The useful question is not "how do we use AI?" but "which repeated task costs us the most time and has the lowest cost of error?" That framing filters out the demonstrations and points to work worth doing.

Worth automating first

1. Enquiry acknowledgement and routing

When a form is submitted, send an immediate acknowledgement that sets expectations, create a record in the CRM, assign an owner based on rules, and raise an alert if nobody responds within an agreed window. This is mostly conventional logic, with a language model only if you want a drafted reply. The measurable outcome is response time and the percentage of enquiries answered within target.

2. Lead qualification triage

A model can read an enquiry and extract company size, service interest, location and urgency, then flag obvious mismatches for a polite decline or route to the right person. Keep the human decision on whether to pursue; automate the reading and sorting. Include a confidence threshold: low-confidence items go to a person rather than being guessed at.

3. Internal summaries and first drafts

Summarising a long email thread before a call, drafting a first pass at a proposal section, or converting call notes into action items saves genuine time. These are drafts with a human editor, and that is the point — the value is in the time saved, not autonomy.

Where it usually goes wrong

  • Pricing decisions. Rule-based calculation with human sign-off on anything binding.
  • Anything regulatory or medical. Needs qualified review, always.
  • Irreversible actions such as refunds, cancellations or account changes — require explicit confirmation.
  • Customer communication with no review in a sensitive industry, where one wrong answer has lasting consequences.
  • Processes that are inconsistent in the first place: automation needs a defined input, and will standardise your inconsistency.

The guardrails that make it reliable

  1. Defined input and output: exact fields, allowed values, length limits.
  2. Grounded source content: approved documents only, not open web retrieval for customer-facing answers.
  3. Review step for low confidence, high value or irreversible actions.
  4. Failure handling: what happens when the API is unavailable, times out or returns nonsense.
  5. Logging: store inputs, outputs and decisions so issues can be diagnosed.
  6. Data minimisation: send only the fields the task requires, and know where they go.

The boring test

Before automating, write the process as a numbered list a new employee could follow. If you cannot, the automation is not the problem to solve.

What it costs to run

Automation platforms, CRM tiers and model usage all carry subscription costs that scale with volume. Estimate them before building: an automation that saves four hours a month but adds two subscriptions and ongoing maintenance may be neutral at best. We usually recommend starting with one workflow, measuring it for a month, and expanding only if it holds.

A sensible starting scope

One workflow, one owner, one measurable outcome, reviewed after thirty days. That is enough to prove the value without committing to a platform migration or a transformation programme. Most small businesses get the majority of the available benefit from acknowledgement, routing and qualification — none of which requires anything exotic.