Workflow and process automation between your tools

Workflows small and mid-sized businesses usually automate

The best candidates are processes that run many times a week, follow clear rules and currently depend on someone remembering to do them:

• Order to invoice: an order confirmed in the CRM or online shop creates the sales order, the delivery and the invoice in Xero, Sage or Business Central.
• Job sheets to billing: a completed job in your field app becomes an invoice with the right labour and parts, and the customer gets the report.
• Supplier set-up: a form collects bank details and documents, someone approves them and the supplier is created in the accounts system.
• Starters and leavers: accounts, licences and equipment requests triggered from HR, and removed again when someone leaves.
• Credit control: reminders by ageing band, with stop-on-account alerts to sales.
• MI reports: figures pulled from several systems every Monday morning instead of being pasted into a spreadsheet.

Where AI adds value in a workflow and where it does not

Many processes pitched as an ‘AI project’ are mostly integration between tools, with one or two AI steps. Treating them that way makes them cheaper and more predictable. Rules are better than AI when the logic is clear: if an invoice is over a set amount, it needs a director's approval. AI earns its place where a person currently has to read something:

• Classifying free text: which department an enquiry belongs to, whether a review is a complaint.
• Extracting data from documents and emails, covered in depth on our AI document extraction page.
• Summarising long threads or call notes before they go into the CRM.
• Drafting a first version of a document that a person then checks.

What makes an automation reliable: retries, idempotency and alerts

The difference between an automation that helps and one that causes trouble is what happens when something fails. A tool is down for ten minutes, an API returns an error, a connection expires on a Friday night. This is what we build in from the start:

• Clear triggers: the workflow starts from an event (a form, a new order, an email) or at a fixed time, and you can see which.
• Retries without duplicates: if a system does not respond, the step is retried with increasing waits. Each write is idempotent, so a retry never creates the same customer twice or sends the same email twice. Some APIs support this directly; Xero's accounting API, for instance, accepts an idempotency key for exactly this purpose.
• Alerts that say something useful: if it still fails, a named person gets an alert with which step failed, on which record and what to do next.
• A queue for failures: failed runs wait in a list where they can be fixed and replayed, rather than disappearing.
• Human review where it matters: AI steps carry a confidence score; below the threshold you set, the case waits for a person to confirm it.
• A log of every run: what came in, what each step did and what came out.
• Documentation: a map of what each workflow does, so your team or any other supplier can maintain it.

Zapier, Make, n8n, Power Automate or bespoke

For many businesses an off-the-shelf automation platform is the right answer, and we will say so. The way each one charges matters as your volume grows:

Credentials, access and data

Automations hold the keys to your systems, so they are treated as such. Each integration uses its own service account with the minimum permissions it needs, credentials are kept in a secrets store rather than in the workflow, and everything sits in accounts in your company's name. When a step sends data to an AI model, we follow the approach to training and data location described on our private ChatGPT for business page. If your workflows mostly move data between finance and operations systems, our ERP and CRM integration service goes into more depth.

How we would run it and what drives cost

We start by mapping the process with the people who run it, then automate one complete workflow before moving to the next. You see a working demo every two weeks, you get one fixed price for the project, and the code and workflows are yours. The rest is on how we work. What moves the cost most:

• Number of systems involved and whether each has a usable API.
• Exceptions: the cases that do not follow the normal path take more work than the normal path.
• AI steps, and how much review they need.
• Reliability requirements: a workflow that bills customers needs more safeguards than one that sends a Slack message.
• Running costs: platform subscriptions or hosting, plus AI usage where it applies. Market references are in our guide to the cost of implementing AI in a business.