AI
Generative AI in operations: 6 practical uses beyond the chatbot
When people talk about AI in a company, the first image is the chatbot on the website. It has its place, but it is only one application, and not always the most useful. The most consistent gains tend to be in internal, unglamorous tasks that eat up hours of the team's time every day.
What generative AI does well
Language models are good at one specific kind of work: reading text, understanding what is in it and producing other text from it. Summarizing, classifying, extracting, rewriting, answering based on source material. Before them, these tasks required a person, because no fixed rule could handle the variety of language.
This widens what can be automated. The McKinsey Global Institute estimated in 2017, in its report "A Future That Works", that about 60% of occupations had at least 30% of activities that could be automated with the technology of the time, and that fewer than 5% could be automated entirely. The study predates today's models, so it serves as a floor for reference, not as a measure of the present. The conclusion that still holds is the second one: technology changes pieces of the work, and almost never the whole job.
What AI does not do well: guarantee a fact is correct without a source, do math accurately without a supporting tool and decide sensitive cases. The uses below respect those limits.
1. Triage of what comes in
Emails, messages and forms arrive mixed together: quote requests, complaints, questions, résumés, billing. Someone reads everything and routes it. A model reads each message, classifies it by topic and urgency and delivers it to the right queue, already with a one-line summary.
It is a good first project. A mistake is cheap, because a message in the wrong queue is fixed in seconds, and the gain shows up in first response time.
2. Summaries of conversations and meetings
After a sales call or a customer meeting, someone should record what was agreed. They rarely do it well. Automatic transcription and summarization create the CRM record: what the customer wants, objections, next steps and owners.
Salesforce's State of Sales research (2022) indicates that sales reps spend about 28% of their week selling. Part of the rest is record keeping. This is the kind of task worth taking off their hands, as we discuss in How to build a CRM your team actually uses.
A necessary precaution: tell participants about the recording and define what is kept. In many cases, keeping the summary is enough, not the audio.
3. Extracting data from documents
Invoices, contracts, receipts, forms filled out by hand. Someone reads them and types the fields into a system. A model extracts the data in a structured format, and the person checks only what came back with low confidence or outside the pattern.
The rule here is never to skip checking amounts and dates until you have measured the accuracy rate on your own material. The gain comes from swapping typing for review, which is faster and less tiring.
4. Drafting replies and documents
Proposals, customer replies, product descriptions, meeting minutes. The model writes the first version from the case data and a template approved by the company. The person reviews and sends.
It works well when the text follows a pattern and poorly when it requires taking a position or negotiating. Never let the model make up a price, deadline or condition: that data should come from the system, not from the model's "memory".
5. Searching the internal knowledge base
Where is the refund policy? What is the procedure for changing a plan? Instead of asking the longest-tenured coworker, the person asks an assistant that answers based on the company's documents and shows where the answer came from.
This use reduces interruptions and speeds up training for new hires. It depends entirely on the quality of the material, which is the subject of How to train an AI agent on your company's knowledge base.
6. Checking and review
The model as a second pair of eyes: it compares a contract with the standard template and points out differences, checks that an order has all its data before moving on, verifies that a text follows the brand's style guide. It does not decide. It flags what deserves human attention.
How to choose where to start
The same criteria as for any automation apply here, with one more.
- Volume: does the task repeat many times a week?
- Cost of error: if the AI gets it wrong, what happens and who notices?
- Available material: are there examples, documents and rules to guide the model?
- Human review: can you keep a person checking at the start?
Triage and summaries almost always score well on all four. For the overall view on prioritization, see What to automate first.
The best first use of AI is one where a mistake becomes a ten-second fix, not an apology to the customer.
Precautions that apply to every use
- Personal data. Know what is sent to the model, for what purpose and how long it is kept. Use the business versions of the tools and read the data terms of use.
- Facts come from a source. Price, deadline, stock and rules come from the system or an official document. The model writes, it does not invent.
- Test before you release. Build a set of real cases with the expected answer and run the model on them after every change.
- Cost under control. Every use has a cost per run. At volume, this needs to be measured, as we show in What it costs to run an AI agent.
- A person in the loop. At the beginning, someone reviews. Autonomy grows as accuracy is measured.
Where to start
Pick a reading-and-classifying task that takes time today, set aside fifty real examples and test it. Within a week you know whether it works in your context, without having bought any platform.
If the question comes earlier, about where the operation is losing time, delpra™'s Operational Maturity Assessment answers it in 4 minutes.
Written by the delpra™ teamProcesses, automation and AI so operations scale efficiently.