66 VIEWPOINT Proving the return on enterprise AI with Algoworks’ The same question is being asked in nearly every boardroom: we’re spending on AI, and usage is up, so where’s the financial return? It’s a fair question, but few companies know how to get the answer Derek Harrar is CEO at Algoworks Every enterprise has an AI story now. And yet, most companies don’t have a way of measuring its worth. Use can climb every month, while revenue doesn’t budge. This isn’t a success story, it’s a warning sign. Microsoft made moves on its own side of the table this year. Partner performance on Microsoft 365 Copilot is now scored on paid monthly active usage rather than seats sold, and customer references have been replaced by an independent third-party audit. Seats gave way to measured use, and assertion gave way to evidence. Enterprises should make these moves too. Part of the problem is that we’re applying old software habits to something fundamentally different. When we used to build software, we shipped it and it stayed built – the code didn’t change. AI isn’t like that. Models continuously change and get retired, so a workflow that worked six months ago can get worse without anyone touching it. Most companies can’t track that. The fix isn’t another dashboard. It’s a measurement discipline, and ours is called AI:CI – continuous improvement applied to AI itself. It starts with a better question: how much faster is the work getting done, compared to how long it used to take? We call that leverage – if a task used to take 10 hours and now takes two, that’s five times the leverage. The important part is what you’re comparing against. A lot of AI claims compare today to some imaginary “we’d have needed twice the headcount without AI” scenario – a number nobody can prove and a board will pick apart in seconds. Leverage compares two things you can actually count: how long a task took before, and what it takes now. That number holds up. But leverage doesn’t keep climbing forever just because the model gets smarter. It gets capped by friction – the extra work someone still must do to prepare the input and clean up the AI’s output before it’s usable. The arithmetic is unforgiving: a five per cent prep-and-rework tax caps your realised return at 20 times and a 10 per cent tax caps it at 10 times, no matter how good the model becomes. Beyond that, buying a faster model stops paying off. The real return comes from reducing that manual cleanup work, not from chasing the next model release. Agents make both problems harder rather than easier. When a workflow runs as a chain of agent steps instead of a single prompt, friction stops being the cleanup on one output and becomes the verification of a sequence of them. The question of which number in that chain was measured rather than inferred becomes the whole argument. The governance conversation the industry is now having about agents is, underneath, the same measurement conversation on steroids. That’s also why token spend – what you’re actually billed for running the AI – needs its own scrutiny. A workflow can look like a win on paper if people are getting more done, while quietly costing more in AI fees. That’s AI that works but doesn’t pay. You need to see both numbers side by side.
RkJQdWJsaXNoZXIy NzQ1NTk=