm’s own rules. Permissions, information walls, audit logs and human sign-off are enforced on every read and every action.

These five depend on each other. Proof needs the model of the firm, because a number is only provable when the system knows which fact was meant. Action needs rules, because an agent that runs overnight has to be held to the same access as its owner.

How the term is used elsewhere

“Enterprise intelligence” gets used loosely. Databricks describes it as an organization-wide capability that combines business intelligence, knowledge management, enterprise search and AI on a governed data foundation. Some research vendors use the phrase as a product name for searching a customer’s internal documents, and older usage treats it as a synonym for enterprise business intelligence.

Our definition is narrower, and it’s the standard we built Kepler to meet: a system qualifies as enterprise intelligence when its AI can show its work down to a single number, apply the firm’s definitions without being told each time, and act only inside the firm’s permissions.

Why now: intelligence is getting cheap, trust and context are not

Two trends cross in 2026.

Reasoning is becoming a commodity. Several labs ship frontier models, each release beats the last on public benchmarks, and the price of a given level of capability keeps falling.

Meanwhile, the problems that block AI from high-stakes work don’t shrink with model size. A frontier model on its own:

• writes numbers rather than pulling them, so a wrong figure can look exactly like a right one;
• attaches citations after writing, so a citation can prove a source exists without proving the number came from it;
• reads only the passages it retrieves, so “every fund agreement with a key-person clause” comes back as the top few hits;
• doesn’t know the firm’s funds, portfolio companies, internal code names or definition of adjusted EBITDA;
• gives every firm the same answer to the same question.

Vinoo Ganesh, our CEO and co-founder, sums up the result in one line: “AI is everywhere, except where it matters most.” It drafts emails and summarizes meetings. It isn’t yet trusted with the investment decision, the report to investors or the compliance check.

In enterprise intelligence, context comes from the firm’s own data, organized into a model that only that firm has. Trust comes from an architecture where code handles the facts and the model handles the language. Both persist when the models change. We think of it as each firm’s own harness: the models are rented, and the ontology, facts, rules and playbooks stay the firm’s own and outlive every model.

The eight parts of an enterprise intelligence system

Part	What it does	How to test it
Connect	Brings in public data and the firm’s own systems, read-only, kept in sync	Add a document this morning; ask about it this afternoon
Read	Reads every page in full at ingest, scans and tables included, and keeps each fact beside its passage	Ask for a complete list and count it against one built by hand
The firm’s model	Resolves every name to one record and stores the firm’s definitions	Ask about a portfolio company by its internal nickname
Facts	Holds each number with its unit, scale, period and source, as filed and as restated	Ask for a figure as first reported and as restated
Reason, with proof	The model plans and writes; code retrieves and computes; citations are checked	Click ten numbers; ask the same question twice
Rules	One access check on every read, walls between deals, nothing read taken as an instruction	Ask as a user who shouldn’t see a deal
Act	Agents on schedules and events, acting as their owner	Set an alert on a filing type and see what arrives
Everywhere	The same answers in the app, Office, chat assistants, Slack and an API	Pull one number in Excel and in the app, then compare

Reading every document before anyone asks is what makes complete lists possible: a question becomes a lookup over facts already extracted, instead of a fresh search that ranks a handful of passages. Facts stored with their units, periods and sources let code compute answers a person can audit. For the auditing itself, see how to check AI-generated numbers in financial research.

The six rules it keeps

Any system that touches a firm’s decisions needs rules as concrete as the ones a compliance team applies to people. We hold every part of Kepler to six:

1. One access check. Every read, by a person or an agent, passes the same permission check, mirroring the firm’s own permissions.
2. Every number cited. Each figure is pulled from a source, computed from cited inputs, or clearly marked as an estimate.
3. Every piece of work is a run. Each task is recorded, tied to the person who started it, and checked against that person’s access.
4. Nothing read is an instruction. Documents, emails and web pages are data. Text inside them can’t redirect an agent, which is the defense against prompt injection.
5. Every edit is a version. Each answer records the version of the data and definitions it used, so it can be reproduced later.
6. People decide. Nothing leaves the firm or gets shared without a person approving it.

These map closely to what examiners and auditors ask about AI use. Our guide to AI governance for an SEC exam shows how.

How it differs from enterprise search, BI, RAG chatbots and copilots

Approach	Built for	Where numbers come from	Knows the firm’s definitions
Enterprise search	Finding documents and passages across systems	Whatever the passage says	Matches text only
Business intelligence	Dashboards on structured warehouse data	Queries over modeled tables	Only metrics someone modeled by hand
RAG chatbot	Answering from passages retrieved at question time	Written by the model beside a citation	Only what the retrieved text says
Copilot	Helping one person inside one app	Written by the model	Only what’s open or pasted in
Enterprise intelligence (Kepler)	Work the firm can stake a decision on	Pulled from source and computed by code	Stored once in the firm’s ontology

Enterprise search finds things. It returns ranked hits, which is the wrong shape for a question that needs every match, and it doesn’t know that “SpaceX” in the CRM and “Space Exploration Technologies Corp.” in a fund report are the same company.

Business intelligence is deterministic and governed, which is why finance teams trust it. It covers only data someone has already structured into tables, and most of an investment firm’s knowledge sits in PDFs, emails and spreadsheets. Enterprise intelligence applies BI’s discipline (defined metrics, computed numbers) to that unstructured majority.

RAG chatbots retrieve a few passages when you ask and have a model answer from them. They’re quick to stand up. Anything the retrieval step misses doesn’t exist for the answer, and the numbers are still typed by the model. Our guides to RAG for financial data and whether ChatGPT can do financial analysis walk through where that breaks.

Copilots help one person with one task inside one application, and their output depends on what the user brought to them.

Enterprise intelligence connects to the same sources and shows up in the same places people already work, including Excel, Claude and ChatGPT. For a survey of the products in this market, see the best AI tools for investment research.

What to ask a vendor

These eight questions separate enterprise intelligence from a chatbot with connectors. Our guide to evaluating AI vendors has the full set of 25, with what a good answer looks like.

1. Where does each number come from: written by the model, or pulled and computed by code?
2. What does a citation point to: a document, a passage, or the exact line or cell? Is it checked before I see it?
3. If I ask the same question twice, do I get the same number with the same sources?
4. Is every document read in full before I ask? Can it return a complete list and say so?
5. Where do our definitions live, and who maintains them when they change?
6. How does it handle a restated figure, or two sources that disagree?
7. Do agents pass the same permission check as people, and is every run logged?
8. Who builds the model of our firm, and how long does it take?

How Kepler builds enterprise intelligence

We started Kepler to build this category for the work where a wrong number costs the most. Our engineers built this layer before at Palantir. We run for investment firms first (hedge funds, asset managers, private equity and credit, allocators, banks), and the same platform runs for law firms, insurers, manufacturers and healthcare teams.

Your firm’s ontology. Our engineers build each firm’s ontology, which we call Kepler Orbit, from its own documents and systems. Every company, fund, vehicle, manager, deal and person resolves to one record, and the firm’s terms, from net IRR to its own adjusted EBITDA, carry one definition. Every link is a claim tied to the sentence it came from. The graph keeps time, so any figure can be read as filed, as restated or as of a date, and it grows richer every week the firm uses it. Details are on our ontology page.

Every document, read in full. Public data comes preloaded for 14,000+ companies across 120+ countries and 30+ languages: 26M+ SEC filings in 20+ form types, 40M+ investor relations documents and earnings call transcripts, and market data and news. The firm’s own sources (SharePoint, OneDrive, Box, Egnyte, Google Drive, Outlook, Slack, CRM, data rooms, models and notes) connect read-only and stay in sync. Extraction models we train read every page at ingest and keep each fact, entity and relationship beside its passage. A request for every side letter with a most-favored-nation clause comes back complete, and a follow-up question is a lookup.

Numbers computed by code. On Kepler, the model never writes a number. Our deterministic engine pulls each figure from its source and runs every calculation as code over cited inputs; financial statements assemble from filers’ own XBRL with no model in the path. Click any figure to land on its page, line or cell, and a calculated figure opens as a small sheet with its formula and inputs. We check every citation against its source while the answer is written and strip any that don’t match. When two sources disagree, you see both.

The right model for each job. Small models we train do the reading, and frontier models handle the question and the write-up. We route each task to the best model for it and swap in better ones as they ship, so every new model improves a firm’s deployment without a rebuild.

Agents you can leave alone. Agents run on schedules and on events: a new 10-Q, a guidance change, news on a holding, a capital call notice, a file landing in a data room. They act exactly as the person who set them up, and nothing they read can steer them. Results arrive as alerts with the exact quote, digests before the morning meeting, updated models and drafted memos. The firm’s playbooks, such as the IC memo in house style, are saved once and applied every time, and each correction improves the firm’s models privately.

Everywhere, and built to be built on. The same intelligence works in the Kepler app, Excel, PowerPoint and Word, inside Claude and ChatGPT, in Slack and email, and through our API, MCP server, Python SDK and webhooks. Firms with engineering teams build their own agents and apps on Kepler, and every number still comes back with its source; our guide to build vs buy covers which parts a firm should own.

Rules, people and security. One access check covers every read, with walls between deals and every query, calculation and run logged with its lineage. Kepler is SOC 2 Type II audited and GDPR and CCPA compliant. Each firm’s data is isolated at the identity layer and never trains models other firms use, and we deploy in our cloud or the firm’s. More on our security page.

Forward-deployed. Our engineers work on site with each firm. Public data is ready on day one, and in the first week we connect the firm’s sources, index every document and build its ontology. Whatever we build for one firm that the next would need goes back into the platform, which is why each new deployment starts further along.

Questions firms ask on day one look like these: “What’s our look-through exposure to this company, across every vehicle?” “Tell me when guidance changes for anyone we cover.” “Draft the IC memo the way we write them, with every figure cited.” The full stack is laid out on the platform page, and our comparison page sets it against a frontier model on its own.

Try Kepler at app.kepler.ai, or book a demo to see enterprise intelligence on your firm’s own data.

Frequently asked questions

What is enterprise intelligence?

Enterprise intelligence turns a firm's documents, systems and institutional knowledge into one living model of the firm that AI can reason over, prove and act on, under the firm's own rules. Kepler builds it for investment firms: answers in the firm's own terms, every number computed in code and traced to the line it came from, agents that run on schedules and events, and one access check on everything the AI reads.

How is enterprise intelligence different from business intelligence?

Business intelligence computes governed metrics from structured tables in a data warehouse. Enterprise intelligence brings that same discipline (defined metrics, numbers computed by code) to the documents, emails and spreadsheets where most of a firm's knowledge lives, and adds AI that can answer questions, cite sources and act on the results.

Is enterprise intelligence the same as enterprise search?

No. Enterprise search finds documents and passages and returns ranked hits. Enterprise intelligence resolves every name to one record, applies the firm's definitions, computes numbers from cited inputs, returns complete lists, and runs agents under the firm's permissions. Search is one ingredient of it.

What is an enterprise intelligence platform?

Software that connects a firm's sources, reads every document in full, builds a model of the firm (an ontology), computes and cites every number, enforces the firm's access rules and audit trail, and delivers answers and agent work wherever people work. Kepler is an enterprise intelligence platform built first for investment firms.

Does enterprise intelligence depend on which AI model you use?

It shouldn't. A well-built system routes each task to the best available model and swaps in better ones as they ship, while the firm's ontology, facts, rules and playbooks stay in place. Kepler works this way: extraction models we train read the documents, frontier models take the question and draft the answer, and each better model improves a firm's deployment without a rebuild.