Digital transformation in the AI era
Agentic AI is the next chapter of digital transformation, and the businesses that already digitised and connected their operations are the ones positioned to benefit. A worked example from Si Novi's own books.
A few years ago I wrote about digital transformation: digitising the processes a business already runs on, replacing the awkward spreadsheet, connecting the accounting software, building the integration that means data is entered once instead of three times. That has been the shape of the work for the best part of a decade. Agentic AI doesn't replace any of it. It changes what becomes possible once it's done.
What actually changed
For a long time, automating a process meant encoding rules. If you could write the process down as a series of definite steps, you could build software to follow them. If you couldn't, it stayed with a person. The awkward middle - the tasks that need a little judgement, that involve reading a messy email or a supplier's oddly formatted invoice, that throw up an exception every third time - is where automation tended to give up.
That middle is exactly where agents are good. A large language model can read the messy input and make the routine call, handing back only the cases that genuinely need you. An agent goes further again: give it a task, a set of tools and a connection to your systems, and it can pull the data, work through it and propose what to do.
Why the groundwork matters
Here's the catch, and it's why a decade of unglamorous integration work suddenly matters a great deal. 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. If your operation still runs on disconnected spreadsheets and manual re-keying, there is nothing for it to take hold of.
I'll give you an example from my own week. I do the bookkeeping for Si Novi, and like anyone who keeps their own books there are a few jobs I actively dread: processing draft bills, working out the real sterling cost of bills from suppliers who invoice in dollars, and reconciling the bank. This week I built a small suite of AI-assisted tools to take those three off my hands. Underneath it is nothing exotic - a collection of scripts, mappings and config files wrapped around the Xero Node SDK - but I run it inside a Claude Code session, which lets the agent pull data from the Xero API, analyse it, and then present me with clear decisions to approve rather than acting on its own. I used it to clear a backlog of bills and get our reconciliation together in time for a catch-up with our accountants, and, if I do say so myself, it works brilliantly.
None of it would have worked without the groundwork, and the groundwork is really about reach. What made it possible is that Xero exposes an API, something I've built integrations on for years: the data was already digital, structured and reachable by software, rather than sitting in a drawer or a desktop spreadsheet. Keeping our books there over the years meant it was clean and well-organised too, which helps, but the real enabler was simply that the business runs on connected systems an agent can get to. The agent didn't transform the business. It stood on a business that was already connected.
Where a person still belongs
That approval step matters more than it might sound. I am not letting an agent post entries into the accounts unsupervised, and I wouldn't advise anyone else to either. The pattern that works, at least for now, is the agent doing the reading, the sifting and the proposing, and a person making the decision - particularly anywhere the cost of a confident wrong answer is high, and in bookkeeping it certainly is. Agents 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.
I'll also be honest that this is addictive in a way ordinary software rarely is. "One more prompt and then I'll stop for the day" is how a session begins, and seven hours later you surface. We're buzzing at Si Novi with ideas for where else an agent could sit in the loop, and that enthusiasm is genuine. The discipline is in pointing it at the jobs that are actually worth automating rather than the ones that are simply satisfying to automate.
The pattern I keep returning to is that agentic AI rewards the businesses that did the boring work first. If your systems are connected and your data is in order, there is an enormous amount an agent can now take off your plate. If they aren't, that is where to start, and it's the same advice I'd have given five years ago. The difference is that the payoff is now a great deal bigger. If you're wondering where AI might fit into how your business actually runs, I'm always happy to talk it through.