What AI can actually automate in a service firm

AI reliably automates the repetitive middle of a service firm: enriching and verifying prospect data, drafting first versions of emails and documents, triaging and routing enquiries, summarising calls, and running follow-up sequences nobody remembers to send manually. It does not reliably automate judgement, relationships, pricing, or accountability. Knowing which side of that line a task sits on is worth more than any tool subscription.

I build automations for UK B2B service firms daily, so this list comes from production systems, not vendor decks. The full argument — including the failure modes and the economics — lives in AI Automation for B2B: what actually works; this piece is the practical inventory.

Which tasks can AI genuinely handle today?

The dependable wins share three properties: they are frequent, they follow rules, and a mediocre output costs little. In a typical 5–50 person firm that means:

• Data work. Finding, enriching, and verifying prospect records; deduplicating CRM entries; flagging records that have gone stale. Tedious for humans, trivial for machines.
• First drafts. Outbound emails, proposals from a template, meeting summaries, job ads. AI produces the 70% version fast; a human takes it to 100%.
• Triage and routing. Reading an inbound enquiry, classifying it, assigning an owner, and triggering the right next step within minutes rather than days.
• Follow-up. Sequenced touches on a schedule. Most firms stop at two follow-ups while deals typically need five or more; software does not get bored at touch three.
• Reporting. Pulling numbers from the CRM into a dashboard nobody has to compile by hand.

Notice what these have in common: none of them is glamorous, and every one of them leaks money when done inconsistently.

Where does AI still fail?

It fails wherever the task needs context it does not have or stakes it cannot carry. Pricing a complex engagement. Reading the politics of a renewal. Deciding whether a lead is worth a partner's time. Writing anything where a wrong claim damages the relationship. AI can assist all of these — summarise, suggest, prepare — but the moment you remove the human, quality decays quietly and you find out from a customer.

It also fails at accountability. When an automated process breaks — and all of them eventually do — a system cannot own the consequence. Someone in the firm has to. If nobody is named, the automation is a liability wearing a cost-saving costume.

A specific example of overreach: most B2B websites get too little traffic for a chatbot to earn its keep, yet chatbots are the first thing many firms ask for. I unpack that one in the chatbot your website probably doesn't need.

How do you decide what to automate first?

Use this mechanism, task by task. When a task happens more than roughly ten times a week and follows the same steps each time, then it is an automation candidate. When the candidate task can be written down as trigger → steps → output on one page, then it is buildable; if you cannot write it down, the process is not ready — fix the process first. When the cost of an occasional wrong output is low (an internal summary, a draft), then automate it fully. When a wrong output is expensive or public (anything sent to a customer, anything involving money), then automate the preparation and keep a human on the approval. When none of the above holds, then leave it alone.

Run your week's tasks through that chain and you will typically find three to five solid candidates — usually in lead handling, follow-up, and reporting. That is a system view of the business rather than a task view, which is precisely the shift described in A Systems-Thinking Guide for Founders: look for the recurring flows, not the individual annoyances.

What results should you actually expect?

Modest, compounding ones. A firm that automates enquiry routing and follow-up does not double overnight; it stops losing the leads it already paid to generate. A firm that automates list building and verification runs outbound at 25–40 emails a day per inbox without an admin drowning in spreadsheets. Hours come back — commonly several per person per week on the affected processes — and, more importantly, things stop falling through gaps.

What you should not expect: AI running your sales function unattended, or "an AI employee". Anyone promising that is selling the demo, not the system. The honest framing is that AI removes the repetitive middle so your people spend their hours on the judgement calls only they can make.

Should you build this yourself or bring someone in?

Either can work; both are covered properly in build vs buy: sales systems for service firms. The short version: buy commodity tools for commodity problems, build (or commission) systems for the processes that are specifically yours, and whoever builds it, keep the credentials, the documentation, and a named owner inside the firm. The tools are the easy part. The discipline around them is what decides whether the automation still works in a year.

---

Next step: the Growth System Audit — £450, seven days, credited against any build — maps where your growth system leaks and what to build first.