The safest first AI project in financial services is rarely an autonomous decision-maker. It is usually a controlled assistant that prepares information, records its work and hands an accountable decision to a person.
01
The pressure to use AI is no longer theoretical
Financial-services firms are being asked to move faster while handling more information, more customer expectations and more scrutiny. Staff still open documents, copy details between systems, search long case histories and prepare the same kinds of summaries every day. AI appears to offer an obvious answer.
Adoption is already well under way. In the Bank of England and FCA 2024 survey, 75% of responding firms said they were using AI and another 10% planned to use it within three years. Operations and IT accounted for the largest share of reported use cases.
The same survey gives a useful warning against rushing. A third of AI use cases involved third-party implementations, 46% of firms reported only a partial understanding of the AI technologies they used, and four of the five highest perceived risks were related to data.
The choice is therefore not between adopting AI and standing still. It is between introducing it as a controlled part of a real operation or allowing experiments, subscriptions and copied customer data to spread without clear ownership.
02
Start where AI can prepare the work
Imagine an insurance or lending operations team receiving customer emails, application forms, supporting evidence and broker documents. A person identifies the case, works out what each file contains, checks whether information is missing, reads the history and prepares the next action.
Some of that work needs experience and judgement. Some of it is repeatable preparation. AI can classify an incoming document, extract named fields, summarise a case history, compare the evidence with a checklist and draft a request for missing information. The output can then sit inside the existing case workflow for a member of staff to review.
This is a stronger starting point than asking a chatbot to run the process. The business can see the input, proposed output and final human decision. It can measure whether preparation becomes faster without pretending that a fluent answer is automatically a correct one.
The running example also keeps the project narrow. The goal is not to make the whole firm AI enabled. It is to reduce one visible piece of repeated work while protecting customer outcomes and the operation around it.
| Workflow task | Possible AI assistance | Human or system control |
|---|---|---|
| Incoming documents | Classify the document and extract agreed fields | Validate required fields, retain the original and route uncertain results for review |
| Case preparation | Summarise history, evidence and unresolved questions | Reviewer checks the source record before relying on the summary |
| Customer communication | Draft a clear request or update from approved case data | Authorised staff approve the message before it is sent |
| Knowledge retrieval | Find relevant wording from approved policies and procedures | Show the source and prevent answers from unapproved material |
| Operational reporting | Group themes and explain changes in workload or exceptions | Managers can trace figures back to the underlying records |
03
Keep accountable decisions with people
A human-review label is not enough by itself. If staff approve almost every recommendation because the queue is too large or the reasoning is hidden, the person has become a rubber stamp rather than a meaningful control.
The boundary should be explicit. AI may prepare evidence for a lending, underwriting, claims, complaint or suitability decision. It should not quietly become the person making that decision unless the firm has deliberately assessed the legal, regulatory, data and customer consequences of automated decision-making.
A useful review screen should show the source information, the AI output, uncertainty or exceptions and the action being requested. The reviewer needs enough time and authority to disagree. Important corrections should be captured so repeated failure patterns can be found.
The FCA has said it plans to rely on existing frameworks rather than introduce a separate set of AI rules. Consumer Duty, senior-manager accountability and expectations around governance and controls therefore remain relevant. Calling a process an AI pilot does not put it outside the responsibilities the firm already has.
04
Use ordinary automation when the rules are ordinary
Not every awkward process needs AI. If an approved application must always create the same document, update the same record and send the same notification, normal software rules or an API integration will usually be easier to test and explain.
AI earns its place when the input contains variation that ordinary rules handle badly. Free-text emails, differently formatted documents, long case notes and natural-language questions are sensible examples. Even then, the uncertain step can be surrounded by deterministic checks.
For example, AI might extract a policy number and document type. The application can then check the format, confirm that the case exists and reject an unsupported category using ordinary code. The combination is often safer than asking the model to complete the entire workflow.
This distinction also keeps costs under control. A reliable API connection may solve duplicate entry without introducing model usage, prompt behaviour and an extra third-party dependency. The business problem should choose the tool, not the other way round.
05
Put the controls around the whole workflow
Model accuracy matters, but it is only one control. A useful financial-services implementation also needs a named business owner, an agreed purpose, approved data access, documented instructions, output logging, exception handling and a way to stop or roll back the feature.
Data deserves particular attention. The firm should know which customer, commercial or special-category information reaches the provider, whether it is retained, where it is processed and whether it can be used to improve somebody else's model. Least-privilege access and data minimisation matter more than connecting every available record for convenience.
Third-party due diligence should cover more than a security questionnaire. The team needs to understand service availability, model changes, contractual terms, incident routes, audit evidence and what happens if the provider is replaced. A convenient managed model is still part of the firm's supply chain.
The ICO recommends a proportionate, risk-based approach to AI and data protection. That includes assessing effects on people, applying data protection by design and reconsidering the organisation's risk appetite where AI introduces new or harder-to-manage risks.
06
Test the workflow before asking customers to trust it
A controlled pilot begins with examples, not a live switch. Gather a representative set of past cases using anonymised, synthetic or appropriately governed data. Agree what a good output looks like and include difficult, incomplete and unusual cases rather than testing only the neat ones.
Run the AI step in shadow mode first. It can classify, extract or draft while staff continue using the existing process. Compare its output with the decisions and corrections made by experienced reviewers. This reveals where the model is useful, where instructions need improvement and where the task should remain manual.
Set acceptance measures before seeing the result. Depending on the workflow, they might include extraction accuracy, missing-field detection, time saved per case, correction rate, false-confidence rate and the number of cases sent for human review. Average accuracy can hide a serious problem if the model regularly fails on one customer group or document type.
Only move into limited live use when the business owner, risk and compliance colleagues, technology team and actual users understand the boundary. Start with a small group, monitor outputs and keep a clear route back to the original process.
07
Measure whether the operation became better
An AI project can look impressive while making work slower. Staff may spend longer checking unreliable drafts, correcting extracted fields or working around a tool that does not fit the case system. Usage is not the same as value.
Measure the existing process before changing it. Record how long preparation takes, where cases wait, how often information is rekeyed, which errors create rework and how many exceptions need senior attention. The pilot then has a real baseline rather than a promise to improve productivity in general.
The outcome should include quality and customer impact as well as speed. A faster document review is not an improvement if more incomplete cases reach the next team. A quicker draft is not useful if staff rewrite it or customers receive less clear information.
If the evidence is weak, stop or narrow the project. A small failed experiment with good records is cheaper than a wide rollout held together by enthusiasm and sunk cost.
08
Move forward one accountable workflow at a time
Financial-services firms do not need to choose between reckless adoption and endless caution. They can start with work that is repetitive, measurable and valuable, then place AI inside a process with visible sources, clear ownership and a genuine human decision point.
The document-intake example is useful because it separates preparation from accountability. AI handles variation in the material. Ordinary software applies known rules. A person reviews the evidence and owns the decision. Logs and measurements show whether the change is actually helping.
That pattern can extend into other areas once it has been proven. The firm learns how to assess providers, govern data, evaluate outputs and support staff before taking on a more material use case.
DanJMills helps businesses map repeatable work, identify where AI is genuinely useful and build controlled automations around the systems they already use. If one financial-services workflow is consuming time without needing a machine to make the final decision, that is usually the right place to begin.
Useful questions
Before approving a financial-services AI pilot, confirm:
- Which stable workflow problem is the pilot solving?
- What evidence shows the current cost, delay or error rate?
- Which output will AI prepare and which decision remains with a person?
- Who owns the use case and is accountable for its outcome?
- Which data can the model access, retain and process?
- Can every output be traced back to its source record?
- What happens when the result is incomplete, uncertain or wrong?
- How will staff contest, correct or override the output?
- Which accuracy, quality and customer measures decide whether the pilot continues?
- How can the feature be paused, rolled back or moved to another provider?
