TL;DR
- AI adoption is as much habit formation as technology implementation, and the change is gradual, daily and mostly welcome. That is unlike ERP, where change arrived on one dreary winter morning.
- Traditional return on investment frameworks were built for capital projects, not capability projects. AI behaves like a new way of working rather than a new machine on the plant floor.
- Look for variance reduction at high leverage points. A model that makes a crucial decision consistent is usually worth more than one that automates interview scheduling.
- Do not sequence department by department. Sequencing builds data and integration debt and the flywheel never reaches a steady state.
Putting AI to use in an organization is as much about habit formation as it is about the technology implementation itself. Enterprise tech vendors talk about ‘change management,’ but the way ‘change’ is framed is the same way it was framed when they rolled out an ERP.
There are months of slow march toward a rollout and one dreary winter morning your organization cuts over to a new version of SAP. Change would be there, just as slippery as the black ice that you navigated to work.
With AI though, the change is different. It is gradual. It is every day. More importantly, it is a joyful kind of change. For the user, AI fills the gap, removes the grunt and makes the interface familiar, unlike software where change means learning new interfaces and new methods.
AI is intuitive in a way software never was.
AI creates value in places spreadsheets don’t naturally capture. A decision twin that reduces rework might not show up immediately in EBITDA, but it collapses hidden decision choke points. A retrieval layer that connects siloed knowledge doesn’t look like cost reduction or revenue increase. AI changes the quality of work that was sloppy to begin with.
Why the old return on investment frameworks do not fit
Traditional ROI frameworks were designed for capital projects, not capability projects. AI doesn’t behave like a new machine on the plant floor. It behaves like a new way of working – one that alters throughput, reduces variance, and unlocks latent value that was previously trapped in institutional knowledge or inconsistent execution.
At Moative, we attempted to look at ROI for AI projects from a new perspective.
Variance reduction
We come across variance reduction consistently. Variance tells you if the company’s performance is becoming more predictable, more stable, and more controllable. AI shines here. A single model that reduces decision inconsistency at a crucial point at a workflow by 30% often creates more enterprise value than a pilot project that automates scheduling candidate interviews. Pick workflows where variance reduction leads to outsized business outcomes.
Measure outcomes in terms of fewer executive hours needed for a decision or elimination of human judgment that needs middle management attention. Most work for leadership is just sloppy judgments that are pushed upwards for resolving. The mechanics of spotting those points before you spend anything are in Find the five gaps before spending on AI.
Control expansion, not cycle time compression
The second aspect we see is control expansion. How long does it take to close the loop between signal and action? AI shrinks this everywhere: in scheduling, in forecasting, in customer support, in engineering change orders. But the real impact isn’t the hours saved – it’s the fact that faster loops allow more loops. Increased iteration is the true productivity gain. Velocity frees up time for new questions. Most departmental leadership time goes into banal work - chasing people, dashboards, and untangling mess that comes from variance at work. Remove banal work and replace it with outcome driven projects. Measure AI not by what time it saved but also by how the time is used.
But here is the problem. Most of what I described above is viewed simply as “cycle time compression” and not “control expansion.”
Let me explain. AI that improves workflows tends to increase the surface area a team can handle without additional headcount. A planner managing 3 sites can suddenly manage 7. A support analyst who handled 20 cases a day can now oversee 50. This isn’t about replacing people. It’s about increasing the capacity of the existing system.
Adjacent gains
The other less intuitive aspect to look for to justify AI investments is adjacent gains.
Most pilots demonstrate isolated gains. Most orgs pilot AI at a function level. When they succeed they sequence the roadmap – one function after the other. We have talked about this before. Do pilots this way to demonstrate value but go enterprise-wide sooner. Unlock data across systems sooner.
Make the flywheel work. A string of adjacent gains compound value for the enterprise. One mini-rollout after the other, at one department at a time creates data and integration debt for your IT teams. They are not equipped. Besides, every new function throws new challenges that the sequence approach creates. Work is not defined by departments. Work simply flows. Its the human constraint that forced work into specialist functions.
AI systems that conform to human-led management style and its constraints repress value – not worth building them. The related point about what a first application is actually for is in Your first AI app is not a moat, it is a lever.
To sum up,
- Pick projects that deliver the best variance reduction at the highest leverage points in your workflows
- Cycle time compression is great. But go beyond that and ask how the quality of your middle management has increased with fewer distractions for them
- Allow adjacent gains to kick in fast. Don’t sequence your projects. They will never reach a steady state.
Enjoy your Sunday!
Sathish Rangarajan contributed to this article.
Where Moative fits
Moative looks for the decision that gets made inconsistently rather than the task that takes the longest, because the first one compounds and the second one is usually just a time saving. That tends to mean starting further upstream than a client expects, at the point where judgment enters the workflow. In healthcare revenue cycle management the highest variance is rarely in the submission itself, it is in how differently two people handle the same edge case before it ever gets there.
Frequently asked questions
How do you measure return on investment for AI projects?
Not the way you would measure a capital project. Look at variance reduction at high leverage decision points, at how much faster the loop from signal to action closes, and at what compounds when several improvements sit next to each other. Hours saved is the least interesting of the available measures.
What is variance reduction in an AI context?
It is making a repeated decision more consistent, so that the outcome depends less on who happens to be handling it. It matters because predictability is what lets you plan, and because inconsistent decisions are what generate escalations upward into management time.
Should you roll out AI one department at a time?
Use a single function to demonstrate value, then go wider sooner than feels comfortable. Sequencing department by department accumulates data and integration debt, and each new function introduces problems that the sequencing itself created.
What is control expansion?
It is the increase in surface area a team can cover without adding headcount, as opposed to simply doing the same work faster. A planner covering more sites and an analyst overseeing more cases are both examples.
Why is AI adoption different from an ERP rollout?
ERP change arrived as a cutover on a set date and required people to learn new interfaces. AI change is gradual and daily, and it usually removes grunt work rather than adding new steps, which is why habit formation matters more than change management theater.
What kind of AI project is not worth doing?
Anything built to conform to the existing departmental structure and its constraints. Work flows across an organization, and specialist functions are a human constraint rather than a property of the work, so a system that respects those boundaries suppresses most of the available value.
