As AI becomes ubiquitous across the asset management space, firms are increasingly faced with choosing technology and tools that will maximize their AI capabilities. The right choice can improve efficiency, accuracy, and decision-making, while a misstep could mean major operational and reputational risk.
So, how can technology and operations teams assess their readiness for AI across their investment operations?
We published five questions on this two years ago but are now updating it in 2026 as two considerations that barely registered in 2024 now sit at the center of the decision: how your own agents talk to your platform and where your data goes when a model processes it.
Here are the five questions, updated for where the industry and the technology have moved.
1. Do You Have the Data Infrastructure in Place to Support AI?
AI models are only as good as the data they can reach. That means data which is accessible, clean, pre-processed and held in well-organized repositories, available continuously instead of once a day after an overnight cycle.
Data spread across legacy systems limits what any model can do and the limit is structural, so no amount of tuning gets around it. Without cloud-native foundations, integrating modern AI tools and running machine learning or large language models against your own operational data stays difficult no matter how capable the models get.
Assess: Is your data accessible and ready for processing as events arrive? If not, data management and cloud-native infrastructure need to be done as a prerequisite and doing that first makes everything after it easier.
2. Does Your Tech Infrastructure Support AI Now and as It Evolves?
Infrastructure has to accommodate the AI you are deploying today and the agentic operating models arriving next. A modern cloud-native stack supports both models that summarize and suggest, which need access to good data, as well as agents that need to respond in real-time, trigger workflows, and prove what they’ve done, because a cloud-native stack gives you elastic compute for variable AI workloads, event-driven processing so data does not wait for end-of-day cycles and the API coverage agents need to operate. Fixed-capacity on-premise infrastructure makes a month-end validation sweep across millions of positions either unaffordable or impossible.
Ask yourself: Can your stack scale with what AI becomes over the next three years, or was it built for a batch world that agents cannot operate inside?
3. Is Data Quality and Reliability Guaranteed?
Reliable data is the foundation of accurate AI. Cloud-native technology makes quality easier to maintain because ingestion, validation and processing can run continuously, with anomalies surfaced as they appear instead of in a variance report the following morning.
On-premise and cloud-enabled platforms make this harder. Limited automation and delayed access to updated data mean quality checks happen in batches and problems get found after they have already propagated downstream.
When every correction and restatement is preserved against both its economic date and its system date, your models learn from the difference between a record that was wrong and later corrected and one that was right all along. Overwrite-based systems throw that distinction away, along with the audit trail your regulator will ask for.
Consider: Are your data management processes designed for consistency and continuous validation? Does your platform preserve the history AI needs to explain itself?
4. How Will AI Tools and Agents Integrate With Your Existing Systems?
Two years ago this question was about databases, operational software and getting data to flow cleanly between them. That still matters but the surface area has grown considerably.
Integration now runs across three layers. There is the data layer, where databases sit alongside data lakes and platforms like Snowflake and Databricks. There is the API layer, where systems exchange information programmatically. And on top of both sit the agent protocols, which is where most of the change has happened.
Model Context Protocol (MCP)
MCP is the open standard that gives AI agents a universal interface to data, tools and workflows, replacing a bespoke integration for every tool. For asset managers this matters because most firms are now building their own AI capability and those tools need to reach your accounting data and they need to reach it under the same entitlements as your people.
Platforms that expose MCP let your own AI tools speak to it in a language they already understand, whereas without it you end up with every internal AI project devolving into a custom integration project.
Agent-to-Agent (A2A)
A2A is the protocol that lets agents share context and execute sequences together, including across organizations. FundGuard provides agents your agents can talk to, so your team can build its own AI tools and its own agents and have them work directly against the accounting data, under governed permissions, without waiting for anyone’s roadmap.
Bottom line: If data is siloed or moved by hand, AI-driven decisions slow down and if your platform has no answer on MCP and A2A, your own agents are locked out of your own accounting data.
5. How Will AI Models Be Managed, Maintained and Kept Secure?
Managing models is partly a question of platform choice. Traditional machine learning runs well on environments like Databricks with compatibility across cloud providers. Large language models are served through cloud offerings including Azure OpenAI and Amazon Bedrock.
The evolved question for 2026 is where your data goes when a model processes it.
Public model endpoints are the sticking point for most institutional clients and reasonably so. Sending position data, investor information or fund documentation to a public model provider is a hard conversation with a CISO and a harder one with a regulator. This is why production deployments increasingly call models through Amazon Bedrock, which serves foundation models from providers including Anthropic inside your own AWS environment. Amazon states that customer data submitted through Bedrock is not shared with model providers and is not used to train the underlying models, so a query against Claude never leaves the boundary you control.
FundGuard runs this way for exactly that reason. Client data stays inside a walled garden, retrieved through secure APIs and never used to train our models. Confidentiality comes from the architecture instead of a clause in a contract.
Review: What are your maintenance needs, where will your models be served from and can your vendor show you the boundary your data never crosses?
Get the AI Foundation Your Firm Needs This Year and Beyond
Adopting AI was never mainly about picking the best tools. The foundational elements decide the outcome: data infrastructure, data quality, integration and the security model around it all. Open agent protocols so your AI strategy stays yours, and a serving model that keeps your data inside your own boundary.
Get those right and every model you deploy afterwards works better.
Frequently Asked Questions
What should asset managers assess before adopting AI in investment operations?
Five things: whether data is accessible and continuously available, whether the infrastructure can scale with AI as it evolves into agentic operations, whether data quality and history are reliable enough to train and explain models, how AI tools and agents integrate through data, APIs and agent protocols, and how models are served and secured.
Why does cloud-native architecture matter for AI in asset management?
AI needs continuous data, event-driven workflows and elastic compute. Cloud-native platforms provide all three. Legacy platforms hosted in a cloud environment still run batch architecture underneath, which means models can only evaluate what already happened.
What is Model Context Protocol (MCP)?
MCP is an open standard giving AI agents a universal interface to data, tools and workflows, so a firm building its own AI capability can connect it to a platform without a custom integration. For asset managers it decides whether internal AI tools can reach accounting data directly, under existing permissions.
What is Agent-to-Agent (A2A) communication?
A2A is the protocol letting agents share context and execute sequences together, including across organizations. It allows a firm’s agents to work with a provider’s agents, so cross-firm workflows like reconciliation queries or audit requests can run agent to agent with governance and audit trails intact.
How do firms use large language models without sending data to public providers?
Services like Amazon Bedrock serve foundation models inside a firm’s own cloud environment. Amazon states that data submitted through Bedrock is not shared with model providers and is not used to train the models, so queries against a model like Claude stay within the boundary the firm controls. FundGuard uses this approach.
What changed in AI assessment between 2024 and 2026?
While some of the questions stayed the same, two considerations moved to the center: agent interoperability through MCP and A2A, because firms are now building their own AI tools and need them to reach accounting data, and model serving security, because institutional clients will not send operational data to public model endpoints.
Does bitemporal data matter for AI?
Yes. Preserving every correction against both its economic date and its system date gives models a clean training signal and gives regulators a reconstructable audit trail. Platforms that overwrite prior state lose both the moment a correction lands.
Intelligence at the Core
Our Intelligence at the Core whitepaper goes through why architecture decides what AI can deliver in investment accounting, with the production evidence, the five pillars of AI-readiness and seven questions to put to any vendor.
