Experience in finance, audit and business control
I start with one recurring task from the finance team's current workflow. The team practises with the selected tool, learns how to check its output and records only the necessary data rules and review steps, so it can repeat the workflow itself. Depending on the task, we use Claude, ChatGPT or a local model.
Why this approach
I have 15 years of experience in finance, including audit and business control, and 5 years in data and IT. I therefore start with the existing finance process, its review points and the people who perform the work. Model choice comes afterwards. The team carries out the workflow and decides at each step what further support it needs.
Services and deliverables
Training, pilot and implementation are separate steps. For each step, we agree the task, output, team input, timing, price, permitted data and reviewer. After each step, the team decides whether to continue.
The team works on recurring tasks from reporting, analysis, planning and control using approved tools. Participants supply sources, write instructions, check results and repeat the workflow themselves.
We record the current process, acceptance criteria and known failure cases. The current and proposed workflows are tested on the same representative examples, including exceptions.
If the pilot meets the agreed criteria, I implement the smallest workable solution and define the permitted inputs and users, approval points, failure handling and operational owner needed for reliable use.
Bounded use cases
These are pilot candidates, not guaranteed automations. Select one recurring task with stable inputs, a defined output and a qualified reviewer.
Controls and accountability
Before choosing a product, we define the task, permitted data, valid sources, reviewer, approval step and owner after handover.
Record which data classes may enter which approved environment, who has access and which source versions the workflow may use.
Calculations remain in a controlled model or source system. AI may draft an explanation; finance reconciles every value before use.
Compare the current and proposed workflows on the same examples using measures such as factual errors, completeness, review time and total task time.
A named finance owner checks source data, calculations, references and assumptions. The model is not the approver.
Claude, ChatGPT/GPT, local
The permitted Claude, ChatGPT/GPT and local or self-hosted options are compared on the same tasks. Products, plans, APIs and deployment forms are assessed separately because their controls and operating requirements differ.
A local model is not automatically safer. Hosting, access control, logging, updates, backups and incident handling remain part of the assessment.Project approach
A pilot records only what is needed to compare the current and proposed workflows: the baseline, representative examples, success criteria, data rules and reviewer.
Document inputs, output, systems, current time spent, common corrections, process owner and approval step.
The team practices approved use while representative examples, exceptions and acceptance criteria are prepared.
Compare accuracy, completeness, review effort, total task time and cost with the current process, and record errors and limitations.
If the criteria are met, document access, review steps, exceptions, version changes and the maintenance owner before handover.
Frequently asked questions
No. Training can be the end of the engagement, or it can be followed by a defined pilot and implementation if the use case meets the agreed criteria.
Only in an approved environment and for an approved purpose. Before real data is used, we record the data classes, product, plan or API, relevant retention and training settings, access rights and connected services. Confidential or personal data is not entered into consumer accounts or other unapproved tools. Specialists are involved where internal policy requires them.
No. The permitted products and deployment options are tested against the same tasks and operating requirements. No vendor is selected in advance.
No. A named employee reviews and approves output used in financial reporting, controls or decisions. The model is not the approver.
The free proof of concept described in my project approach is a limited demonstration without confidential data. A scoped pilot uses agreed examples, acceptance criteria and a documented test report, and may be paid work.
No. We begin with one team and one recurring task. We add only the data rules, controls and documentation needed for safe, repeatable use. Further support remains available, but the goal is for the team to own the workflow.
Introductory meeting
Bring one recurring task: its input, desired output, frequency, review step and systems involved. We determine whether AI fits the task and whether training, a pilot or implementation is the right first step. Describe the workflow without sharing confidential data.