How AI is put to work in real operations.

People often ask me how AI is actually being used inside an operating company. My answer is that the most valuable applications are rarely the flashy ones. They sit inside the everyday work of running the business and the science.

Over the past year I've put AI to work across seven areas:

• Accounting and vendor tracking. Turning invoices, vendor records and budgets into one current, professional view of spend, commitments and budget-versus-actual.
• Project planning. Designing Gantt charts that are visually appealing and easy for any audience to understand at a glance.
• Scientific computation. Setting up molecular dynamics simulations so scientists can evaluate more designs, faster.
• Meetings. Moving from recording to transcript, summary and presentation-ready slides automatically.
• Custom models. Building and training models for prediction and automated workflow analysis.
• Private RAG models. Giving teams answers drawn from their own documents, with sources cited and the data kept private.
• Governance. Putting rules, review and documentation around all of it so the work can be trusted.

What ties these together is that the technology is not the hard part. The hard part is understanding the operation well enough to describe what it needs. When you can do that, one person can accomplish what used to require a team, and a team can accomplish far more than it thought possible.

The notes that follow describe how I approach each one. If you're thinking about how AI fits into your organization, I would love to talk with you.

Every growing organization reaches the point where its financial information lives in too many places: accounting software, email attachments, spreadsheets and someone's memory. The cost isn't only time. It's not knowing, at any given moment, what has been committed, what has been paid, what is coming due and how spending compares with the plan.

The approach I use starts with the questions leadership actually needs answered each month. Who are we paying, and for what? What is outstanding? How are we tracking against budget, by program and by category? From there, AI does the heavy lifting: reading invoices as they arrive, extracting the vendor, amounts, dates and terms, coding each one to the right account and commitment, and flagging anything that looks inconsistent or out of policy.

The same structure supports professional accounting and budget tracking across the organization. Expenses roll up into budget-versus-actual views, vendor histories stay complete, and month-end preparation starts from clean, organized records instead of a scramble to assemble them.

People still approve the payments and own the books. AI removes the hours of gathering and reconciling that used to come before every decision. The lesson for any leader: define the answers first, then let the system collect the data to support them.

Most project timelines fail for a simple reason: people don't read them. A dense grid of dates and task names may be accurate, but if a leader, a funder or a new team member can't understand it at a glance, it isn't doing its job.

That is why I built my own approach to Gantt charts. The goal was a timeline that is visually appealing and easy to understand, so the plan communicates on its own.

Every design choice serves that goal. Each owner gets a color, so anyone can see who is carrying what without reading a single label. Stages sit in softly tinted bands, so the phases of a project are obvious at once. Milestones and decision gates stand out as clear markers, and a today line shows exactly where the work stands. Branding, a title and an accent color are applied automatically, so the chart looks finished, not like a working file.

Because a timeline gets shared far beyond the project team, the same chart exports cleanly to a styled spreadsheet, a presentation-ready image or a printed page that fits properly. It reads as well in a board update or a funder report as it does in a weekly stand-up.

What I've learned is that clarity is a form of alignment. When everyone can see the plan the same way, the conversation moves quickly from "what are we looking at?" to "what do we need to decide?"

Molecular dynamics simulation lets scientists watch how a molecule is likely to move, fold and interact before anything is made in the lab. It is powerful, but traditionally slow to set up and expensive to run, which limits how many ideas a team can test.

My role has been to make that work operational. With AI, the setup, configuration and analysis steps that once required a dedicated computational specialist can be scripted, standardized and run on scalable cloud computing. Simulations launch consistently, results come back in a standard format, and the analysis is ready for scientists to review instead of waiting to be assembled.

The scientists stay in charge of the science. They decide what to test and judge what the results mean. What changes is the pace: more designs can be evaluated, comparisons are cleaner, and the path from a question to an answer gets shorter.

For research organizations, this is where AI creates real value. Not by replacing scientific judgment, but by removing the computational bottlenecks that keep good ideas waiting in line.

Most organizations lose a surprising amount of value after their meetings end. Decisions live in someone's notes, action items drift, and preparing the next update means starting from scratch.

I've built a workflow that closes that gap. Meetings are recorded and transcribed automatically. AI then produces a structured summary: the decisions made, the open questions, the action items and who owns them. From there, the same content becomes a short slide deck in a consistent, branded format, ready for the next leadership review or partner update.

The value isn't just saved time. It's continuity. Everyone works from the same record of what was decided, and nothing important depends on one person's memory. When a new team member joins, the history is already documented.

Two principles keep it trustworthy. A person reviews every summary before it is shared, and sensitive conversations follow clear rules about what gets recorded and where it is stored. Automation handles the drafting. Judgment stays with the team.

General-purpose AI tools are remarkably capable, but every organization eventually reaches questions they can't answer well, because the answers depend on its own data, processes and expertise.

That is where custom models come in. I build and train models for two kinds of work: prediction, where the model learns from an organization's historical data to estimate outcomes for new cases, and automated workflow analysis, where it reviews results in a standard way so people can focus on the exceptions.

The most important decisions happen before any training begins. What exactly should the model predict? What data is reliable enough to learn from? How will we know whether it is right? And who reviews its output before anyone acts on it? Getting those answers right matters more than the choice of algorithm.

A custom model is also not a one-time project. It needs to be measured, retrained as new data arrives and documented so others can understand what it does. Treated that way, it becomes an asset that grows more valuable with every cycle of the work.

Every organization has knowledge that no public AI model has ever seen: its protocols, reports, research history, contracts and the hard-won lessons in years of internal documents. A private retrieval-augmented generation model, often called private RAG, puts that knowledge to work without sending it anywhere it shouldn't go.

The idea is straightforward. Instead of retraining a model, the system searches the organization's own approved documents for the passages relevant to a question, then uses them to write an answer that cites its sources. Staff can ask in plain language and see exactly where each answer came from.

The word that matters most is private. The documents stay in the organization's own secure environment. Access follows the same permissions people already have, so no one can retrieve what they couldn't open themselves. And every answer points back to its source, so a qualified person can verify it before acting.

Done well, private RAG becomes an organization's institutional memory: new team members get up to speed faster, experienced people stop answering the same questions twice, and decisions draw on everything the organization already knows.

As AI moves from experiments into daily operations, the question shifts from "can it do this?" to "can we trust what it did?" Governance is how an organization answers that.

The framework I use is practical. First, sort the work by risk. Low-stakes, reversible tasks can run on AI with light oversight. Anything that reaches customers, regulators, funders or financial decisions needs a named, qualified person to review it before it leaves the building.

Second, make review visible. Log who checked the work, what they verified and when, so oversight becomes something you can show, not just something you assume. Third, document the systems themselves: what data each tool uses, where that data is stored, what each model is designed to do and what it should never be used for.

Good governance doesn't slow adoption down. It is what allows a team to move faster with confidence, because everyone knows where the guardrails are. It is also the heart of the Operational Credibility Square: in the age of AI, credibility is the corner that has to be measured.