AI Operational Resilience: Business Continuity for Investment Offices
Imagine an investment committee meeting tomorrow. An analyst has used AI to compare manager responses, review fund terms and draft the recommendation, and the team has come to reach much of its research through a conversational interface. Today the AI provider has an outage. The documents are still there and the meeting still needs to happen. The practical question is whether the team can retrieve the evidence, finish the review and show what has already been approved.
AI operational resilience is the ability to continue essential work, preserve decision records and recover interrupted workflows when an AI service becomes unavailable or unreliable.
As investment offices use AI across research, due diligence and investment workflows, continuity planning needs to cover the failure of models, interfaces and the services beneath them.
Large outages across cloud, enterprise software and public infrastructure have shown how one dependency can stall many organisations at once. AI brings the same exposure into investment workflows.
AI is not the only way into most investment offices’ systems today. But as teams use it more, they move up a dependency curve: from assistance with individual tasks to reliance on automation across connected workflows. An interruption that once delayed a summary could eventually disrupt research, approvals and follow-up actions.
Business continuity planning (BCP) and disaster recovery (DR) need to keep pace. A provider’s ability to restore service is only part of the test. The office also needs to continue critical work, recover interrupted actions and reconcile decisions made during the disruption.
A pilot shows how a tool performs when it is available. A continuity exercise helps the office see how it would carry on when it is not.
1. Preserve access to investment research without AI
A conversational interface can make years of research easier to use. As these interfaces improve, more work could move behind a single point of access.
The dependency develops gradually. Teams become accustomed to asking for information rather than navigating the underlying systems. Direct access may still exist, but fewer people know how to use it.
An unavailable AI interface could then make otherwise accessible information difficult to retrieve when it matters.
Investment offices should preserve independent access to source documents, review status and approved decisions. Employees should know how to use that route. Where AI supports an independently accessible system of record, the office can lose speed while retaining its ability to operate.
2. Maintain an audit trail beyond the AI conversation
A saved conversation may show the question and answer without establishing which document versions were reviewed, what the analyst changed or who approved the result.
When AI takes action, the record should show how far each step got: drafted or sent. After an outage, the team then knows what to redo and what to leave alone.
An AI interface can support strong auditability, but saving the conversation alone does not establish it. The diligence record should connect material prompts, sources, document versions, model details and outputs to reviewer judgment and final approval.
That evidence needs to remain accessible when the AI is unavailable. Another reviewer should be able to reconstruct the decision and establish which actions actually completed.
3. Assess shared LLM dependencies across vendors
An office may use different vendors for diligence, research and reporting, yet all three can rely on the same large language model (LLM) provider. That is concentration risk, whatever the vendor names suggest.
Know which model providers sit behind each application. Then ask each vendor three questions: can it move a critical workflow to another provider during an outage, has that alternative been tested for the workflow, and does switching need manual steps?
A list of supported models does not by itself show that a fallback has been tested. Also check what the backup models share: the same routing layer, or the same cloud AI service.
Teams that build their own AI workflows on LLMs, or set up workflows with agent tools such as Cowork, carry more of this responsibility themselves. Business continuity and disaster recovery planning should extend to the AI infrastructure too, including the model provider, the hosting, and the connections to internal data. Workflows set up by individual teams also need a named owner and a documented fallback.
Regulatory frameworks such as the EU’s DORA and the UK’s operational resilience rules already expect firms to map critical third-party dependencies. AI services fit naturally into that exercise.
4. Test alternative AI interfaces and data access
An alternative interface can help preserve access when a vendor’s embedded AI is unavailable. Where supported, the Model Context Protocol (MCP) can allow another approved enterprise AI application to connect to the underlying system of record.
Authorized data access through Snowflake or Databricks can provide another route for analysis and reporting. The available information depends on the integration, its scope and its permissions.
These alternatives need to be configured and tested before an incident. They must preserve permissions, source attribution and appropriate activity records.
They also need to avoid the failed dependency. An MCP connection may still rely on the unavailable platform. A live data connection may depend on the same provider, while a separate copy may be hours old. Either may exclude attachments, reviewer comments or approval history.
The office should know what each route contains, how current it is and which tasks it can support. Access to data does not automatically recreate the workflow around it.
5. Plan for unreliable AI, not only unavailable AI
Outages are the easier failure to spot. AI can also stay available while becoming less reliable: a model update that changes outputs, a connector that returns stale data, or a retrieval gap that omits a key document.
Because nothing visibly stops, work can carry on using flawed outputs. Define what gets checked routinely: sources cited, data freshness, model or version changes, and spot checks against source documents for critical conclusions.
Decide in advance who can declare AI outputs untrusted, what that triggers, and which decisions made in the interim need re-review.
6. Practice AI continuity and recovery
A continuity plan is easier to trust once the team has tried it. A walk-through tends to bring small gaps to light: access may be missing, key documents may sit outside the backup, or the manual approval route may rely on the same email system that is down.
A day-without-AI exercise can expose these gaps. Ask the team to prepare for a manager meeting, retrieve the evidence behind an investment recommendation and complete a time-sensitive review using the approved fallback.
The exercise should also test who makes the decisions. Someone must have authority to pause an agent, activate an alternative and approve the return to normal operations. Include the scenario from section 5, where AI stays available but cannot be trusted.
Recovery deserves equal attention. Tasks may be partially completed, approvals recorded elsewhere and automated requests still queued. Without reconciliation, restarting can create duplicate actions and conflicting records.
Assign an incident owner and backup, track work completed during the interruption and establish what is safe to resume. Repeat the exercise as AI takes on more responsibilities. A fallback tested for document summarization may be insufficient for a workflow that also updates records or initiates actions.
7. Build continuity into AI adoption
Start with three essential workflows. For each, identify the accountable person, required information, longest acceptable interruption and approved fallback.
Build those requirements into vendor discussions. Ask vendors to demonstrate what still works when their primary model is unavailable and what changes if the underlying platform goes down.
AI can help investment offices work faster and use their knowledge more effectively. As reliance grows, the office should retain the ability to retrieve evidence, exercise judgment and complete essential work through an interruption.
Investment offices already look closely at portfolios whose diversification rests on a single counterparty. The same lens is useful for the technology the office relies on.
8. AI business continuity checklist
| Area | Practical test | Action |
|---|---|---|
| Critical workflows | Does each of your top three workflows have an owner, a maximum downtime and a fallback? | Define them. Build them into vendor reviews. |
| Access without AI | Can an analyst pull the latest manager review, documents and approved recommendation with AI off? | Keep direct access to records. Train the team to use it. |
| Audit trail | Can another reviewer reconstruct a decision and see which actions completed? | Log sources, versions, material prompts, edits, approvals and action status outside the AI tool. |
| Vendor concentration | Which vendors share an LLM provider, routing layer or cloud AI service? | Map dependencies. Ask for evidence of tested failover. |
| Alternative access | Can MCP or a data-warehouse route supply the records, and how current are they? | Test before an incident. Document scope, freshness and permissions. |
| Output reliability | Who can declare AI outputs untrusted, and what happens next? | Set routine checks and a re-review rule. |
| Drills and recovery | Can the team finish a time-sensitive review without AI, and restart safely after? | Run a day-without-AI drill. Name an incident owner. Reconcile before resuming automation. |
Planning a day-without-AI exercise for your diligence workflow? We’re happy to compare notes on what to include.