Most companies that ask us about artificial intelligence have already tried it informally. Someone on the marketing team drafts product descriptions with a chatbot, someone in finance pastes a supplier invoice into a model to pull out the totals, and nobody has written any of this down. The technology is not the hard part. The hard part is deciding which tasks deserve it, who checks the output, and what happens when the output is wrong.
Start with tasks, not tools
A useful first question is not “which AI platform should we buy” but “which piece of work do we repeat every week that nobody enjoys”. The answers are usually unglamorous: summarising long email threads before a handover, tagging incoming enquiries by service type, turning meeting notes into a structured task list, converting a scanned delivery order into rows in a spreadsheet. These jobs share three traits that make them good candidates. They happen often, the input is reasonably consistent, and a mistake is visible rather than silent.
Contrast that with pricing decisions or contract interpretation. Those also involve text, but an error there is expensive and difficult to spot after the fact. We tend to advise clients to leave that category alone until the easier ground has been covered and the team has built up some judgement about how the models behave.
Content work still needs an editor
For written material, language models are strongest at the middle of the process. They are poor at knowing what your company actually did last quarter and poor at judging what a Malaysian buyer in your sector cares about. They are good at restructuring a rough braindump into something ordered, producing four different headline options, or checking a draft against a style guide.
The practical arrangement that works is a human brief in, a machine draft in the middle, and a human edit out. Skip the first step and you get generic copy that could belong to any competitor. Skip the last step and errors of fact travel straight to your customers. Companies that publish unedited output are easy to identify from the outside, and it damages credibility faster than publishing nothing at all.
Operations work needs clean data first
Inside back-office systems, the constraint is almost never the model. It is the state of the records. If customer names are entered inconsistently across three systems, if half your service tickets have no category, if product codes were typed by hand for six years, then anything built on top of that will produce confident nonsense.
This is why our AI and automation projects often begin with a fortnight of unglamorous tidying: deduplicating contacts, agreeing a single list of service categories, adding validation to the forms that feed the database. Clients sometimes find this frustrating because it looks like a delay. It is not. It is the part that determines whether the intelligent layer above it can be trusted.
Decide who owns the output
Every automated step should have a named person accountable for its results, in the same way a manual step does — the discipline that keeps automation from becoming an invisible process. When an enquiry is auto-routed to the wrong department, someone must be responsible for noticing the pattern and correcting the rules. When a generated summary misrepresents a client conversation, someone must be able to override it and log why.
We also recommend keeping a plain record of where AI touches the business: which process, which tool, what data it sees, who reviews it, and how to switch it off. This takes an afternoon to write and answers most of the questions that come up later from auditors, insurers and enterprise customers running vendor assessments. It also prevents the situation where a departing employee takes an undocumented workflow with them.
Measure the boring things
Judge these projects on the same terms as any other operational change. How many hours a week did the step take before, and how many now? What proportion of generated output gets accepted without edits? Did the error rate go up or down? Those numbers are dull and they are the only reliable evidence that the work was worth doing.
Handled this way, AI stops being a strategic talking point and becomes an ordinary part of how a company runs — narrow, documented, supervised and quietly useful. That is a much better position to be in than owning an expensive platform nobody has integrated into daily work.
