Webbit

Where AI Belongs in a Workflow, and Where Plain Automation Wins

Most working AI is a single step inside a larger process. A simple test for which steps need AI, which need plain rules and which need a person.

Many owners ask for "AI" before they know which part of their work needs it. Sometimes the answer is a lot. Sometimes it is none, and a simple rule would do the job better. The skill is knowing which is which, step by step.

What one automation consultancy found

Flow Digital, a Brooklyn-based automation consultancy, published an analysis on 12 November 2026 and the Manila Times carried it the next day. It reviewed six months of its own sales and client-call transcripts and its delivered work. It is a press release from a company with something to sell, and the numbers describe its own clients, so read it as one practitioner's view.

The findings are still useful. AI came up in 49% of conversations, but only 20% of organisations asked for help with a specific AI use case, and few of those had a solution deployed. Of the AI work Flow Digital delivered, 47% came from clients who first described an operational problem and did not ask for AI. About half of its live AI work extracted information from documents or emails, and 61% of the AI components ran as one step inside a Zapier workflow. In some projects the team removed the AI after review. One matching step went back to ordinary code because the rules were stable, and one failed AI calculation became a formula.

AI usually works as one part of a process, so choosing the right part is most of the job.

A three-question test for each step

Take any process and list its steps. Then ask three questions about each step.

Can the rule be written down so that two people would always get the same answer? If yes, use a plain rule, a formula or a filter.
Is the input messy, such as free-text emails, scanned invoices or voice calls? If yes, AI can read it and turn it into structured data.
Does a wrong answer carry real cost, such as money, legal exposure or a customer's trust? If yes, put a person in the loop to approve it.

Step	Best tool	Why
Route an order over a set value to a manager	Rule	The condition is exact and easy to check
Pull the total and supplier from a supplier invoice PDF	AI extraction	Layouts vary and the text is messy
Classify an inbound email as sales, support or spam	AI	Meaning matters more than keywords
Calculate sales tax on an order	Formula	The answer must be exact every time
Approve a refund above a limit	Person	Money and customer relationship are involved

Why rules still win where they fit

They give the same output every time, so you can test them.
They cost almost nothing to run.
They are quick, and they do not need a model provider to be available.
When one breaks, you can see why.

Using AI for a step that a rule could handle adds cost and a new way to fail. It is common to see an AI step quietly replaced by a formula once the team understands the pattern, which matches the project reversals in the analysis above.

Where AI is worth it

Extracting fields from invoices, forms, contracts and emails.
Sorting and summarising incoming messages.
Drafting a first reply for a person to approve.
Answering questions from your own documents, with sources shown.
Understanding a phone call and writing the notes into your CRM.

Notice that most of these turn unstructured information into structured information. That is the part rules cannot do, and the part where AI saves the most typing.

How to start

Pick one process that wastes time every week and write its steps on a page.
Mark each step as rule, AI or person using the test above.
Build the rule steps first. They are cheap and quick.
Add AI only to the steps that need it, and measure its accuracy on 50 real examples before you trust it.
Keep a person on any step with real cost if it goes wrong, and review a sample each month.

A worked example: supplier invoices (illustrative)

Step	Tool	Reason
Invoice arrives by email as a PDF	Rule	A filter moves it to the invoices folder
Read supplier name, date, total and line items	AI extraction	Every supplier uses a different layout
Match the supplier against your list	Rule	An exact lookup, with a fuzzy match flagged for review
Check the total against the purchase order	Formula	Arithmetic must be exact
Approve anything over your limit	Person	Money leaves the business

Only one of the five steps needs AI. Before trusting it, run 50 real invoices through the extraction step and count the fields it gets wrong. If 47 of 50 are fully correct, that is 94%. Decide in advance whether 94% is enough, and route low-confidence invoices to a person.

Pay for AI only where it earns its place

Adding AI to the wrong step costs money and creates new ways to fail. Webbit maps your process step by step and tells you where AI, plain rules or a person fits best, then builds, tests and documents the result. The first conversation is free.

Tell us about the process that wastes the most time each week.
We reply with a plain-language view of which steps could be automated and what it would involve.
You decide, with a clear scope confirmed in writing before any work starts.

Tell us about your process or see our business process automation service.

Frequently asked questions

Should I add AI to every step of a process?

No. Use AI where the input is messy or the meaning matters, and use rules where the answer can be written down exactly. Fewer AI steps usually means lower cost and fewer surprises.

How do I know if an AI step is accurate enough?

Test it on 50 or more real examples and count the errors. Decide in advance what error rate is acceptable, and keep a person reviewing anything where a mistake is costly.

Do I need Zapier or a similar tool for this?

Not necessarily. Workflow tools such as Zapier, Make and n8n are a common way to connect the steps, but the right choice depends on your systems, volume and security needs.

Home · Blog · Book a Free Consultation