Start With the Work, Not the Technology
Many people begin by asking, "Where can we use AI?" That question is hard to answer. It sends you looking for a place to put a new gadget.
Try a better question: "What does my team do every day that is mostly reading, copying, checking, and chasing people?"
That points you at real work. Work you can watch, time, and test.
Here is an easy way to begin. Spend one hour with the people who do the work. Ask them:
- What do you type in from an email, over and over?
- Which websites do you log into every morning?
- What job do you wish you never had to do again?
Some jobs are annoying but rare, like a monthly report that takes two hours. Other jobs take only five minutes but happen 80 times a day. Start with the second kind. Small savings add up fast when they repeat all day.
Example: It is 7:40 on a Monday morning. The shared orders inbox already has 23 new emails. Most have a customer's purchase order (PO) attached. One is a phone photo of a printed PO with a coffee ring on it. The order desk coordinator opens the first one. She finds the customer in your business software. She types in each item number, quantity, and price, one line at a time. One customer wrote "same as last time" next to a line, so she has to dig up last month's order. By lunch she has entered 30 orders. Almost none of that time went to a decision that only she could make.
Pick the Right First Job
Your first job for AI should be:
- Frequent. It happens often enough that saving time really matters.
- Easy to check. You can look at the result and say "right" or "wrong."
- Small enough to manage. One team, or your 20 biggest customers, not the whole company.
A good first job has a clear start and a clear finish. For example: "When a customer PO comes in, prepare the sales order." You can test that.
A poor first job is "help the sales team move faster." Nobody can tell when that is finished.
Example: A sales manager has three ideas and room for only one. Quotes take the most time, but every request is different. One customer wants 500 embroidered shirts by Friday. The next wants "something close to this part number." Chasing suppliers is the buyers' least favorite job, but the answers are scattered across six websites and a dozen inboxes. PO entry wins. Every order starts as an email with a PO attached and ends as a sales order in the system. That makes it the one project where you can say after ten days, "Yes, this is working."
Here is how those three jobs compare:
| Entering POs | Preparing quotes | Chasing suppliers | |
|---|---|---|---|
| How it starts | Customer emails a PO | A request by email, phone, or rep | A list of open orders each morning |
| What comes in | Mostly the same details | Different every time | Spread over websites and inboxes |
| How it ends | Sales order ready in your system | Quote sent to the customer | Updated ship dates |
| Easy to check? | Yes | Partly | Partly |
What Kind of Work Is a Good Fit?
The work does not have to be perfectly neat. It just needs three things:
- The same basic steps most of the time.
- Information your team can get to. For example, in email, your business software, a supplier website, or a spreadsheet.
- A few clear moments where a person has to make a judgment call.
Those judgment calls can stay with your people. The searching and looking things up around them can be handled for you.
Example: A vendor's invoice says "HX BLT 3/8-16 x 2 ZP." Your PO says "Hex bolt, zinc, 3/8 x 2." The accounts payable clerk opens the PO, checks that 40 boxes were received, then searches old invoices. She finds that this vendor has described the item this way four times before. It takes 11 minutes to reach an answer that was obvious once she had all the facts in one place. AI can gather those facts for her. She would only look at the matches that are truly unclear.
Write down how the work is really done, not how the manual says it is done. Note what starts the job, every screen your people open, the different ways information arrives, and what happens when something is missing.
Example: A rep gets an email: "Can I get a quote on the usual mugs, but with the new logo, for the spring event?" Written out, that is six steps. They include looking up the customer's price, checking stock on a supplier website, and working out shipping. Writing it down also shows what is missing. There is no quantity. There is no ship date. The new logo is sitting in someone's downloads folder. Those questions should go straight back to the customer, not sit in the rep's inbox for two days.
Decide What AI Does and What People Approve
You do not have to hand over everything. Sort each job into three piles:
- AI does it on its own. Safe, simple, low-risk steps.
- AI prepares it, and a person approves. The AI does the legwork. Your person says yes or no.
- A person does it. Some things should stay fully in human hands.
Keep a person in charge of anything involving money, promises to customers, inventory, or production. AI can prepare a sales order and still wait for a person to release it.
Give AI more freedom only after it has handled the messy cases your team really sees, not just the easy ones.
Example: Three POs arrive before 9 a.m.
- The first has 14 lines. Every item is known and every price matches the customer's contract. It becomes a sales order ready to go before anyone finishes their coffee.
- The second has two lines priced 6 percent below the contract. It goes to a review list, with those lines highlighted and last quarter's prices attached.
- The third asks for 4,000 units of something the customer usually orders 200 at a time. It waits for a manager to confirm it is not a typo before anything is sent.
What Are the Different AI Tools?
"AI tool" can mean many different things. You may already have tried one or two. Here is what each kind does well, and where it stops.
Chat assistants such as ChatGPT, Claude, Microsoft Copilot, and Gemini. These are very good at reading a messy PO, writing a supplier email, or summing up a long email thread. But they work one conversation at a time. A person still has to paste in the document, check the answer, and type the result into your business software. The copying is still there.
AI built into your business software. Many systems, such as Salesforce, NetSuite, Microsoft Dynamics, and SAP, now include AI helpers. They work well on information that is already inside that system. But most distributor work starts outside it, in an email inbox, a PDF, or a supplier website.
Document reading tools. These turn PDFs and scans into typed information. That solves the reading part. But deciding that "HX BLT 3/8-16 x 2 ZP" is item 44871, or that a price is too low, is still left to a person.
Tools that connect your apps. Examples are Zapier, Make, n8n, and Power Automate. They pass information from one app to the next. They work well when the information is tidy. Someone on your team has to build each connection, fix it when something changes, and handle cases it was not made for.
Screen-clicking robots (called RPA). Tools like UiPath and Automation Anywhere click through screens the way a person would. They help with old systems that cannot connect to anything else. But they follow one exact path. If a PO layout changes or a button moves, they can stop working until someone fixes them.
AI agents for operations. These combine all of the above. An agent reads emails and documents, looks things up in the systems you already use, prepares the work, and sends unusual cases to a person to approve. They fit best when one job crosses an inbox, a document, your business software, and a supplier website.
| Tool type | Good at | Where it stops |
|---|---|---|
| Chat assistants | Reading, writing, summing up | A person still types the result into your system |
| AI inside your software | Work on information already in it | Emails, PDFs, and websites outside it |
| Document reading tools | Turning PDFs and scans into typed text | Checking, matching, and the next step |
| Tools that connect apps | Passing tidy information between apps | Messy inputs, and every connection needs an owner |
| Screen-clicking robots | Old systems that cannot connect | Breaks when screens or layouts change |
| AI agents for operations | The whole job, start to approval | Needs a clear job and review points to begin |
Example: A distributor has already tried three tools for PO entry. The coordinator pastes POs into ChatGPT and gets clean lines back, but still types them into the system. A document reader handles PDFs from the five biggest customers, but gets stuck on the phone photo and the "same as last time" note. A Zapier connection creates orders from the website order form, but that is only 12 percent of orders. Each tool did its small part. None of them handled the order from the inbox all the way to a finished sales order.
It is fine to use more than one tool. The simple way to choose is to look at the finish line. If the job ends with a sales order ready in your system, pick a tool that can actually get it there.
If you are comparing a specific product, we have side-by-side guides for ChatGPT, Claude, Agentforce, Zapier, and n8n.
You Probably Do Not Need to Replace Your Systems
Most distributors already own the software they need. Your business system holds customers and items. Your customer system holds notes. Your supplier websites hold stock and delivery dates.
The real time drain is usually people carrying information from one place to another by hand. The software itself is rarely the problem.
Before anyone talks about replacing anything, look at how information moves today. Most systems can be connected in some way, even older ones. Some connect directly. Others can share files, read emails, or be operated on screen, often with a person approving the important steps. The question to ask is simple: can the work read and update information reliably, without adding risk?
Example: Follow one order at a regional distributor and count the windows. The PO arrives in Outlook. The coordinator copies the account number into the business system. She flips to the customer system to check whether this customer needs a liftgate truck. Then she drags the PO into a folder named by year, month, and customer. That is four systems and a lot of switching back and forth for one order, repeated 60 times a day. The answer is not a new system. The answer is to have the order information move once, while all four systems stay right where they are.
This is the kind of work Trelium is built for. Trelium agents work inside the systems you already use. They read the inbox, check your business software, and prepare the work. Your team approves anything that matters. There is no move to a new system, and no big replacement project. Your current system stays.
Is Your Job Ready for AI? A Simple Checklist
Go through these questions. The more you can answer "yes," the better your chances.
How often, and how the same?
- Does this job happen often, so that saving a few minutes each time really adds up?
- Do the steps look about the same most of the time, even if each order or document looks a little different?
- Is there a clear starting point, like a new PO, an invoice, a quote request, or a supplier email?
Information and results
- Can your team get to the information needed, in email, your business software, a supplier website, a spreadsheet, or a shared folder?
- Can you describe what a correct result looks like?
- Can you list the unusual cases, even if AI cannot handle all of them?
Checking and measuring
- Can someone on your team look over the unclear cases while you are testing?
- Would a mistake be caught before it causes a serious problem?
- Do you know how much time and rework this job causes today?
How to Tell If It Is Working
Count results that matter to your business. Do not count how many tasks the computer did.
Good things to track:
- Time spent on each order
- Hours given back to your team
- Orders waiting in the pile
- How fast customers get an answer
- How often unusual cases come up
- How often something has to be fixed afterward
- Cost to process each document
Measure these before you change anything. Then you have something to compare against.
Be careful about "time saved" numbers. If order entry drops from eight minutes to two, but someone spends five minutes fixing the result, you have not saved much.
Also look at normal orders and unusual orders separately. Say AI handles 70 percent of orders cleanly and safely hands the other 30 percent to your team. That is often better than a system that tries to do everything and is quietly wrong some of the time. Hidden mistakes tend to be found at the worst moment, by a customer.
Example: A distributor handling 80 POs a day checks four numbers every week: average time per order, share of orders needing a person to look, mistakes found afterward, and total daily processing time. When the number needing review drops but the mistakes stay the same, things are improving. When both drop, it is time to give the system more orders.
Watch the Unusual Cases, Not Just the Average
A single "95 percent accurate" number hides what matters most. Make a list of what can go wrong:
- An item number nobody recognizes
- A price that does not match the contract
- The same order sent twice
- An order that ships in parts
- A customer who changes their PO
- A supplier who never confirmed
- An invoice with no matching PO
Your team should be able to see what went wrong, why, and what information the system was using at the time.
Every unusual case needs a backup plan. If a supplier website is down, the system should mark the order "not confirmed." It should never quietly assume the order is on time.
A system that tells you loudly when it is stuck is one you can trust. A system that fails quietly teaches your team to stop checking. That is how a missed confirmation turns into a late delivery, and then into a lost customer.
Example: Your buyer chases three vendors. One has no way to connect to it. One needs a website login. One only answers by email. The system checks all three every morning. When a check fails, it is shown to the buyer on a list instead of being marked as a success. The buyer starts the day with eight orders that need attention, not 300 that might.
Common Mistakes to Avoid
Starting too big. "Automate customer service" cannot be tested or finished. "Answer order-status questions using our system and shipping data" has clear edges. Clear edges are what make a first project succeed.
Automating a job nobody understands. Suppose three employees each fix invoice problems in three different ways. The problem is not missing technology. The problem is that nobody has agreed on how to do the job. First find out why they differ. Sometimes all three are right for different customers, and that becomes a rule. Sometimes two of them are workarounds for a problem that no longer exists.
Taking people out too soon. Human review matters most where the money and relationships are: prices, promises to customers, payments, and unusual orders. If you take your people out before the system has proven itself, you may spend all your savings on apologies.
Example: A company starts an order-status helper that only answers questions when your system shows a confirmed ship date and a tracking number. Anything unclear goes to customer service. At first it covers 60 percent of incoming questions. That means 60 percent fewer phone calls. As the rules improve, the number climbs.
Questions People Often Ask
What is the best first AI job for a distributor?
Pick a job that happens a lot and has a clear start and a clear finish. Entering customer POs is the most common choice. It starts with an email or PDF and ends with an order in your system. The best answer is different for every company, so compare PO entry with quotes, invoice matching, and supplier follow-ups. Ask which one takes the most time, follows the most regular steps, and has the lowest risk if something goes wrong.
Does the job have to be exactly the same every time?
No. A job can have some variety, as long as you can describe the normal way and the main exceptions. POs come in dozens of layouts, yet they all contain the same basic things: customer, item, quantity, price, ship-to address, and date wanted. When something is unclear, it goes to a person instead of the system guessing.
Should we hand over the whole process at once?
Usually not. Start with the repetitive part that eats the most time, and keep a person approving the riskier steps. For example, AI can read an order, find the customer, match the products, and prepare the entry. A person then approves price differences or unusually large orders. Give AI more freedom once it has proven itself on your real-life cases.
How long until we know if it is working?
A few weeks, if you have real examples to test with. In the first month, the goal is not to handle every unusual case. The goal is to learn whether the system saves time on normal cases and safely hands off the rest. If you cannot tell after 30 days of measuring, testing on past orders, and running side by side with your current way, the problem is probably the job you picked, not the timeline.
What should we measure?
Measure the thing you want to improve, and write down where you stand before you start. For order entry, that means time per order, orders handled per person, how often fixes are needed, how many orders are waiting, and how many need a person to look. For supplier follow-ups, that means hours spent chasing, overdue confirmations, and how quickly buyers hear about problems.
What if our business software is old?
That rarely stops the project. Depending on the system, the work can use shared files, scheduled reports, direct access to the data, reading emails, or working on the screen like a person would. A person can approve the important steps. The real question is whether the work can read and update information reliably without adding risk. Older systems pass that test more often than software vendors admit.
Can we just use ChatGPT or Copilot?
For single tasks, yes. A chat assistant is a quick way to pull the lines out of a PO or write a supplier follow-up. The catch is that it works one conversation at a time. Someone still has to give it the document, check the answer, and type the result into your system. That is fine for occasional work. For a job that happens 80 times a day, most of the time goes into the copying around the AI. That is the part you wanted to get rid of.
Will AI replace our team?
A good first project takes repetitive work off your team. It does not take their jobs. People still handle the unusual orders, the customer decisions, the price disagreements, and the supplier problems, where experience and judgment matter. The goal is for your buyers to review a short list of problems instead of checking every open PO by hand. That is a better job, and a harder one to cut.
How do we know when a job should stay manual?
Keep it manual when it rarely happens, depends on negotiation or personal judgment, has unclear information, or would be very costly if done wrong. A yearly supplier negotiation is a poor fit. Checking 300 open purchase orders every morning for missing confirmations is an excellent fit.
Download this guide
Get a PDF to share with your team. Enter your name and email and the file downloads on this page.
Get started
Ready to see Trelium in action?
Schedule a 30-minute conversation about the workflow you want Trelium to handle.
Talk to a Human→