How to Use AI at Work Without Sacrificing Quality: An Interview With an AI Workflow Specialist

Almost every team I talk to has the same story. Somebody discovered an AI assistant, showed everyone how to use AI at work to get a draft in thirty seconds that would have taken an hour, and the whole office got excited. Then, a few weeks later, a client email went out with a made-up statistic in it. Or a report landed on a manager’s desk that looked polished and said absolutely nothing. The excitement turned into suspicion, and now half the team uses AI quietly while the other half refuses to touch it.

We sat down with an AI Integration and Workflow Specialist who spends most of the week inside companies untangling exactly this problem. Their job is not to sell software. It is to figure out where AI actually belongs in a workflow, where it does not, and how people can learn how to use AI at work without their output getting worse. What follows is our conversation, lightly edited for length.

“Most People Are Asking the Wrong First Question”

When a company brings you in, what is usually going on?

Usually somebody at the top has read that AI makes people faster, so they bought licenses for everyone and waited for the magic. Six months later they are confused, because the numbers did not move and some of the work actually got sloppier.

The first thing I tell them is that they asked the wrong question. They asked, “Which AI tool should we buy?” The right question is, “Which parts of our work are we willing to hand to a machine, and who checks it when it comes back?” That second question is the real starting point for how to use AI at work. It is also boring. Nobody writes press releases about it. But it is the whole game.

The Evidence Behind the Hype

Is the productivity boost real, though? Or is it hype?

It is real, and we have decent evidence now. The study I point to most often is the one Harvard Business School ran with Boston Consulting Group. They gave 758 consultants realistic tasks, some with access to GPT-4 and some without. On tasks that fell inside what the AI was good at, the consultants using it finished about 12% more tasks, worked about 25% faster, and their work was rated more than 40% higher in quality.

That is a huge result. But here is the part people skip. On a task that was deliberately designed to sit just outside the AI’s abilities, consultants who used it were 19 percentage points less likely to get the right answer than the people working alone. The AI gave them a confident, persuasive, wrong answer, and they believed it.

The researchers called this the “jagged frontier.” AI is brilliant at some things and quietly terrible at others, and the line between the two is not where you would guess. That single idea explains most of the quality problems I see when people use AI at work.

Where AI Actually Earns Its Place

So how do you decide what goes to the AI?

I have people sort their weekly work into three buckets. I do this on a whiteboard with the actual team, not in a strategy deck, because the people doing the work know things managers do not.

The Three Buckets

Bucket one is blank-page work. First drafts, outlines, summaries of long documents, meeting notes, turning bullet points into a readable email, rewording something for a different audience, even the first outline of a business plan. AI is excellent here, and the risk is low because a human is going to reshape it anyway. It is also a useful push if you are trying to stop procrastinating on a task you keep avoiding.

Bucket two is structured grunt work. Reformatting data, writing a formula you half remember, cleaning up a spreadsheet or a monthly budget, drafting a standard reply to a common customer question, building a checklist from a policy document. AI is very good here too, but the output needs a spot check, because small errors hide easily in structured work.

Bucket three is judgment work. Pricing decisions, anything legal, anything medical, sensitive personnel matters, strategy, facts you plan to publish, and anything that goes out under your name to a client who trusts you. AI can help you think here, but it should never have the final word. The same goes for personal money decisions: AI can explain how big an emergency fund should be, but your own situation decides the number. This is where the jagged frontier bites hardest.

When teams actually do this exercise, they are usually surprised by how much sits in bucket one. That is where the easy wins are, and it is where I tell people to start when they learn how to use AI at work.

A Real Example From the Field

Give me a real example of bucket one done well.

I worked with a small logistics firm where the operations lead was spending close to five hours a week writing the same kind of shipment delay notices. Every one was slightly different, but the structure was identical. We built a simple routine. She pastes in the carrier update and the customer details, the AI drafts the notice in the company’s tone, and she edits it. Her edit takes about two minutes. The whole thing went from five hours to under one. That kind of win matters most when you are starting a small business and every hour counts.

Notice what did not change, though. She still reads every single one. She still decides what the customer is told. The AI does the typing. She does the thinking.

The Quality Problem Nobody Wants to Name

You mentioned sloppy work. How bad does it get?

There is a name for it now. Researchers from Stanford and BetterUp Labs, writing in Harvard Business Review, called it “workslop.” It is AI-generated work that looks finished but does not actually move the task forward. Around 40% of the workers they surveyed said they had received it from a colleague in the previous month, and each incident took close to two hours to sort out.

Think about what that means. The sender saved twenty minutes. Meanwhile, the receiver lost two hours. Overall, the company lost time, and the relationship between those two people took a hit, because people who receive workslop start seeing the sender as less capable and less trustworthy.

Why Workslop Happens

Why does it happen? People are not lazy on purpose.

No, and I want to be fair to people here. Most workslop comes from three places, and none of them are about whether people know how to use AI at work in a technical sense.

1. Pressure. People are told to “use AI more” without being told what good looks like. So they use it everywhere, including places it does not belong, because usage is what gets noticed.
2. Fluency disguised as quality. AI writes smoothly. Smooth writing feels finished. Your brain reads a clean paragraph and assumes the thinking behind it is also clean. It often is not.
3. People stop checking. This one worries me most. Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers and found that the more confident people were in the AI’s ability to do a task, the less critical thinking they reported putting in. Trust goes up, scrutiny goes down. That is a very human reaction, and it is exactly backwards from what quality needs.

How to Use AI at Work: The Habits That Protect Quality

Let’s get practical. If someone reading this wants to know how to use AI at work tomorrow without embarrassing themselves, what do they do?

I give every team the same short list. None of it is complicated. The hard part is doing it every time.

1. Own the Brief Before You Open the Tool

Before you type anything into an AI, write one or two sentences for yourself. What is this for? Who reads it? What does a great result look like? If you cannot answer that, the AI cannot either. It will just produce something generic, and generic is the first symptom of workslop.

I tell people to imagine handing the task to a smart new hire on their first day. You would never just say “write the report.” You would tell them the audience, the goal, the deadline, and what to avoid. Give the AI the same courtesy.

2. Give It Your Material, Not the Internet’s

The single biggest quality jump I see comes from this. Instead of asking the AI to write about a topic from scratch, paste in your own notes, your data, last quarter’s report, or the client’s actual email. The output becomes grounded in your reality instead of the average of everything online.

Just be careful about what you paste. Most AI assistants are cloud computing services, which means whatever you share leaves your machine. Check your company’s policy on confidential data first. If there is no policy, that is a conversation to have with IT before you share anything sensitive.

3. Treat Every Output as a First Draft From an Intern

This is the mindset shift that fixes most problems. The AI is a fast, tireless intern who has read everything and experienced nothing. Interns are useful. You still read their work before it goes to the client.

Concretely, that means you read the whole thing, not just the first paragraph. You check every number, name, date, and quote against a source you trust. You cut anything that sounds impressive but says nothing. And you ask yourself whether you would be comfortable defending every line if someone questioned it.

4. Verify Facts in a Separate Step

AI tools can invent sources, merge two statistics into one, or state an outdated figure with total confidence. So fact-checking has to be its own step, done on purpose. I have teams highlight every factual claim in a draft, then confirm each one against the original source. If they cannot find the source, the claim comes out. No exceptions.

5. Put Your Voice Back In

Clients and colleagues can feel when something was not written by you. The tone is a little too even. The examples are a little too general. Every paragraph is the same length.

So after the facts are right, I ask people to do a voice pass. Add the specific example only you would know. Cut the phrase you would never say out loud. Break up a long sentence. Say the thing directly. This takes five minutes, and it is the difference between content people trust and content people skim.

6. Keep a Record of What Works

Every team I work with ends up with a shared library of prompts and routines that actually produced good results for their specific work. Not generic prompt templates from the internet. Their own, tested on their own tasks. When something works, write it down. When something fails badly, write that down too, because that is how you find where your jagged frontier sits.

The Role of Experience

Does AI help junior people and senior people equally?

No, and this matters for how you roll it out. A large study by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond looked at more than 5,000 customer support agents. With an AI assistant, agents resolved about 14% more issues per hour on average. But newer and less skilled agents improved by around 34%, while the most experienced agents saw very little gain.

The AI was basically spreading the best practices of top performers to everyone else. That is wonderful for onboarding. But it also means junior people are the ones most likely to accept AI output without questioning it, because they do not yet know what “wrong” looks like.

So my advice on how to use AI at work across different experience levels is simple. Pair them up. Let junior people use AI heavily, and build in a senior review for anything in bucket three. The junior person learns faster, the senior person catches the mistakes, and the company gets both speed and quality.

Winning Over the Skeptics

What about senior people who refuse to use it at all?

I get it. Some of them have been burned. But I usually ask them to try one thing: use AI only to critique their own work, not to create it. Paste in a draft and ask what is unclear, what is missing, and what a skeptical reader would push back on. Most experienced professionals find this surprisingly useful, and it keeps them firmly in control. Once they see the value, they usually start experimenting on their own.

Building Guardrails That People Actually Follow

Should companies write an AI policy?

Yes, but keep it short enough that people actually read it. The best policies I have seen fit on one page and answer five questions:

1. What data can never go into an AI tool?
2. Which tools are approved?
3. Which types of work always need a human review before they go out?
4. When do we tell clients or readers that AI was involved?
5. Who do I ask when I am not sure?

If you want a more formal foundation, the NIST AI Risk Management Framework and its Generative AI Profile are free and very thorough. You do not need to adopt the whole thing. But reading through the risks they list, such as confabulation, data privacy, and information integrity, is a good way to make sure your one-page policy is not missing anything important.

The Manager’s Role

What about the managers? What is their job here?

Managers shape how people use AI at work more than any policy does. If a manager praises someone for “using AI a lot,” people will use it a lot, whether or not it helps. If a manager praises someone because “the analysis was sharp and it was on time,” people will use AI where it helps and skip it where it does not.

I also tell managers to model the behavior. Say out loud in meetings, “I used AI to draft this, and here is what I changed.” That makes review normal instead of something people hide. The worst environment is one where everyone uses AI secretly, because secret use is never reviewed. That matters even more on remote teams, where staying productive working from home already depends on clear expectations.

Microsoft’s Work Trend Index has been tracking this for a few years now, and the pattern is consistent: leaders tend to be further along with AI than their teams, and employees often bring their own tools when the company does not provide clear guidance. That gap is where quality problems grow. Close it with honest conversation, not just more licenses.

Measuring Whether It Is Working

How do you know if AI is actually improving work rather than just speeding it up?

When teams use AI at work, speed is the easiest thing to measure, which is exactly why it is dangerous to measure only speed. I have teams track three things together:

• Time saved. Roughly how long did this take before, and how long does it take now?
• Rework. How often does AI-assisted work come back with corrections from a manager, a client, or a colleague? If rework goes up, your time savings are fake.
• Outcome quality. Are clients happier? Do fewer errors reach customers? Are proposals still winning at the same rate?

If time saved goes up and the other two stay stable or improve, you have found a good use case. If time saved goes up and rework goes up with it, you have found a workslop factory, and you should pull that task back into bucket three until you fix the process.

What is a warning sign that someone is leaning on AI too much?

When they cannot explain their own work. If I ask someone why a report reached a particular conclusion and they say “that is what the AI said,” we have a problem. The person should always be able to walk me through the reasoning in their own words. If they cannot, the AI did the thinking, and nobody checked it.

Looking Ahead

Where do you see this going over the next couple of years?

The tools are moving quickly toward what people call agents: AI that does not just draft something but takes actions, like updating records, sending messages, or running multistep tasks on its own. That is exciting, but it raises the stakes on everything we just talked about. A bad draft gets caught in review. A bad action might already be done by the time anyone notices.

So the habits behind how to use AI at work matter even more, not less. Clear briefs. Human review at the points that matter. Honest measurement. Teams that build those habits now, while the stakes are lower, will be the ones who can safely hand over bigger tasks later.

Last question. If you could give one piece of advice to someone learning how to use AI at work, what would it be?

Stay the author. Use AI for the typing, the drafting, the reformatting, and the first rough pass at thinking. But keep your name on the judgment. The moment you stop reading what goes out under your name, quality starts slipping, and it slips quietly. The people who do well with AI are not the ones who use it the most. They are the ones who never stop checking.

Frequently Asked Questions

What is the best way to start using AI at work?

Start with low-risk, repetitive writing tasks like meeting summaries, first drafts of routine emails, and outlines. These give you quick time savings while a human still reviews everything before it goes out. Once you know how the tool behaves, expand into more structured tasks.

How do I use AI at work without making mistakes?

Treat every AI output as a first draft. Read the full response, confirm every fact and number against a trusted source, and rewrite anything that does not sound like you. Keep judgment calls, sensitive topics, and final decisions with a human. The HBS jagged frontier study shows AI can hurt accuracy on tasks just outside its strengths.

What tasks should not be given to AI at work?

Avoid handing AI final decisions on legal, medical, financial, or personnel matters, and never publish facts from AI without verifying them. Also avoid pasting confidential client or company data into tools your organization has not approved.

What is AI workslop?

Workslop is AI-generated work that looks polished but lacks real substance, pushing the effort of fixing it onto the person who receives it. Research published in Harvard Business Review found it is common and costs recipients close to two hours per incident.

Does using AI at work make people less skilled?

It can if people stop thinking critically. A Microsoft Research and Carnegie Mellon survey found that higher confidence in AI was linked to less critical thinking. Using AI to critique your own work, rather than replace it, helps keep your skills sharp.

Who benefits most from using AI at work?

Less experienced workers often see the largest gains. A study of customer support agents found AI raised productivity by about 14% overall and around 34% for newer workers.

Does my company need an AI policy?

Yes. A short, clear policy should cover what data cannot be shared, which tools are approved, what work needs human review, and who to ask with questions. The NIST AI Risk Management Framework is a helpful free reference.

References

1. Dell’Acqua, F., McFowland III, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., and Lakhani, K. “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality.” Harvard Business School Working Paper, 2023. hbs.edu
2. Harvard Business School AI Institute. “Navigating the Jagged Technological Frontier.” aiinstitute.hbs.edu
3. Niederhoffer, K., Kellerman, G. R., Lee, A., Liebscher, A., Rapuano, K., and Hancock, J. T. “AI-Generated ‘Workslop’ Is Destroying Productivity.” Harvard Business Review, September 2025. hbr.org
4. Harvard Business Review. “Why People Create AI ‘Workslop’ and How to Stop It.” January 2026. hbr.org
5. Lee, H. P., et al. “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.” Microsoft Research / CHI 2025. microsoft.com
6. Brynjolfsson, E., Li, D., and Raymond, L. “Generative AI at Work.” National Bureau of Economic Research, Working Paper 31161. nber.org
7. MIT Sloan. “Workers With Less Experience Gain the Most From Generative AI.” mitsloan.mit.edu
8. Microsoft. “2025 Work Trend Index Annual Report.” news.microsoft.com
9. National Institute of Standards and Technology. “AI Risk Management Framework.” nist.gov