The AI conversation in fund administration has moved quickly from whether to when. Vendor decks promise transformed operations, and the underlying technology is genuinely capable. Yet the evidence from early adopters is sobering: most enterprise AI projects fail, and they rarely fail because the model was not good enough.
MIT's 2025 State of AI in Business research found that around 95 per cent of generative AI pilots deliver no measurable impact on profit and loss. The researchers attributed the failures not to model capability but to poor workflow integration, weak data readiness, and the absence of a defined outcome. UK financial services tell a similar story from the other direction. The Bank of England and FCA's most recent survey found that 75 per cent of firms already use AI, and that the constraints they report are dominated by data quality and governance rather than the technology itself.
For a regulated fund administrator, the lesson is direct. Readiness is organisational, not technological. In regulated firms, most AI project failures are control failures and data failures wearing a technology costume. So before you evaluate any platform, evaluate your firm. The checklist below covers six dimensions that determine whether an AI deployment lands as a controlled operational improvement or becomes an expensive pilot that never leaves the sandbox.
1. Data quality: do you have a golden source?
AI amplifies whatever data it is given. Feed it a clean, authoritative record of your entities, clients, and vendors and it will prepare work quickly and accurately. Feed it three conflicting spellings of the same client across three systems and it will industrialise the confusion.
Ask yourselves:
- Is there a single golden source for entity and client data? One system that is definitively right about names, structures, directors, bank details, and fee arrangements, or does the truth live in a spreadsheet on someone's desktop?
- Who owns data quality? A named individual with the authority to fix it, not a shared inbox.
- How fast do changes propagate? When a director resigns or a bank mandate changes, how long before every system reflects it?
- Could you export a complete, accurate list of your live structures today without a week of manual clean-up?
If the honest answers are uncomfortable, that is not a reason to abandon AI. It is your first work package, and it pays for itself whether or not you ever buy a platform.
2. Process definition: can you write it down?
An AI platform automates the process you actually have, not the one described in the procedures manual. In most firms those are different processes, and the gap between them is where deployments quietly fail.
- Can you write the workflow down end to end? From the trigger (an invoice arrives, a client requests a payment) to the final ledger entry, including every handoff.
- Who approves what, and at which thresholds? If approval authority is "ask whoever is around", it cannot be configured into a system.
- Where do exceptions go? Every workflow has a path for the case that does not fit. If that path is undocumented, it is invisible to any platform.
- Do two experienced administrators describe the process the same way? If not, you have two processes, and you need to choose one before automating either.
"A process that cannot be written down cannot be safely automated. The distance between the documented procedure and daily practice is where AI projects go to die."
3. The control model: what must never be automated?
Automation should strengthen a control environment, not thin it out. Before any vendor conversation, map your controls and decide, deliberately, which of them are load-bearing.
- Where are your four-eyes points? List every step that currently requires a second person, and confirm each one survives automation intact rather than collapsing into a single click.
- Is segregation of duties enforced or merely recorded? The person who prepares a payment must not be the person who releases it. A platform should make the breach impossible, not just log it.
- What is on your never-automate list? Payment release, client acceptance, suspicious activity decisions, and regulatory notifications are common entries. Write the list before you see a demo, so the vendor's defaults do not write it for you.
- Can your controls be tested? If an auditor asked you to demonstrate a four-eyes control firing, could you trigger it on demand and show the evidence?
A useful discipline: treat every control as non-waivable by default, and require a documented decision to relax one. Firms that discover their controls only exist on paper tend to discover it at the worst possible moment.
4. Decision governance: who is accountable for each output?
This is the dimension regulated firms most often skip, and the one supervisors increasingly ask about. The emerging expectation, visible in the Bank of England and FCA's survey work and in board-level guidance across jurisdictions, is that firms can show how their use of AI is governed: who is accountable for each AI-assisted output, and how the human decision is evidenced.
- For every AI-assisted output, is a named individual accountable? Not a team, not "the system". A person whose name appears in the record.
- How is the human decision evidenced? Who reviewed the output, what they saw, what they changed, and when they approved it. An email trail is not an evidence model.
- Does your governance survive staff turnover? If the one person who understands the AI workflow leaves, does accountability transfer cleanly?
Regulatory note: under Article 38 of the Data Protection (Jersey) Law 2018, individuals have the right not to be subject to decisions with legal or similarly significant effect based solely on automated processing. In practice this means a person, not a machine, makes the decision, and your records must be able to prove that the human involvement was real rather than a rubber stamp. Equivalent provisions exist under the UK and EU GDPR.
This is the design principle behind CoreAdmin: the AI prepares the work, a named officer makes the decision, and the platform records both sides of that exchange automatically. Whatever platform you choose, insist on the same shape.
5. People: champions, training, and honest framing
The MIT research is blunt on this point: the pilots that succeed are pulled by the people who do the work, not pushed by an innovation function. Technology adoption in a fund administration team is a human problem long before it is a technical one.
- Do you have champions inside operations? At least one respected administrator who wants this to work and will invest time in shaping it.
- Is training time budgeted? Real hours in real diaries, not a lunchtime webinar.
- Have you framed the change honestly? The credible framing is workload relief: the AI removes rekeying, chasing, and formatting, and leaves judgement with the team. If staff suspect the real agenda is headcount, they will quietly ensure the pilot fails.
- Is there a feedback loop? A named route for staff to flag wrong or odd AI output, with visible evidence that flags change things.
6. Vendor due diligence: six questions for any AI platform
Once the internal dimensions are in reasonable shape, the vendor conversation becomes far more productive, because you know what to demand. Put these questions to every platform, and treat vague answers as answers.
Using the checklist
Score each dimension honestly with the people who do the work, not just the management team. Gaps in the first three dimensions (data, process, controls) are sequencing questions: fix the worst of them first, then start with one narrow, well-controlled workflow rather than a firm-wide programme. Gaps in decision governance need closing regardless of your AI plans, because the accountability expectations behind them are arriving either way. Gaps in the last two shape your rollout plan and your procurement questions.
Firms that pass this checklist do not avoid AI risk entirely, but they take it with their eyes open, with controls that hold, and with evidence that stands up to inspection. That is the difference between the 5 per cent of projects that compound and the 95 per cent that stall. If you want to see what a governed deployment looks like against your own workflows, with the AI preparing, named officers deciding, and every interaction recorded, a 30-minute CoreAdmin demo is the quickest way to test the standard this checklist sets.
Sources
- MIT NANDA, The GenAI Divide: State of AI in Business 2025, as reported by Forbes and Yahoo Finance
- Bank of England and Financial Conduct Authority, Artificial intelligence in UK financial services, 2024
- Jersey Office of the Information Commissioner, rights relating to automated decision-making; Data Protection (Jersey) Law 2018