Every business owner has now sat through a pitch about AI “transforming” their company. Most have also tried a chatbot that confidently gave a customer the wrong answer. Both experiences are real, and the gap between them is where the money is.
This is our honest map of where AI is paying off for businesses right now — based on what we build, not what gets announced on stage.
Where AI reliably pays off
1. Reading and sorting documents
Invoices, contracts, application forms, delivery notes, emails with attachments. If people in your business spend hours reading documents to pull out a few facts and type them somewhere else, this is the single highest-ROI use of AI today.
Modern models can classify a document, extract the vendor, amounts, dates and line items, check them against your rules, and push clean data into your systems — with uncertain cases routed to a human. We’ve seen turnaround go from days to minutes and the error rate fall, because the model doesn’t get tired at 4pm.
2. First-line answers from your own knowledge
Not a generic chatbot — an assistant grounded in your documentation, policies, product details and past tickets. Done properly (the technique is called retrieval-augmented generation), it answers from your material and says “I don’t know, here’s a human” when it doesn’t. The win isn’t replacing support; it’s that your team stops answering the same forty questions and your customers get answers at 11pm.
3. Summarising and drafting inside existing workflows
Call notes turned into CRM updates. Long email threads summarised for the next person. First drafts of proposals, job descriptions, product descriptions and follow-ups. The key phrase is inside existing workflows — the value comes when it happens automatically in the tools people already use, not when someone has to remember to open a chat window.
4. Routing and triage
Which team should this enquiry go to? Is this review positive or a complaint? Does this lead match our ideal customer? Classification tasks like these are fast, cheap and accurate with modern models, and they remove a surprising amount of human sorting from the front of every process.
Where it usually disappoints
Replacing judgment
Hiring decisions, credit decisions, medical or legal conclusions, anything where being wrong is costly and explanations matter. AI can assist — surfacing information, flagging risks — but handing it the decision creates liability and erodes trust. Keep the human in the loop where the stakes are high.
“Add AI” with no process behind it
Buying an AI feature because a competitor announced one. If you can’t name the task, the person doing it today, and the hours it takes, there’s no ROI to measure and the project drifts. Start with the process, then ask whether AI helps.
Public chatbots with no guardrails
The classic failure: a bot on the homepage that invents a refund policy. Without grounding in your real content, limits on what it will discuss, and a handoff to a person, a chatbot is a reputational risk, not a feature.
The questions to ask before any AI project
- What is the task, and who does it today? If you can’t point at a person and a process, stop.
- How will we know it worked? Hours saved, turnaround time, error rate, response time — pick one and measure it before you start.
- What happens when it’s wrong? Every model is sometimes wrong. The design question is whether a wrong answer is caught (review queue, confidence threshold, human handoff) or acted on.
- Where does our data go? Use providers and configurations that don’t train on your data, or self-hosted models where confidentiality demands it. This is a settings decision, not a mystery.
- Do we have the data at all? You usually don’t need much — but an assistant can’t answer from documentation that doesn’t exist.
How to run a low-risk first project
Pick one document-heavy or question-heavy process. Build the smallest version: ingest, process, output, with a review step for anything the model isn’t sure about. Run it alongside the current process for a few weeks and compare. If the numbers are good, widen the scope; if not, you’ve learned something for a modest cost.
That first project usually takes weeks, not months, and it answers the question every owner actually has — “is this real for us?” — with data instead of a pitch.
If there’s a process in your business that involves a lot of reading, sorting or answering, describe it to us. We’ll tell you honestly whether AI is the right tool for it — and, if it is, what a measured first step looks like.