AI for Quotes and Order Entry: How to Approach It
From incoming request to order in your ERP or TMS: where AI saves time, where a person signs, and how to start without automating the old process.
From request to order: where the time goes
Every company that sells to other companies runs the same six steps: capture, qualify, specify, price and quote, negotiate and approve, order and handover. Together with Dagmar Ritterbex of The Lean Kitchen, I worked that chain out in the Lead to Order case.
Ask how long the chain takes and the answer is weeks, sometimes months. Ask how much of it is actual work and the answer is hours or days. The rest is waiting: in an inbox, for a meeting, for a sign-off, for someone to re-type what someone else already typed. In the fictional example from the case, one hospital tender takes 24 working days. Three of them are work.
That gap is where AI promises the most. It's also where AI can do the most damage if you get it wrong. Most companies start at two steps: the quote, which is at the top of my list of AI use cases for SMEs, and order entry.
What AI does with an incoming request
A request arrives by email, through a web form, as a PDF attachment or as notes from a phone call. An AI agent reads that request and turns it into structured data: who the customer is, what they want, how much, when, where to and on what terms.
In transport it looks something like this. A shipper emails an order with loading and delivery addresses, time windows, the number of pallets and a reference. The agent extracts those details, looks up the customer in your system, checks whether the addresses are known and whether the order is complete, and prepares a draft order in the TMS. If something is missing, it sends the question back to the customer instead of guessing.
That's re-typing work that often sits with planners or customer service today, and it's exactly what a language model is good at: reading messy text and turning it into fields. How an agent works safely in your own systems, with permissions per user and a log, is covered in AI on your own business data.
Creating quotes with AI: the agent drafts, a person decides
At the quote, the role of AI changes. The agent can do the reading, looking up and drafting: finding the right items or rates, checking earlier quotes for the same customer, writing the quote text and checking that everything matches the request. It saves the most on quotes that resemble each other, and most companies have plenty of those.
The price is a different story. In the case we show what happens when an agent sets the price alone: it optimises for the chance of winning and prices below cost. With a tender, you're then locked into that price for years. That's why, in the case, a named person always signs off the final price.
What does help is agreeing margin rules once. In the case, that's the fix at the pricing step: finance sets the rules once and from then on only sees the exceptions. An agent applies those rules and only escalates the quotes that fall outside them. The decision stays with a person, without every quote waiting for a meeting slot.
Automating order entry with AI in your ERP or TMS
Once the customer accepts, the order has to go into the system. In many companies that means: open the quote, open the ERP, re-type. That's where the errors come from that you later see as wrong deliveries and credit notes. The case names exactly those two wastes at this last step: re-typing and defects.
Order entry is the step best suited to run on rules. The data is already structured and approved, so the agent doesn't have to decide anything. It enters the order in the ERP or TMS, links it to the quote and flags anything that doesn't fit. People handle the exceptions. Every time I connect an order agent to an ERP, the same thing happens: most of the time goes not into the technology but into agreeing what counts as an exception.
Mind that word, exceptions. The case also shows what happens when an agent resolves an exception by itself: to hit a delivery date, it ships a batch with three months of shelf life against a twelve-month requirement. Nobody approved that deviation. An order agent can handle the standard. A deviation from the standard is a decision.
The trap: making the old process faster
It's tempting to put an agent on every step and change nothing else. In the case we calculated exactly that. The tender is found the same morning and the brief is ready in an hour. But it still waits for the go/no-go meeting, for finance's Thursday slot and for three sign-offs in turn. 24 working days become 18. Same waste, faster.
Take out the waste first, still without AI, and those 24 days become 9. The tender desk checks the portals daily, go/no-go moves into the morning stand-up, margin rules are agreed once and one approval pack goes to all approvers at the same time. Only then come the agents, and 9 days become 5.
The figures are illustrative, but the order is the point. Agents on a cleaned-up process deliver far more than agents on the old one.
Where the human stays in the loop
The rule from the case: automate what you can undo, sign what you can't. Every step gets three questions.
- Can the decision be undone?
- Does it commit the company?
- Does it need judgment?
What a decision needs
Logging a request commits nobody to anything, so it runs on rules. A price, a delivery date or a promise to a customer does commit the company. That's where a person signs.
Where a person signs, the decision needs three things: a named owner (not a department, and never the agent), the evidence it rests on, and a record you can still audit months later. What the agent read, what it proposed and who signed stays with the decision instead of in someone's inbox.
Could it be faster if nobody decides? Yes. In the case the tender is then done in under one working day, with five commitments nobody signed. A few days faster is not worth that.
Getting started with AI in sales and logistics: a step-by-step plan
Don't start with a tool. Start with one flow, from the first request to the order in the system, and measure it end to end.
- Pick one flow and one owner. For example transport orders from regular customers, or quotes for standard products. One person owns the whole flow, across sales and operations.
- Put lead time next to work time. How long does it take from request to order, and how many hours of work are in it? That gap shows you where to look.
- Take out the waste. Cut approval layers that add nothing, agree on one source of truth for the customer's request and stop re-typing between systems.
- Add agents, with decision points. Let agents read, draft and check. Use the three questions to decide, step by step, where a person signs.
- Measure again, and only then expand.
The whole flow, not one step
I call this a model line: one value stream that you redesign from scratch, with an owner, a team, a mandate and a clear agreement on what done means. Not automating one step, but the whole stream. Once the model line works, you have a pattern for the next flow. Which flow becomes the first model line is a decision for the leadership team. In the boardroom workshop we make that decision together. The broader step-by-step plan is in how to implement AI in your business.
Pitfalls with AI in quotes and order entry
Four things keep coming back.
- Messy master data. An agent that can't find a customer because they're in your ERP three times, with three different addresses, picks the wrong one or gets stuck. Clean up customer, item and rate data before you automate. That work pays for itself even without AI.
- Exceptions nobody wrote down. Every planner knows one customer always wants to load a day early and another never on a Friday. As long as that only lives in people's heads, the agent doesn't know it. Write the exceptions down, or have the agent recognise them and pass them to a person.
- No owner across sales and operations. Sales writes the quote, operations executes the order, and the errors happen at the handover. If nobody owns the whole flow, each department optimises its own part and the waiting in between stays.
- Taking the human out too early. Start with an approval step on every action that leaves the building. Only remove it for steps that run on rules and can be undone.
Key takeaways
- Most of the time between request and order is waiting, not work. That's where the gain is.
- AI reads requests from email, forms and PDFs, drafts quotes and prepares orders in your ERP or TMS.
- A person signs everything that commits the company: the price, the delivery date and every deviation from the standard.
- Agents on the old process make the same waste faster. Take out the waste first, then add agents.
- Start with one flow and one owner, put lead time next to work time and redesign the whole stream.
Ready to rework your own flow from request to order?
I map your flow from request to order, take out the waste and build the agents, with decision points where a person signs.
View AI Implementation