Key ideas
- Choose a recurring workflow with a clear owner and measurable friction.
- Give models, business rules and people distinct responsibilities.
- Measure completed work, corrections and operating cost against a baseline.
Choose the work you want to change
Start by following one case through the organization. Where does it arrive? Which information is copied between tools? Who has to ask for clarification? What allows the case to leave the queue? These questions reveal the handoffs that an isolated model demonstration can miss.
Useful candidates include assembling an application pack, preparing an investor-service response, investigating a payment discrepancy or collecting evidence for a claim. Prefer an initial scope where the team can recognize a correct result and explain the exceptions.
Record the current process before changing it. The baseline should include the time spent doing the work, time spent waiting, common corrections and cases that return for another review. A faster extraction step matters only if it helps the wider process.
Assign the right work to each part of the system
A language model can interpret varied documents or correspondence. Calculation, permissions and required-field checks can remain explicit in application logic. People provide domain judgment, resolve ambiguity and authorize the decisions assigned to them.
| Part | Useful responsibility | Result to inspect |
|---|---|---|
| AI model | Extract, classify, retrieve or prepare a draft | Facts and text linked to their sources |
| Application rules | Validate fields, calculate values and enforce permissions | A reproducible check and its outcome |
| Responsible team | Resolve exceptions and make designated decisions | A review or approval attached to the case |
Anthropic's architectural distinction is useful here: workflows follow predefined paths, while agents can choose more of their own steps. Our starting point is a defined process, adding model-directed choices where the task benefits from that flexibility.
A practical example: preparing an application
- 01
Receive
Create a case and preserve the incoming files and correspondence.
- 02
Interpret
Classify the material and extract the relevant application details with source references.
- 03
Validate
Check completeness, compare repeated information and identify conflicts.
- 04
Clarify
Prepare a request for missing information, with review before sending where the process requires it.
- 05
Review
Give the responsible team the application, supporting evidence and unresolved questions.
- 06
Record
Write the accepted result to the authoritative system and retain the case history.
The first release might cover one applicant type and a small document set. The team should be able to correct a field, reject a prepared response and reopen a case when new evidence arrives. Those paths are part of the product experience.
Connect the result to existing operations
Decide which system owns the case and which changes the automation is permitted to make. A model-generated summary becomes operationally useful when it reaches the right queue, includes the evidence and gives the next person an understandable action.
Define what the user sees when a connected service is slow, unavailable or returns an uncertain result. Preserve a pending state when completion cannot be established. The integration design should explain how the team investigates and resumes that work.
The processing environment also follows the information involved. Agree which documents a model may receive, how long outputs are retained and who can access them. A private endpoint is one design choice; the actual data path and operating responsibilities still need to be established.
Measure the change in completed work
Evaluate the proposed workflow on representative examples, including incomplete files, conflicting values and requests outside the expected path. Keep the same business definition of completion when comparing it with the baseline.
- Quality: source accuracy, reviewer corrections and inappropriate next actions.
- Throughput: cases completed, manual touch time and elapsed time to the agreed endpoint.
- Exceptions: unresolved items, repeated follow-ups and cases requiring escalation.
- Economics: model and infrastructure costs together with review, integration and ongoing support effort.
An extraction accuracy score alone will not describe the operating result. A workflow can produce plausible text while creating more review work. Acceptance should reflect the quality and effort required to finish the case.
Define a first delivery the team can assess
Choose one process owner, one case type, representative evidence and an agreed output. Build the full path to that output, test the exceptions and review the results with the people who do the work.
At Sfxdx, this is how we scope financial AI delivery: understand the operational task, define the system boundaries, build a working pilot and use the evaluation to decide the next step. Expansion can then follow the evidence and the client's priorities.
Sources and further reading
Primary references checked on 15 November 2026. The implementation examples and selection guidance are Sfxdx's engineering analysis.
Explore the workflow
The linked capability example uses synthetic data and local rules to show the interaction pattern.
Explore the onboarding example ↗