Category: AI Tools for Business

What better way to learn about chatbots than to interview one? For my AI Tools for Business class, I asked ChatGPT to play the subject matter expert. We discussed customer service, helpdesks, business costs, and the risks of letting a confident machine represent your organization.

But the interface was part of the lesson. I was using GPT-Live, OpenAI’s new voice model family, introduced in July 2026. Its defining feature is full-duplex conversation: it can listen and speak at the same time. The two versions introduced were GPT-Live-1 and GPT-Live-1 mini. OpenAI’s announcement explains the change.

That sounds like a small improvement until you think about how people actually talk. We pause without surrendering the floor. We offer little acknowledgments while somebody else speaks. We interrupt because we missed a word, disagree with an assumption, or need the explanation to slow down. Sometimes we need a moment to find the phrase we wanted. My students are familiar with this particular feature of their instructor.

Earlier voice systems depended on detecting when you had finished speaking. Guess too early and the machine cuts you off. Guess too late and everybody waits. GPT-Live continuously processes incoming audio while generating speech, allowing the model to decide whether to speak, listen, or pause. OpenAI’s engineering account describes how it removed the separate turn detector from the live audio path.

In my demonstration, I stumbled over “by the dozen” and “in droves.” The chatbot supplied the phrase with the hilarious aside “English. Lovely language.” LOL. Later, the audio broke up during an explanation of chatbot return on investment. I interrupted and asked it to start again at a more beginner-friendly level. It did. The technology still had an imperfect moment, but conversation gave me a straightforward way to recover.

There is another interesting piece behind the voice. GPT-Live can delegate harder reasoning and searches to another model while remaining available for conversation. The system separates the immediate demands of talking from work that takes longer. That helps explain how a responsive voice interface can connect to more substantial AI capabilities. OpenAI’s architecture explanation covers this division of labor.

For work and learning, I see possibilities in rehearsing a presentation, practicing a difficult customer conversation, or talking through an unfamiliar concept. You can steer as you go: “Give me an example.” “That assumes too much background.” “Wait, let me finish.”

The charm deserves scrutiny, too. GPT-Live makes talking to a computer feel more natural. We still have to decide whether the computer has said anything worth believing. A fluent explanation can make it sound authoritative. Neither establishes that its answer is correct. In the interview, I assigned the chatbot the role of expert; that assignment did not give it credentials. My practical suggestion is to try a short conversation about something you know well. Interrupt, change direction, ask for clarification, and check its claims. Judge both the interaction and the information.

ChatGPT now asks us to make a decision before we ask it to help us make decisions: to Chat or to Work? That is the question! Apparently even the chatbot needs to know whether this conversation could have been a one-line email.

The basic rule of thumb: Use Chat to think with ChatGPT. Use Work to delegate a job to ChatGPT. Chat is useful when you want an explanation, a brainstorm, a comparison, or help finding the right words. You are directing the conversation as it develops. Work is designed for a larger assignment with an outcome you can deliver and share: a report, a presentation, a spreadsheet, or even an application prototype/demo. That is the practical distinction in OpenAI’s guidance.

“Help me think through an employee onboarding session” belongs comfortably in Chat. You can discuss the audience, reject boring activities, and figure out what people actually need to learn. “Use these policies and notes to create the onboarding presentation, check that every required topic is covered, and return an editable PowerPoint” is a Work assignment. The useful distinction is how much of the process you want to hand over. Both modes use the same model. So what actually changes?

The (multimodal) Large Language Model (LLM) is only part of the system. The software program around it determines which tools it can use, what information it can retain during a task, and how it continues after an intermediate result. Developers call this surrounding machinery an agent harness. An agent loop can repeatedly ask the model what to do next, execute a tool call, return the result, and continue until it reaches a stopping point. Work provides an environment for carrying out an assignment across multiple steps. Cloud Work can continue supported tasks while you are away; local Work can use files and applications made available through the desktop app.

When the assignment is fuzzy, talk it through with Chat. Once you can describe the deliverable and what would make it acceptable, Work becomes easier to direct and evaluate. OpenAI recommends reserving Work for substantial tasks and using Chat for quick questions and short rewrites. Delegation still leaves you with a job. Review the data sources, check and improve output quality, and decide whether the result meets your requirements. Choose Chat when you want to work through the thinking together. Choose Work when you can assign the job, have clear requirements, and intend to eventually share an artifact with others.

Editable writing blocks may be the best thing since sliced bread. That sounds like a ridiculous amount of enthusiasm for a text Edit box, but the feature removes one of the most persistent frustrations of writing with ChatGPT: the awkward distance between discussing a document and actually editing it.

OpenAI’s older Canvas interface pointed in the right direction. A longer draft opened in a separate pane, giving the document more room and providing editing tools outside the ordinary chat stream. It was useful, but it also felt clunky. The conversation lived in one place and the document lived in another. Revising often felt like stepping out of the discussion and into a second workspace.

The newer writing block, with its Edit control, is far more natural. The document appears directly inside the conversation. I can ask ChatGPT for a draft, click Edit, rewrite a sentence myself, delete an awkward paragraph, or add a detail that only I know. Then I can continue the conversation from the version I actually edited. There is no need to copy the entire document into another program or paste the revised version into a new prompt. This makes iteration feel less like repeatedly ordering new drafts and more like genuine collaborative editing. I can preserve what works, fix small problems myself, and ask ChatGPT to focus on the parts that still need help.

For an IT example, put a draft network-change plan in an editable writing block. Correct the device names, IP addresses, maintenance window, and rollback steps directly. Then ask ChatGPT to review the edited plan for unsupported assumptions, missing validation steps, unclear instructions, and anything that would confuse a junior technician. The human supplies the operational truth; ChatGPT helps inspect and improve the document.

The first response is still only a draft. The difference is that the draft no longer feels trapped inside a chat response or exiled to a separate Canvas. The writing block keeps the document and the conversation together. That small interface change makes revision faster, clearer, and much more pleasant—and keeps human judgment in control.

The first response is slop. It is not the one to publish or share. It is a shot in the dark. It must be improved, edited, and/or regenerated with better prompting, and often by adding more context, before letting anybody else see it.

Iteration is the difference between AI slop and AI-human quality. The latter is the one that genuinely saves time and advances work tasks. After the first response, pause, tell ChatGPT exactly what you like and dislike. “The opening works, but the middle sounds inflated.” “Keep the example and replace the conclusion.” “This is too polite. Say what the problem actually is.” The model needs your expert steering and context awareness. If you find yourself correcting ChatGPT the same way multiple times, consider adding it to your Custom Instructions (in Settings -> Personalization). Your future self will thank you.

Ask for multiple tries when you do not yet know what quality looks like. Generative AI is based on statistical probabilities so even the exact same request will produce different results, some better than others. Or you can request three genuinely different versions, structures, or approaches—not three lightly rearranged copies of the same draft. Compare them. Keep one opening, another explanation, and a third conclusion. A failed attempt can still show you what you do not want.

When a response misses, do not merely say, “Try again.” Name the failure. Identify the requirement it ignored, the assumption it invented, the tone it misunderstood, or the detail it must preserve. If part of the answer is correct, say so. Otherwise, the next attempt may repair one section by breaking another.

Sometimes the conversation gets stuck in a groove. Stop. Wait. Read the result later with fresh eyes. Start a new chat with a cleaner prompt, provide a better example, or change the order of the instructions. You can also try a different model. A more capable model may help with difficult reasoning or synthesis; another model may simply approach the wording or structure differently. Compare the results against your actual requirements rather than assuming that newer, larger, or slower thinking automatically means better.

Iteration is not endless polishing. Watch out for AI’s tendency to worsen your perfectionism. Short and authentic beats ‘shock and awe’ almost every time. Imperfection can make you sound more human and real. So keep the loop of iteration/improvement/regeneration short, tidy and still human. Experiment, change one or two things, and try again. Done. Eventually the document meets the requirements, the remaining flaws do not matter, or more revisions stop contributing value.

ChatGPT can assist with consequential work, but it cannot assume responsibility for the result. Check important facts, calculations, citations, assumptions, privacy risks, and recommendations before using an output. Document verification when the stakes are high.

Do not upload confidential, regulated, or personally identifiable information unless the organization has approved the tool and the handling process. For an IT example, before using ChatGPT to prioritize security incidents, verify the severity calculations, protect names and device identifiers, check the supporting evidence, and require a qualified analyst to approve any action that would isolate a device, disable an account, or reset credentials.

Human review is not a ceremonial final click. A person remains responsible for decisions affecting customers, employees, money, access, compliance, or business operations. Good governance means deciding in advance what the AI may recommend, what it may draft, what it may never do on its own, and who is accountable for checking the result.

A successful prompt is useful once. A documented workflow can be useful every week. Turn repeatable ChatGPT work into a process that specifies inputs, steps, quality checks, outputs, ownership, and the points where a person must review or approve the result.

An IT service team, for example, could create a weekly ticket-triage workflow that imports de-identified support records, checks required fields, groups tickets by affected service, counts recurring issues, flags possible security incidents, drafts a manager summary, and routes urgent items using an approved template and review checklist. Projects, skills, templates, and approved business tools can help another team member run the same process consistently.

The real productivity gain comes from improving the whole workflow, not merely generating text faster. Measure time saved, error reduction, consistency, and output quality. If the workflow saves ten minutes but creates an hour of verification work, it is not an improvement. A repeatable AI process should make the work more dependable as well as faster.

Uploading a spreadsheet is not the same as defining an analysis. Start with the business question, then provide the relevant spreadsheets, reports, PDFs, logs, or customer feedback. Ask ChatGPT to inspect data quality, identify patterns, calculate appropriate results, create useful visualizations, and explain the evidence in plain language.

For an IT example, upload a de-identified spreadsheet of monthly help-desk tickets by service, location, priority, resolution time, and reopen status. Ask ChatGPT to flag missing or inconsistent values, calculate month-over-month changes, identify services with rising ticket volume or repeat incidents, create two suitable charts, and produce a one-page briefing with three evidence-based actions.

Require a useful deliverable—a cleaned table, dashboard outline, executive summary, or prioritized action list—rather than a collection of observations. Then spot-check the calculations and make sure the analysis does not confuse correlation with cause. Useful analysis connects data to a business decision without pretending that the data says more than it does.

Research with ChatGPT should begin with current, trustworthy sources. Ask it to identify publication dates, authors or issuing organizations, methods, versions, and limitations. Request citations or links, distinguish established facts from interpretations, and compare conflicting evidence rather than forcing a single conclusion.

This discipline is especially important in IT, where vendor documentation, product versions, vulnerabilities, and recommended configurations change. Suppose a manager is deciding whether to replace an on-premises file server with a cloud document platform. Ask ChatGPT to compare current vendor documentation, government security guidance, and independent technical analysis; identify which claims come from vendors; explain where the sources disagree; and list the questions that still require a pilot or local test.

The goal is not a pile of links. The goal is a decision-ready summary that explains what is known, what remains uncertain, why the differences matter, and what action the evidence actually supports.

A useful ChatGPT prompt starts by defining the job. State the goal, audience, context, source material, constraints, and desired output before asking ChatGPT to begin. Describe the work in concrete terms, provide relevant background and reference samples, and specify requirements such as length, tone, format, deadline, evaluation criteria, and prohibited content. Explicit constraints reduce guesswork and avoidable revisions.

That matters in IT because a vague request can conceal important operational choices. Instead of asking, “Write an MFA rollout plan,” ask, “Create a 60-day MFA rollout plan for a 150-person company using Microsoft 365. Use the attached device inventory and support-hours schedule. Include pilot users, communications, enrollment support, exception handling, rollback steps, owners, deadlines, and success metrics. Do not assume that every employee has a company-owned phone.”

The longer prompt is not automatically better because it is longer. It is better because it tells ChatGPT what problem it is solving, what evidence it may use, what boundaries it must respect, and what a usable result looks like.