Everyone has access to AI, getting started is the easy part. What’s harder is measuring what AI delivers – and if you’ve been struggling to work out the ROI of your AI investment you’re not alone.
At our recent Managed Services User Group I shared this concept which is one of the most talked about topics in enterprise AI at the minute – the rise of tokenmaxxing.
Tokens are the basic unit of data that an AI model uses to read and generate information. A large language model (LLM) consumes tokens in every prompt, agentic action, and automated decision. The value of what a business pays for token consumption makes AI spend a measurable operating cost, similar to cloud compute spending a decade ago.
There’s a growing trend of organisations adopting ‘tokenmaxxing’ which incorrectly treats token spend as a productivity metric. But this isn’t the route to success – we believe the winners in AI will be the organisations that can name the agents, workflows, and processes that turn token spend into measurable value.
Why measuring AI impact is the hard part
Nearly six out of ten (59%) organisations have moved agentic AI beyond the pilot stage, but only 9% have made significant progress in creating autonomous, multistep workflows, according to ServiceNow’s 2026 AI Maturity Index.
Most are paying for a capability they haven't yet unlocked, the report finds. The spending behind that gap is significant and growing. AI investment rose 110% in a single year and is projected to rise a further 81% by 2027, taking it past 20% of the average IT budget.
Here’s the hard part: that cost accrues where most organisations can't see it. Token consumption is a live budget line that is not fully accounted for, drawn down with every prompt, agentic action, and automated decision. Yet few organisations can track what each agent or workflow consumes or returns, let alone measure its impact.
Without that visibility, costs climb while the return stays unproven. It's the same pattern Gartner points to: governance gaps, escalating costs and unclear business value are what get agentic AI initiatives cancelled. Let's look at what organisations can do instead.
Five things to get right on the path to AI ROI
Based on the conversations I’m having with customers, the organisations that earn compounding returns from AI tend to get five things right.
Nail these, and the rest follows for realising AI value:
- Right use case: start where the impact is visible and undeniable. Favour fast payback and clear business alignment over the most impressive demo.
- Right architecture: fix the foundations before automating, because AI amplifies what is already there, good or bad.
- Right people: serve both human requesters and AI agents. A tool that helps both sides drives compounding value and far higher adoption.
- Right governance: set guardrails from day one. Early governance costs less than retrofitting later.
- Right data: the data feeding AI must accurately represent the business for AI outputs to deliver value.
How to measure AI ROI
Cost savings that come from greater efficiency don’t always tell the full story. A virtual agent “deflecting 40% of cases” means nothing if those users re-log the same issue minutes later. This represents duplicate effort, inflating the apparent cost saving on top. That's the trap of vanity metrics, which can show movement on a dashboard but no change in the outcome.
Effective AI measurement ties every metric back to business value. Rather than just counting what’s been deployed, here are the core AI KPIs we’ve helped implement with many of our customers:
- Decision quality: understanding whether the decisions AI makes or supports hold up, not how many it makes. In the deflection example, the real test is if the issue was genuinely resolved and didn't resurface.
- Intervention rates: how often a human has to step in to correct, override, or take over. This is where human oversight becomes measurable. A falling intervention rate on a stable workflow shows an agent is earning trust.
- Time to value: how quickly a use case delivers measurable benefit once live. Fast payback is the point – a workflow returning value in weeks beats an ambitious one that takes a year to prove.
Every metric you use must tie back to business value, and the real test is whether a use case moved a real outcome, not simply whether it shipped. This is the difference between AI investment ROI and activity.
How to start finding value from your AI investments
Knowing how to measure ROI only matters once you've pointed AI at the right problem. Start by identifying your highest-friction workflows – this is where work slows, queues build, and manual effort concentrates. AI pays back fastest in this area.
Next, match the task with the right type of AI – generative, predictive and agentic AI each serve a different purpose. Check which capabilities you have available, what they consume, and whether a custom skill or even a simple workflow delivers the same value.
For higher-risk actions involving agents, keep a human in the loop. I've seen organisations strip out oversight to chase efficiency numbers, and it nearly always backfires – errors compound, trust erodes, and you end up rebuilding the workflow from scratch. Oversight is a strategy, not a compromise. An agent deflecting 60% of manual work and escalating relevant cases still delivers real value, and you keep the safety net that lets you scale it further.
ServiceNow provides a broad range of AI solutions, and with the recent Australia release, over 100 out-of-the-box AI Agents are now available. UP3's experts can help identify the right AI tools for your needs.
Moving AI from line-item cost to return on investment
UP3 helps you answer the question: are you spending your assets on the right tools in the right way, or is there a better approach?
We can help you realise value from AI by identifying your highest-impact AI use cases. Our AI Unlocked Service Offering gets your first AI agent live in as little as three weeks, while setting up how you measure AI impact effectively.
