AI workflow automation, with people still in charge.
AI is most useful as one step inside a workflow you already run: reading an email, extracting fields from a document, summarising research, drafting a reply. We build those workflows end to end — with approvals for anything consequential, limits on spending and a record of what the AI did and why.
If the work is moving data between systems with fixed rules, it doesn’t need AI — see CRM and system integrations.
Work that’s a good candidate for AI.
- Someone reads every incoming email or form to decide where it goes.
- Information is copied out of PDFs, invoices or documents into another system.
- Research or qualification is done by hand: reading pages, comparing options, writing a summary.
- The same kind of reply or report is drafted from scratch every week.
- People already paste things into ChatGPT, but there’s no process, record or control around it.
What we build.
AI steps in existing workflows
Classifying, extracting, summarising and drafting — connected to the inbox, form, CRM or database where the work already happens.
Approval gates
Anything that sends a message, spends money or changes a customer record waits for a person to approve it.
Spending limits
Each automation runs under a budget cap, so a loop or a spike in volume can’t produce a surprise bill.
Evidence and logs
Every run records its inputs, outputs and sources, so a result can be checked and a mistake traced.
Scheduled and triggered jobs
Automations that run daily, or the moment a new item arrives — with retries and alerts when something fails.
Measured rollout
We test on real samples first, check the accuracy with you, and widen the scope only when it’s working.
Is this a fit?
A good fit when
- The task is repetitive, text-heavy and happens often enough to matter.
- You can describe what a good result looks like, and check a sample of them.
- A short human review step is acceptable for consequential outputs.
- The data lives in systems we can connect to, such as email, forms, a CRM or a database.
Probably not a fit when
- Fully autonomous decisions with legal, financial or safety consequences and no review.
- A task that happens a few times a month — doing it by hand is cheaper.
- A process that’s unclear or changes every week; it needs defining before it can be automated.
- Mass outreach or messaging people who haven’t asked to hear from you.
Not sure which side you’re on? Describe the situation and we’ll tell you honestly — including when the answer is not to build anything.
Examples.
Each example is labelled. Production software built for Satori, where our founder is CTO and partner. Internal tools we built to run our own business. Internal experiments we’re still trialling. And illustrations of the kind of work, which are not client projects.
Internal experiment
Buying-signal analysis
A workflow we’re trialling: a scheduled job reads public job posts each day and uses AI to flag those that describe a real software or automation need. It runs under a monthly spending cap, and a person decides what to act on.
Internal Raul Ventures tool
Opportunity research with sources
Our research workflow uses AI to gather and summarise evidence about business opportunities, keeps the source of every finding, and records decisions only after a person approves them.
Illustrative
Supplier invoices into accounting
Incoming invoices are read, the key fields extracted and checked against the purchase order, then queued in the accounting system for approval.
Illustrative
Support inbox triage
Customer emails are categorised, urgent ones flagged, and a draft reply prepared for a person to edit and send.
The rules we run our own AI under.
Our own research and buying-signal workflows follow the same rules we build for clients: a budget cap on every job, a person approving anything consequential, and a record of every source the AI relied on. They’re internal tools, not client results.
How a project runs.
Pick the workflow
We find the one task where automation saves the most time, and define what a good output looks like.
Test on real samples
Before building anything permanent, we run the AI step on a sample of your real data and review the results with you.
Build it properly
Connected to your systems, with approvals, spending limits, logs and alerts.
Measure and widen
We check accuracy and time saved in use, then extend to the next step — or stop if it isn’t paying off.
Questions
Related:
Which AI models do you use?
Whichever fits the task and the data. We use commercial model APIs and choose by accuracy, cost and how the provider handles your data — not by brand.
Is our data used to train the AI?
We review each provider’s data-use terms for your case before using it, and use API services rather than consumer chat apps. Sensitive data can be excluded or masked before it reaches a model.
What does it cost to run?
AI usage is billed per use. We estimate the running cost on your real volumes before building, and every automation has a spending cap you control.
What happens when the AI gets it wrong?
It will, sometimes. That’s why consequential outputs go through a person, every run is logged with its inputs and sources, and we measure accuracy on samples before widening the scope.
Can you add AI to the tools we already use?
Usually, yes — through the tool’s API, its webhooks or its inbox. If it has no way in, we’ll say so.
Which task does your team repeat every day?
Tell us what arrives, what someone does with it and where the result goes. We’ll tell you whether AI helps — and what it would take to do it safely.
or write to info@raulsoftwareservices.com, or message us on WhatsApp