AI-assisted workflow automation
When document-heavy admin needs software with judgement built in.
I build bounded workflow tools that use AI for extraction, classification, triage or draft preparation while keeping human review, audit trail and business ownership in the design.
Problem diagnosis
The useful work is usually not an autonomous agent. It is a controlled workflow.
- Staff read the same kinds of documents, emails or forms and copy a few important facts into another system.
- Simple automation fails because inputs are messy, exceptions matter and somebody needs to approve the result.
- AI experiments look promising but are not safe enough to become part of the business process.
The risk is creating a black box. The opportunity is a tool that reduces repetitive admin while leaving review, correction and accountability visible.
Responsibility
AI is one part of the system, not the product being sold.
I design the workflow, review points, data boundaries, validation rules and handover. If AI is useful, it sits inside that structure; if it is not suitable for a step, it stays out.
What I would build
The first useful version would be concrete.
- 01
A controlled intake workflow
Emails, forms, PDFs or documents enter a process with clear storage, review and routing rules.
- 02
Extraction or classification suggestions
AI can propose fields, categories, summaries or next actions without silently making final decisions.
- 03
Validation and exception flags
Low-confidence, inconsistent or high-risk outputs are sent to a person rather than treated as complete.
- 04
Human approval screens
People can see what was suggested, what was changed and what was accepted.
- 05
Controlled system updates
Only approved outputs are pushed into CRM, database, order, ticket or reporting systems.
Fit filter
Probably right, probably not.
Probably right when…
- The process repeats often and has a clear desired output.
- Inputs are messy, but exceptions can be described.
- Human review is acceptable and auditability matters.
Probably not me when…
- You want a chatbot, content machine or general AI strategy.
- The process is vague, rare or impossible to review safely.
- You want fully autonomous high-risk decisions.
First step
Start with diagnosis, not a big commitment.
The first conversation is about the repeated workflow, the inputs, the review points and what must never happen automatically.
£150 per hour plus VAT where applicable. Two hours minimum for new clients. No retainer, no minimum term, no lock-in. You own everything I build.
Useful things to send before a call
- Sample documents, forms or emails
- The current manual process
- Desired outputs and exception examples
- Data/privacy constraints and destination systems
Objections
Questions this page should answer before you email.
How do you stop AI from making things up?
By narrowing what it is allowed to do, validating outputs, flagging uncertainty and keeping humans in control of approval.
Will our data go into public consumer tools?
No client confidential or personal data should go into a consumer account or unapproved tool. Data boundaries are agreed before build work starts.
What if AI is not suitable?
Then it should not be used for that step. The useful product is the workflow, not forcing AI into it.
Next step
If this looks like your problem, send me the awkward bits.
The useful first email is short: what is breaking, what product or process cannot handle it, and what working would need to mean.