By Bryan Knipe, founder of TriPinnacle Advisors
Every leadership team is being told that AI will take cost out of the business. Some of that is true. The part that gets left out is where the savings actually come from. In my experience, most of the money is found in the cleanup you do before the AI goes in: duplicate tools, overlapping teams and processes nobody has questioned in years. The AI takes the next layer, the high-volume, rules-heavy work people check by hand. Skip the cleanup and you end up automating a mess, paying for AI on top of everything you were already paying for.
I have done this work from both ends. At Optum, after we took over a large health system's entire analytics organization in an outsourcing deal, I led the transition: reconciling its tools, processes and people with ours and rebuilding how the work ran. More recently, I worked with a startup burning about $150,000 a month. We streamlined operations, cut waste and replaced manual work with AI tools I built, and brought the burn down to about $40,000 a month, roughly a 73% reduction, without stopping the work the company needed to prove. The scale was completely different. The sequence was the same.
1. Map the work before you touch the tools
Start with the work itself, not the software. For each team or process in scope, I want to know what comes in, what goes out, who touches it in between and how long each step takes. The fastest way to get this is to sit with the people who do the work and watch a real week of it, not to read the process documentation, which describes how the work was supposed to run when someone last wrote it down.
Three things always come out of this:
- Work that exists only to move data between two systems that should talk to each other.
- Reports that are built every week and read by nobody, or read by one person who would be fine with something simpler.
- Checking steps added after a past mistake that now check something the system already guarantees.
None of these need AI. They need a decision to stop.
2. Consolidate first
When two organizations come together, or when a company has grown by adding tools and people to each new problem, the biggest savings are usually in consolidation. In the Optum transition, the health system's analytics organization had its own tools, its own reporting and its own way of working. Running both side by side would have doubled the cost of everything. Reconciling them, one tool, one process and one team structure for each piece of work, is where most of the cost came out, while the reporting the business ran on kept working throughout.
The same applies at a smaller scale. Three tools that do roughly the same thing, two people producing overlapping reports, a vendor contract nobody remembers signing: consolidating these is unglamorous, and it is usually worth more than the first AI project.
3. Pick the work AI is actually good at
Once the process is clean, look for work that is:
- High volume. The same kind of task many times a week, so a small saving per item adds up.
- Rules-heavy. Someone can explain what a correct answer looks like, even if the inputs are messy.
- Document-based. Invoices, contracts, claims, filings, reconciliations and management reports, where people spend their time reading and re-keying.
- Checked by hand today. Which means you already know what an error costs, and you have a baseline to measure against.
Avoid starting with work that is rare, high-judgment or politically sensitive. AI can help there too, but it is a bad place to prove the case.
4. Check every output, at least at first
The fastest way to lose the savings is an AI that is confidently wrong. Someone downstream catches the error, trust collapses, and the team goes back to doing everything by hand while still paying for the tool.
The systems I build check every AI output against its source before anyone acts on it, and drop answers that cannot be supported rather than showing them. People stay on the exceptions: the items the system flags as uncertain. As the error rate proves itself on your real data, the share that needs a person shrinks. That is where the savings grow, and you can show the numbers to prove it.
5. Put a cap on what the AI costs
AI cost is easy to underestimate because it is spread across thousands of small calls. A process that costs pennies per document in a test can cost a lot more in production when documents are longer, retries pile up or someone loops the same step. Every run of the systems I build has a hard cost cap, and every call is metered, so the cost per document is a known number before anyone commits to a deployment. If you cannot state your AI cost per unit of work, you cannot prove you saved anything.
6. Plan for the people
Cost reduction that touches people's jobs fails when it is treated as a technology project. Decide early what happens to the time that is freed: does it go to work that was being neglected, to growth, or is it a reduction? Be honest with the team about which. In my experience, the people doing the work are the best source of what to automate, and they will tell you, as long as they trust the answer will not be used against them.
7. Measure the result against the baseline
Agree before you start what the savings will be measured in: cost per transaction, hours per week, cycle time or error rate. Measure the baseline before the change, then measure again at 30, 60 and 90 days. Report the full cost, including the AI, the reviewers and the time spent fixing errors, not just the headcount that moved.
Done in this order, cost reduction holds. The cleanup takes out the waste, AI takes out the volume, the checks protect the quality and the measurement proves it to the board.
Where to start
Pick one process that is high-volume, document-heavy and checked by hand. Map it for a week. Clean it up. Then decide whether AI should touch it. That is the shape of a workflow sprint: two to three weeks, one process, a prototype on your data, and a cost-per-run number before you commit to anything larger.