Almost every business that comes to us has already had a go at this. Someone on the team is a bit technical, they’ve had a play with Zapier or Make, and there’s a half-finished automation sitting in an account that nobody quite trusts.

That’s not a failure. It’s a useful signal — it means the appetite is there. What’s usually missing is a sense of which automation to build first, and that choice matters more than the tooling does.

Here’s what we’d build, in order, for a typical UK SME.

The one everybody tries first (and why it disappoints)

Automated social media posting.

It’s the obvious starting point: visible, low-risk, easy to demo. It’s also the one that delivers the least.

The problem is that content isn’t the bottleneck in most businesses — conversion is. You can triple your LinkedIn output and change nothing about revenue, because the enquiries you already get are the ones going cold. Automating the top of the funnel while the middle leaks is a very efficient way to feel productive without being productive.

Build it eventually if you like. Don’t build it first.

Automation 1: The enquiry triage

What it does: Every enquiry — email, web form, WhatsApp, voicemail — lands in one place, gets read, gets classified, gets a first response, and gets routed to the right person with the right context attached.

Why it’s first: Because response time is the single most leveraged number in most SMEs. The research on this is unambiguous and slightly brutal: the odds of qualifying a lead fall off a cliff after the first hour. Not the first day. The first hour. Most UK small businesses are measuring their response time in working days without realising it, because the enquiry arrived at 4:50pm on a Friday and the person who handles it was on site.

What “AI” adds here, concretely:

A rules-based system can send an autoresponder. That’s not triage — a customer knows an autoresponder when they see one, and it buys you nothing.

A language model can read an enquiry properly. It can tell the difference between:

• “Do you do commercial work?” — a qualification question, answerable instantly, no human needed
• “We had a leak at the Croydon site again, third time this year” — an existing customer, escalating, needs a human within the hour
• “Please add us to your supplier list” — not a lead at all, file it

Three enquiries, three completely different correct responses. The first one gets answered in ninety seconds. The second one pings the ops lead’s phone. The third one gets archived and never wastes anyone’s attention.

Realistic build: two to three weeks, including the fortnight where a human reviews every AI decision before it goes out. That review period isn’t optional and it isn’t wasted — it’s how you find out what the system gets wrong while the stakes are still low.

Automation 2: The document-to-data pipeline

What it does: Takes the PDFs, scans, and photos that arrive in your business and turns them into structured data in the system where you actually need them.

Why it’s second: Because it’s the most reliably underestimated cost in any business that handles paperwork. Supplier invoices. Timesheets. Delivery notes. Certificates. Tenant references. Purchase orders that arrive as a photo of a printout taken at an angle in poor light.

Someone, somewhere in your business, is re-typing these. They’ve been doing it so long that nobody counts it any more.

What’s changed: Optical character recognition has existed for decades and was always slightly rubbish, because reading the characters was never the hard part — understanding them was. A scanner could see “£1,240.00” but couldn’t tell you whether that was the net, the gross, or the previous balance carried forward.

Modern models read documents the way a person does: contextually. They can pull the right total off an invoice that’s laid out in a format they’ve never seen before, flag when the VAT doesn’t reconcile, and match the line items against the purchase order they belong to.

The critical design decision: confidence thresholds. Anything the system is sure about flows straight through. Anything ambiguous goes to a human queue with the relevant bit highlighted. You are not aiming for 100% automation — you’re aiming to shrink the human’s job from “process 200 invoices” to “check the 12 that looked odd.”

That distinction is the whole game, and it’s where most DIY attempts come unstuck. They aim for perfect, don’t get it, and abandon the project. Aim for 85% with a clean exception path and you’ll bank the value immediately.

Automation 3: The proactive follow-up

What it does: Watches your pipeline, notices what’s gone quiet, and does something about it before it dies.

Why it’s third: It depends on the first two being in place. Once your enquiries are properly captured and your data is actually in the system rather than in someone’s inbox, you can finally see the pipeline clearly — and once you can see it, the gaps are embarrassing.

Every business has them. The quote sent three weeks ago that nobody chased. The customer who’s ordered every quarter for four years and hasn’t ordered this quarter. The proposal that got a “let me check with my colleague” and then silence.

These aren’t lost deals. They’re unattended deals, and the difference is one email.

What good looks like: not a blast. The system flags the situation and drafts a follow-up that references the actual conversation — the specific job, the specific concern raised, the date they said they’d get back to you. A human glances at it, adjusts a line, and sends. Thirty seconds of attention on a deal that would otherwise have quietly evaporated.

A caution: this is the automation with the highest potential to damage relationships if you get it wrong. Nothing says “you are a row in a database” quite like a follow-up that gets the details wrong or fires on a customer who cancelled last week. Keep a human in the loop on send, at least for the first few months. The time saved is in the drafting and the noticing, not in the clicking.

What these three have in common

Each one sits at a junction — a point where information has to move between a human and a system, or between two systems, and currently a person is doing the carrying.

That’s the pattern to look for in your own business. Not “what tasks are boring” but “where does information get stuck, and who’s unsticking it?” That person is your automation brief.

Sonwa AI Lab builds and runs these systems for UK businesses. We start with an audit, not a sales pitch — and if the honest answer is that automation won’t pay for itself in your situation, we’ll tell you that. Book a call.