As asset managers move beyond AI pilots and begin integrating agents into investment operations, they're confronting a familiar challenge: the limitations of their existing technology infrastructure. An AI agent may be capable of automating complex workflows, but its effectiveness depends on the quality, timeliness and accessibility of the data it works with.
For investment accounting, this raises an important strategic decision. Can firms extend the capabilities of their existing systems to support AI-driven operations, or does realizing AI's full potential require a different approach to the underlying system of record?
Seema Amble of Andreessen Horowitz recently explored this question in “Is Software Losing Its Head?”, outlining three paths for software buyers:
- Keep the incumbent system and add agents
- Build the system of record entirely in-house
- Buy an AI-native replacement
Her framework has particular relevance for investment accounting, where systems of record must do far more than make data available. They must calculate and maintain accurate positions, cash balances and NAVs, support complex financial instruments and meet demanding regulatory and audit requirements.
Each path offers potential advantages, but also significant trade-offs. Here's how those choices play out for asset managers and what firms should consider as they build AI into their operating models.
An InvestOps Decision Already Taking Place, Whether You Know It or Not
For many asset managers, this decision is already taking shape, even if it hasn't been framed as an enterprise-wide technology strategy. One team may be adding AI assistants to an existing accounting platform while another explores an internal data architecture and a third evaluates newer technology providers. These initiatives often develop independently, but collectively they have implications for the firm's long-term system of record strategy.
Firms have other routes and most of them lead back to one of the three. Outsourcing investment accounting to a fund admin or asset servicer hands the decision to the provider, which faces the same three paths across its own book. Phasing a change by keeping the incumbent for some funds while moving others means running path one and path three side by side for a period. Waiting is also a decision and in practice it keeps firms on path one while that path’s costs keep growing.
What raises the stakes is how much the industry now expects from AI in operations. BCG’s 2026 Global Asset Management Report estimates that agentic workflows could raise investment operations capacity by 55 to 65% and cut operational costs by around 40%, and it points firms toward rebuilding workflows around agentic AI to capture that value. Alpha FMC’s 2026 outlook research finds fee pressure pushing firms to simplify their technology estates while they invest in data foundations and automation. Each of the three paths responds to those pressures in a different way.
Path One: Keep the Incumbent and Add Agents on Top
With path one, there is no migration, agents can reach users within months and the incumbent vendor often supplies them. Amble describes incumbents in this position as going headless, exposing their data and functions through APIs so agents can use the system without its screens. In doing so, she argues, they are betting that the data layer will remain the source of value.
For investment accounting, that bet depends on what the incumbent’s data layer can give an agent, and on legacy batch platforms the answer is limited in four ways.
Current state
A batch platform settles its data into final form after the overnight cycle, so an agent working during the day reasons about last night’s positions while the portfolio has moved on.
Action
Agents need to write back by correcting records, releasing exceptions and triggering the next step, and an API added to a reporting surface gives them plenty to read and little they can safely change.
Explanation
Overwrite-based ledgers replace a balance when a correction lands, so the history an agent needs to justify a decision has to be reassembled from separate transaction stores, when it can be reassembled at all.
Breadth
When listed equities, derivatives and private credit sit in separate systems, an agent sees part of a portfolio whose most useful signals lie in the relationships between holdings.
The impact
These limits tend to stay hidden in a pilot running on a handful of funds with clean extracts and they surface in production, when the agent has to work from the platform’s real data on its real schedule. AI that runs in a reporting layer or a separate analytics environment can only evaluate what has already happened, a point FundGuard made in The Real Work of AI in Investment Accounting, and agents on path one inherit that ceiling.
Path one also adds integration work over time. When accounting, reconciliation, exceptions and reporting sit in separate systems, each agent has to learn each environment and each vendor upgrade can break the automation built against it. Automation tends to expose the architectural complexity a firm already has.
Path one can still be the right short-term call for a firm in the middle of another transformation or close to a contract renewal, as long as it is treated as a bridge and the firm is clear about what agents can’t do on that foundation.
Path Two: Rebuild the System of Record from Scratch
Path two appeals to firms with strong engineering teams, and AI has made it look more achievable than it used to. Coding assistants and agent frameworks shorten the time it takes to stand up a data model, a set of APIs and a working ledger prototype. In Amble’s description, this path means a firm building its own data model, operational logic, permissions, audit trails and integrations, along with its own agents.
AI lowers the cost of recreating the first 80% of a system of record, she writes, and the remaining 20% of exceptions, approvals, compliance requirements and edge-case workflows is still what separates “a useful wedge from a true replacement.”
Investment accounting lives in the last 20% of a system of record, the part AI doesn’t make cheap. Tax lot relief methods that differ by transaction type, corporate actions that ripple across several books, multi-class funds with different fee structures, NAV error policies, private markets waterfalls and the regulatory reporting built on all of them are where the years of engineering go. A prototype ledger that handles a clean equity book in weeks still has most of that work ahead of it.
Timelines are the other concern. In Citi and CREATE-Research’s 2026 survey of 221 asset managers, 57% expected their digital transformation to take more than five years. A full rebuild of the system of record sits at the demanding end of that range, because it also has to prove its accounting from first principles, run in parallel with the platform it replaces and satisfy the same regulatory scrutiny, while the business keeps launching products the new system hasn’t been designed for yet.
A firm that finishes the rebuild then owns it, as a software vendor with one client, responsible for regulatory changes, new product structures and security reviews for as long as the platform runs. Rebuilding from scratch appeals to engineering pride, and it tends to worry anyone who has run a core platform migration and is honest about how long they take.
Path Three: The AI-Native Replacement
Path three means buying a system of record designed for the agent era, which Amble describes as software built to be machine-readable, with agent orchestration treated as a first-class feature.
In investment accounting, AI-native has a specific meaning that goes well beyond a modern interface:
- Event-driven processing, where trades, prices, cash movements and corporate actions update the record as they arrive, so it is current during the day, with private holdings at their latest known state between valuations
- Versioned data, where a correction creates a new version and keeps the old one, so any number can be traced through what changed and when
- One dataset across public, private and digital assets, with the investment book, the accounting book and regulatory views derived from the same events
- Governed access for agents through open standards such as Model Context Protocol and Agent-to-Agent communication, under the same entitlements as the firm’s people
Amble asks a pointed question of incumbents going headless: with the interface stripped away, how is what remains different from a Postgres database, a well-designed schema and an API? The question applies just as well to anything sold as AI-native. In investment accounting, the defensible answer is the accounting engine, the operational logic that turns transactions into positions, cash, accruals and NAV, with controls applied as data enters the books. Amble’s own view is that agents may remove muscle memory as a moat while operational logic and context remain, and in investment accounting that logic is most of the value.
The practical test is whether the platform manufactures its accounting data, calculating positions and NAV from the underlying activity, or aggregates numbers calculated elsewhere and exposes them through an API, which makes it a data layer whatever it is called.
Path three does have a cost of its own, which is the migration. Replacing an investment accounting platform is serious work, with positions, cash, tax lots and history to move while teams learn a new system and keep running the old one. That cost is paid through a defined process with parallel running and acceptance criteria agreed in advance. The costs of path one recur with each new agent, and the costs of path two run on for the life of the platform.
Partnering is already part of how many managers plan to scale. After all, in the Citi survey, 44% saw collaboration with third parties as a way to build scale, enhance capabilities and share risk.
Why Compliance Raises the Stakes for Financial Services
Amble expects compliance to slow the shift in financial services. In deeply compliance-bound categories such as financial services, she writes, the incumbents’ bet on the data layer may hold for some time. That is a fair description of how regulated firms buy technology, cautiously and on long timelines.
The same compliance requirements also decide which path holds up once agents start doing real work. A regulator asking how a NAV was derived, or what the firm knew when an agent approved an exception, is asking for lineage and for the firm’s ability to reconstruct its own state at a past point in time. Accountability for a regulated number stays with the firm, whether a person or an agent produced it. Operational resilience frameworks such as the EU’s Digital Operational Resilience Act add the expectation that a firm can evidence how its systems behave under disruption.
Each path meets those requirements differently. On path one, the incumbent’s data layer was built to overwrite and to settle overnight, so an agent’s decisions rest on records that can’t show what was known at the time. A rebuild can meet them in principle, at the cost of recreating the permissions, audit trails and regulatory reporting that make up so much of the hard 20%. An AI-native platform meets them when it was designed for them, with versioned records, entitlements that govern agents and people alike and data that stays inside a controlled boundary when models process it.
Compliance narrows the decision, because the requirements that matter most to regulators are the same ones an agent-based operating model depends on.
Choosing a Path
Most firms will run more than one path for a while, often without a single owner for the decision. The useful step now is to name the path the firm is on, check what each AI project assumes about the system of record and decide whether that foundation can carry the agents being built on it through the next five years.
Our Intelligence at the Core whitepaper sets out the architecture questions to put to any vendor, and a FundGuard demo shows what path three looks like in a live investment accounting environment.
Frequently Asked Questions
What are the three paths facing asset managers in the AI era?
Keeping the incumbent investment accounting platform and adding AI agents on top, rebuilding the system of record in-house or buying an AI-native replacement.
What does headless architecture mean in financial services?
Headless architecture separates a system’s data and functions from its user interface and exposes them through APIs, so agents and other applications can use the system without its screens. For an incumbent investment accounting platform, going headless typically means wrapping the existing engine in APIs and agent tools while its batch processing and data model stay the same.
Why do AI agents struggle on legacy investment accounting platforms?
Batch platforms settle data overnight, so agents reason about stale positions during the day. Overwrite-based ledgers lose the history agents need to explain decisions, reporting-layer APIs offer little that agents can safely act on and asset-class silos give agents a partial view of the portfolio.
Should asset managers build or buy their investment accounting system of record?
AI makes the first 80% of a system of record cheaper to build, while most of investment accounting sits in the remaining exceptions, approvals, compliance requirements and edge cases. Rebuilds tend to run for years and leave the firm maintaining the platform for as long as it runs. The cost of buying an AI-native platform is the migration, which runs through a defined process with parallel running and acceptance criteria agreed in advance.
What makes an investment accounting platform AI-native?
An AI-native platform uses event-driven processing to keep the record current during the day, with private holdings at their latest known state between valuations. It versions its data so each correction is preserved, holds public, private and digital assets in one dataset and gives agents governed access through open standards such as MCP and A2A, under the same entitlements as the firm’s people. Above all, it manufactures its own accounting data from the underlying activity.
How does compliance affect the choice between the three paths?
Compliance slows adoption, since regulated firms change systems of record cautiously. It also narrows the choice, because regulators expect lineage, the ability to reconstruct what was known at a point in time and clear accountability, which are the same things agents need from a system of record.
Are there other options beyond the three paths?
Firms have other routes and most of them lead back to one of the three. Outsourcing investment accounting hands the decision to a fund administrator or asset servicer, which faces the same three paths across its own book. Phasing a change across funds runs paths one and three side by side and waiting keeps firms on path one.
