In late March, NewTide Chairman Doug Haugh sat down with a Colonial Oil study group to work through what it actually takes to move AI from interest into operation. Our thanks go to Colonial Oil, an existing customer, for the invitation to present, and to Jeff Bernard for facilitating the meetings and helping shape the session. It was a real opportunity to talk through where AI is heading with some of the leading companies in the industry. Study groups run differently than conference sessions. The room is smaller, the people in it run real businesses, and the questions come fast and skeptical. That fit the material, because Haugh’s case was about execution rather than enthusiasm.
He opened with a picture every trading desk will recognize. A trader sits in front of twelve screens and can only act on what those screens show. The opportunities that matter often sit on the thirteenth screen, the one nobody has room for. The point was not that traders need more monitors. It was that the work of watching has outgrown what a person can do by hand, and that is exactly the kind of work that should move to software.
Agents that do the work, not tools that suggest it
The clearest part of the session was Haugh’s distinction between a tool and an agent. A tool answers a question or surfaces a recommendation and then waits. An agent carries a process through to the end, keeps track of where things stand, and acts inside the systems a company already runs on. A person stays in the loop and steps in when judgment is required.
He made that concrete with a trading and supply example built around a small set of agents, each with a defined job. One scans spot versus rack spreads across every terminal around the clock and fires an alert only when a spread moves outside its historical range, with the netback already calculated. Another simulates contract margins against live OPIS, Platts, and DTN feeds and flags the moment a competitor’s rack posting lags a price move, then recommends the adjustment. A third captures trade confirmations out of chat and email, writes them into RightAngle or Allegro, and clears small invoice variances against pre-set rules. None of these replaces the trader. They handle the watching and the keying so the trader can decide.
What the numbers look like in dispatch
The dispatch example is where the operational case got specific. Haugh described a setup running continuous monitoring with a handful of specialized agents, exception response under two minutes, and full delivery coverage. The agents split the job cleanly. One reads live tank levels and generates delivery requests before a runout happens. One runs best-buy math across terminals, accounting for rack price, freight, taxes, and discounts, and catches the cases where a farther terminal actually wins. One solves the routing problem against truck compartments, driver hours of service, and fleet hierarchy. The last one handles exceptions. When a pipeline allocation gets cut, it reruns the best-buy calculation on alternate terminals and hands the dispatcher an approved alternative rather than an error.
The back-office example followed the same shape. Invoices arrive by email, EDI, or portal and get read without manual keying. The system matches them to purchase orders, codes the GL entries from historical patterns, and routes approvals by dollar threshold and cost center so only real exceptions reach a person. Approved invoices post straight into the accounting system with an audit trail. Haugh put concrete figures against it: near zero-touch ingestion, a 95 percent auto-match rate, and roughly 60 percent less processing time, with month-end close pulled in by days.
The thread across all three examples was the same. The value is not in a smarter dashboard. It is in closing the loop so nobody has to sit and monitor an inbox, a tank gauge, or a spread.
Naming a champion and giving them room
When the conversation turned to how any of this gets started, Haugh’s advice was to stop trying to convince the whole organization at once. Every company has a few people already using these tools on their own and asking why a given task cannot be automated. Those people have self-selected, and they are where to put the weight.
His model for it was practical. Find the curious. Give them a real title, something like AI Innovation Lead, because a title creates both accountability and permission. Fund them with a modest budget, in the range of five to twenty-five thousand dollars, and give them ninety days. The small budget and the deadline force focus instead of sprawl. Then show the results, the wins and the smart failures both, in front of the whole company, because what gets celebrated gets repeated.
A plan that fits in a quarter
Haugh closed the working part of the session with a ninety-day plan that any operator in the room could start the next morning. Month one is assessment: audit the AI already sitting unused inside your systems of record, survey the team to find who is already experimenting, draft a first acceptable-use policy, and pick two or three champion candidates. Month two is the first pilot, with a weekly review cadence and the champions running their experiments. Month three is the decision point: measure the pilot’s return, kill it or scale it, stand up a governance committee, and set a twelve-month roadmap with a budget attached.
The framing throughout was that none of this requires betting the company. It requires picking something useful, putting a named person behind it, and measuring what comes back.
Study groups exist so operators can compare notes on what is working and what is not, away from the show floor. That is the conversation NewTide wants to be in, and it is the kind of problem these agents are built to handle in fuel and convenience.