AI automation for business: where it fits, and where it doesn't
AI can automate the judgement-heavy middle of your operations - reading messy inputs, routine decisions, exceptions - but only once your systems are connected. A practical guide to what works, what doesn't, and how to start.
AI can now automate the parts of your operations that used to be too fiddly for software: reading a messy supplier email, making a routine judgement call, handling the exception that comes up every third time. The catch is that it only pays off if your systems are already connected and your data is reachable. Point it at one specific, repeatable, high-friction task, keep a person in the loop to approve anything costly, and the results can be transformative. Try to automate everything at once, on top of disconnected spreadsheets, and you will be disappointed. Here is where AI automation actually fits in a business, where it does not, and how to start.
What AI can actually automate
For years, automating a process meant writing down definite rules. If you could express the task as a series of fixed steps, you could build software for it; if you could not, it stayed with a person. AI changes that by handling the awkward middle - the work that needs a little judgement. In practice, the tasks worth pointing it at look like this:
- Reading and extracting data from messy inputs. Invoices, receipts, emails, PDFs and forms, pulled into your systems without manual re-keying.
- Routine judgement calls. Categorising a transaction, triaging a support message, qualifying a lead - the decisions where the rule is really "it depends".
- Exception handling. The odd cases a rules-based automation would otherwise kick out to a human.
- Drafting and summarising. First drafts of replies and reports, or summaries of long threads and documents, for a person to check and send.
- Reconciliation and cross-checking. Matching records across systems and flagging only what does not line up.
Where a person still belongs
Automate the reading and the sifting, not the final decision - at least anywhere a confident wrong answer is expensive. The pattern that works is the agent doing the analysis and proposing what to do, and a person approving it. I would not let an agent post entries into the accounts unsupervised, and I would not advise anyone else to either. AI systems are fast and tireless and, every so often, cheerfully wrong. Designing the person back into the loop at the right point is most of the work.
What you need in place first
An agent is only as useful as the systems it can reach. It needs clean data to reason over and a connected system to act on. This is why a decade of unglamorous digital transformation work suddenly matters so much: the businesses that already replaced their spreadsheets with proper applications and connected their accounting and other systems are the ones an agent can actually take hold of. If your operation still runs on disconnected spreadsheets and manual re-keying, that groundwork is where to start - and it is worth doing on its own merits, before any AI is involved.
How to start
Start small and low-risk. Pick one task that is repeatable, high-friction and painful, keep a human approval step, and measure the result before you widen the scope. Momentum comes from proving the value on a single job and expanding from there, not from a grand automation programme. It is the same advice I would have given five years ago about automation generally. The difference now is that the payoff is a great deal bigger.
The best place to look is the job you quietly dread: the repetitive, fiddly task that eats an afternoon and never quite gets automated. If you are wondering where AI might fit into how your business actually runs, I am always happy to talk it through.