Putting AI to work the right way

How to read AI claims

The AI conversation runs hot in both directions. A simple filter helps: ask where a claim came from before you decide what to do about it.

Every claim in this document is marked. It is worth applying the same marking to anything else you read this year.

This matters more than usual with AI, because many of the loudest voices teaching it also sell AI training. Their numbers are not necessarily false. They are simply unaudited, and they were chosen because they sound impressive.

What is worth doing

Three ideas in wide circulation are genuinely good. They cost nothing and they work right away.

What to discount

Four claims come up constantly. Here is how much weight to give each one.

The prompt technique

This takes about two minutes and works in any of the major tools. It is the fastest improvement available to you.

Checking the work

This is the part most AI advice leaves out, and it is what separates a tool that saves you money from one that quietly costs you money.

Software teams building AI products use a practice called evaluation. It sounds technical. It is not. The idea is simply that you cannot judge AI output by how impressive it feels, so you compare it against work you already know is good.

Four parts, about an hour to set up

• A reference set. Ten to twenty real pieces of your own work that you were happy with. Proposals, client updates, whatever you produce most often.
• A written standard. Decide before you automate anything. What does a good version contain? What would make you reject it? Write it down, because standards you keep in your head drift.
• A check on the first ten outputs. Put each one next to your standard and count how many pass.
• A record of the failures. Keep them. They tell you what to fix, and they are the most useful thing you will collect in the first month.

If eight out of ten pass, keep the workflow. If four out of ten pass, the task is not ready and more prompting will not fix it. That single number is worth more than any productivity claim you will hear this year.

The accuracy problem

AI tools produce confident, well-formatted, professional-looking output that is sometimes simply wrong. This is not an occasional bug. It is a known property of how these tools work, and the professional consequences are now well documented.

Two details from that record matter to you even though you are not a lawyer.

First, the errors are getting harder to spot. Early cases involved obviously invented sources. More recent ones involve real sources with fabricated quotes attached, or real sources cited for things they never said. A quick glance will not catch those.

Second, specialist tools do not solve it. A peer-reviewed study from Stanford's RegLab found error rates of roughly 17 percent to 34 percent even in paid legal research tools built specifically for the job. Using a more expensive tool does not remove your obligation to check.

The practical rule is short. Anything factual that leaves your desk with your name on it gets verified by you at the original source. Statistics, dates, names, quotes, citations, prices.

Your data

The common advice is "do not use the free version, pay for the business version." The instinct is right. The line is drawn in the wrong place, and the difference matters if you handle client information.

So upgrading from free to a paid personal plan does not, by itself, change anything about your data. Two things actually do.

• Open the privacy settings and look. In whichever tool you use, find the setting about training or improving the model. Set it deliberately rather than accepting whatever it defaults to.
• Move to a business or team plan if client confidentiality is involved. That is where the contractual protection sits. It is usually a modest increase in cost.

These terms change often, and they differ between vendors. Treat this as a reason to check your own account settings, not as a permanent answer.

Your client agreements

This is the question almost nobody raises, and it is the one most likely to cause you a real problem.

Many consulting agreements contain a confidentiality clause. Those clauses commonly say that client information will not be shared with third parties without permission. Pasting a client's document into an AI tool is, on a plain reading, sharing it with a third party.

Whether that actually breaches your agreement depends on how the clause is written and which plan you are on. The point is that it is a question with a real answer, and most people never ask it.

Three things to do

• Read your own confidentiality clause. Look for language about third parties, subprocessors, or written consent before disclosure. If it is strict, you may need permission before using AI on that client's material.
• Decide what you will tell clients. Some now ask directly whether AI was used. Having an answer ready is better than being asked cold. A short line in your engagement letter saying you use AI tools as an assistant, with client material handled under a business plan, covers most situations and reads as professional rather than defensive.
• Ask your insurer. If you carry professional liability or errors and omissions cover, ask whether AI-assisted work product is covered and whether any exclusions have been added. I cannot tell you what your policy says. I can tell you it is worth a short email to your broker.

None of this is a reason to avoid AI. It is a reason to spend twenty minutes on it before a client asks.

Who owns the output

If you deliver work to clients and your agreements transfer ownership of that work, this affects you.

The Office is equally clear that using AI as an assisting tool is fine. Using it for ideas, for editing, or as part of a larger piece you wrote does not affect protection for the work as a whole. They have registered over a thousand works containing some AI-generated material.

What this means in practice for you is straightforward. If a deliverable is largely AI output with light editing, you may not own what you are transferring to your client. If it is your thinking, your structure and your judgment with AI helping along the way, you are on solid ground.

Two habits keep you there. Make sure the substance and structure are genuinely yours. Keep a rough record of which parts of significant deliverables were AI-assisted, in case ownership is ever questioned.

What not to hand over

There is real enthusiasm at the moment for giving AI tools direct access to your inbox so they can read, reply and send on your behalf. This is the one recommendation worth turning down.

The safe version costs you almost nothing.

• Draft with AI. Send it yourself. Reviewing a draft takes a few seconds. Un-sending a message is not possible.
• Read-only access is fine. Having a tool summarise your inbox is a different level of risk from letting it act on your inbox.
• Anything you cannot undo needs a person. Sending, publishing, paying, deleting, agreeing to something.

What this does to your pricing

This is a business question rather than a technical one, and it catches consultants out.

If you bill by the hour and AI makes you three times faster at a task, you have just cut your own revenue on that task by two thirds. You did better work in less time and you got paid less for it. Hourly billing punishes you for getting more efficient.

There is no single right answer here, but there are three common ones.

• Move toward fixed prices or project fees. The client buys the outcome. Efficiency gains stay with you. This is the most common response and usually the right one.
• Keep hourly billing and take on more clients. Works if you have demand waiting. Does not work if your pipeline is the constraint.
• Keep hourly billing and use the time for higher-value work. Spend the recovered hours on the parts of the engagement clients value most.

Whichever you choose, decide it deliberately. The default outcome is a quiet pay cut.

Why saved prompts stop working

Worth knowing in advance, because it will happen and it is confusing when it does.

AI vendors update their models regularly, often without much announcement. A prompt you refined against one version can produce noticeably different results against the next. Nothing broke and you did nothing wrong. The tool underneath you changed.

Two habits handle this. Keep your saved prompts in a plain document you control, not inside a chat tool's memory, so you can read and edit them. Re-run your reference check every few months, using the same examples from the section above. If quality has slipped, you will see it in the numbers rather than sensing it vaguely.

Your first workflow

One concrete project, about an hour, difficult to get wrong.

Questions to answer first

Six questions, about fifteen minutes on paper. They prevent most of the expensive mistakes.

• What do you do every week that you would happily pay someone else to do?
• What does a good version of that output look like? Can you point to a specific example?
• Does any of this touch client information you have agreed to keep confidential? If so, what does your confidentiality clause actually say?
• Are you on a personal plan or a business plan, and have you opened the privacy settings?
• If your deliverables get faster to produce, what happens to how you charge for them?
• What would you do with the time if it worked?

What to expect

The honest summary is this. AI is a meaningful productivity improvement and a genuine disadvantage to ignore. It is not a business model on its own. The people describing it as one are usually selling something. What you were already good at is still what clients are paying for.