Carta

Take the agents out and the platform still runs. Carta operated for over a decade as a capitalization table and fund administration system before any of this existed. Today the general ledger and the cap table work without the agents, and so do the tax product, the compliance tools and the investor portal. That is starting to change.

Unlike a chat layer bolted on top, the agentic ingestion feeds the general ledger directly and the monitoring agents sit in the transaction path, so the AI is moving toward the core rather than staying beside it. If the accounting close stops working without the agents, the AI becomes central to the product.

Agents work from the governing partnership agreement. Each detected signal is processed in that context, so behavior is grounded in the document that actually controls the fund rather than in a generic rule set. The company also states plainly that agents run continuous background health checks while the last mile belongs to experienced human experts. That last mile statement is positioning rather than a described mechanism.

There is no threshold, no review rate and no escalation criterion, and nothing says which agent outputs reach the general ledger without a person seeing them. Set against a stated goal of removing the need to log events manually, that is the gap that matters.

The auditability claim is real, but it covers the underlying data rather than the models: a deep layer of auditable data running from a single company's capitalization table to a fund's general ledger. Nothing describes how the models themselves are evaluated. There are no accuracy or extraction error figures for the ingestion agents that write to the general ledger, no documented failure modes, no drift policy and no model documentation.

The continuous health checks are agents doing monitoring rather than monitoring being done of the agents. For a product whose output lands in fund accounting, extraction accuracy is the standing question and it is unanswered.

The generative products have adoption evidence of their own: more than 1,500 companies and firms using the assistant plugins within months of launch. Platform scale is documented in volume terms that are checkable in kind, with monthly document processing in the tens of thousands and over a million journal entries a year. Customer quotes appear on the platform pages. No named institution is attached to a measured outcome from the agentic products, so adoption is evidenced and performance is not.

The promise is clear and the detail behind it is missing. Customer data is never used to train models and stays private and internal, with contractual protections named with the model providers. That arrives as one marketing sentence rather than as a governance document. There are no enumerated data sources and no stated isolation guarantee between firms. Retention and purge periods, model hosting policy and supplier posture are not published either, so a buyer cannot verify or enforce the commitment.

The central commitment is made plainly and without hedging: customer data stays private and internal and is never used to train models, backed by stated contractual protections with the providers involved. It stands alone. There is no data processing agreement or subprocessor list, no retention or purge terms, and no residency commitment or named privacy regime alongside it.

The attestation is named but loosely worded. The generative material states that the intelligence runs in a closed service organization control (SOC) environment with strict contractual protections. That describes an environment as attested rather than stating that the company has a report of a given type, scope and period. There is no public trust portal, no statement on report availability and no penetration testing disclosure.

A fund administrator of this size holding client books and records would be expected to hold both a SOC 1 report over financial reporting controls and a SOC 2 report over trust criteria, and neither is published.

Carta's supervised status is unconfirmed. The company has historically operated regulated entities in the United States, plausibly including securities transfer agent registration and a broker dealer subsidiary for its private market trading venue. Transfer agent registration would matter in a way that most group licenses do not, because it attaches to securities recordkeeping. That is the product itself rather than an adjacent business.

No fairness testing or governance disclosure. Exposure is lower than in consumer facing lending or pricing, since the subjects are funds and transactions rather than individuals, but it is not absent: the relationship intelligence product ranks leads and evaluates targets from years of historical data, and ranking which opportunities and which people surface first is a consequential ordering even when no individual is being scored for credit.

Stronger than silence and worse for a buyer: the published material carries an express disclaimer that the company does not assume any liability for reliance on the information provided. Nothing allocates responsibility for an agent that misreads a capital account statement and writes a wrong figure into a general ledger, and nothing describes detection or correction when that happens.

The assistant platforms the product interoperates with are named openly, and the open protocol server architecture is named as the mechanism. Contractual protections with providers are asserted. The model behind the proprietary ingestion agents that write to the general ledger is not named. Naming the assistants a customer can drive the platform from is a different disclosure, and it leaves that second question unanswered.

The integration position is structural rather than claimed. The platform is itself the system of record, holding the general ledger and the capitalization table. It connects outward to bank feeds, portfolio company data and investor records, and payment authorization runs through named financial institution partners at around $3 billion of volume.

It also exposes an open protocol server architecture and ships plugins that let a firm run fund administration, cap table and dealflow workflows from inside general purpose assistants, with adoption in the low thousands of firms. A data warehouse serves the same data to natural language queries.

Software as a service with nothing published on region, residency, tenancy or hosting arrangements. For a platform holding the books and records of funds across multiple jurisdictions, and now routing that data through general purpose assistant platforms via plugins, residency is a question a buyer will ask and the material does not answer it.

No price is published for the fund administration, loan operations or analytics lines. The company has historically published tiered pricing for its capitalization table product aimed at early stage companies. If those tiers are still public, at least one product line is priced openly.

Both sides of the private markets relationship are served on one platform. Clients include venture and private equity funds, private credit funds, funds of funds, institutional limited partners and allocators, general partners, portfolio companies and corporates. The work spans fund administration, fund accounting, tax and compliance, plus capitalization table management, loan administration, investor reporting and relationship management. Scale runs to roughly 9,000 funds and more than $203 billion of assets, about 125,000 allocators and more than 50,000 capitalization tables.