Are AI assistants actually saving your agency time? Here’s how to find out.

Somebody on your team said it last week, probably with a grin. “That would’ve taken me all morning. It took four minutes.” You nodded, because it’s true, and because it’s the kind of sentence that makes the whole AI thing feel worth it.

So why did the Job still ship on Thursday, same as always? If the drafts are that much faster, where did the morning go? There’s a gap between how much time AI assistants feel like they’re saving your agency and how much they are, and there’s a plain two-week check that closes it without buying a single new tool. Which, given the year we’ve all had with tools, may be the most appealing sentence you read today.

My guess before you run it: less than it feels like. AI makes the first draft show up fast, and fast feels like saved time. What you can’t feel is the review, the fix, the “wait, is that number right,” and the twenty minutes someone spent explaining to the client why the tone was off. Those land later, on someone else, and nobody files them in the same mental column as the two hours the draft “saved.” To know for sure, you have to look at the whole Deliverable from brief to delivery and compare it with how that same kind of work went before. Two weeks of paying attention does it.

Why the feeling lies to you

The speed is real. A research brief that used to take a morning comes back in four minutes. A first pass at ad copy appears before the coffee’s done. That’s the part you see, and the part your team talks about, because it’s fun and a little bit magic.

The rest is quiet. Someone reads the draft twice because they don’t trust it yet. Someone rewrites the middle section. Someone catches an invented statistic on the second read and goes looking for one that exists. Someone tightens the whole thing so it sounds like your shop instead of a very polite stranger. Each of those is ten to forty minutes, and none of them gets charged to the AI in anyone’s head, because none of them feels like “AI work.” It feels like Tuesday.

Then there’s the work that didn’t exist before. Prompt tinkering. Trying three tools because the first one was weird today. Checking a source the model cited that turns out to be a very confident fabrication. New effort, invisible for the same reason.

So the owner ends up with a strong feeling and no evidence. The feeling says “we’re faster.” The Job timeline says “we shipped Thursday.” Both can be true at once, which is exactly why you need the check.

Where the time goes

Think of any Deliverable as a chain: brief, draft, review, revise, approve, deliver. AI is very good at the draft link and mediocre at the others. Drop it in and the draft link shrinks while the review link often grows. The chain gets a different shape without getting shorter.

Sometimes it does get shorter, and that’s the whole reason to ask. Research summaries, first-pass structure, alt text, meeting notes, variations on something already approved: those tend to come out ahead. Client-facing writing, anything with numbers in it, and anything that has to sound like a specific human tend to come out even or behind once you count the review.

The useful question isn’t “does AI save time.” It’s “on which Deliverables, for which people, does the whole chain take less effort than it used to.” We looked at the usual suspects in Where does AI actually save time in a small agency’s workflow?, and the pattern holds: the savings live in specific places, and they don’t send a memo.

What “effort” has to include

Count only the minutes someone spent with the AI tool open and you’ll get a number that flatters the tool. Count the effort on the Deliverable, whoever spent it, whatever was open on their screen.

The review counts. The rewrite counts. The Slack thread where the account lead asked “did anyone check this” counts. The client revision that came back because the draft was generic counts too, and that’s the sneaky one, because it shows up a week later on a different person’s plate, wearing a disguise.

You don’t need a timer on every breath. You need each person logging effort against the Deliverable, not against “AI stuff” or “writing.” If your team already logs by Deliverable, you’re most of the way there. If they log by day or by client, you’ll be guessing again, just with better-looking spreadsheets.

The two-week check

Pick three Deliverable types your shop produces over and over. A blog post, a landing page, a monthly report, a set of social captions, a functional spec. Things you’ve done enough times that you already know roughly what they take.

Week one, run them the way you’ve been running them. Whoever normally uses an AI assistant uses it. Everyone logs effort by Deliverable, including review, including revisions, including the client round.

Week two, for one of the three types, have the same people do the work without the assistant. Just for the check. Nobody’s taking anyone’s toys away. Log it the same way.

Then line them up. For each type you now have two numbers you can trust more than a feeling: total effort with the assistant, total effort without, client round included in both. If the client hasn’t come back yet, note that and compare what you have.

One more comparison, and it’s the one most shops skip. Pull the effort you logged on the same Deliverable types six months ago, before the assistants showed up in earnest. That’s your baseline, and it already contains your team’s habits, your clients’ habits, and the ordinary mess of a small firm. If the new number isn’t clearly under the old one, the savings are living in someone’s imagination, rent free.

Reading the result without kidding yourself

Three outcomes are common.

The chain got shorter. Less total effort on the Deliverable, and the quality held. Keep doing it, write down how, and make it the default for that Deliverable type. You now know exactly where the win lives, which puts you ahead of most of the industry and every LinkedIn post about it.

The chain moved. The draft got fast, the review got slow, and the total is about what it was. This can be a training problem, a prompt problem, or a reviewer who doesn’t trust the output yet, and it deserves another two weeks with a better setup before you decide. Our piece on keeping quality control when the team uses AI tools covers the review side.

The chain got longer. More total effort, usually with a client revision hiding in there somewhere. Stop using the assistant on that Deliverable type for now. AI isn’t the villain here. It’s just not paying you back on that kind of work yet.

None of these results is embarrassing. The embarrassing version is not knowing, telling clients you’re an AI-forward shop, and absorbing the extra effort in evenings.

The Bottom Line

The feeling of speed is nice. Seeing it is what lets you plan on it. Run the two-week check on three Deliverable types, count the whole chain including the review and the client round, and compare against what the same work took before the assistants arrived. Where the chain got shorter, estimate with the new number and make it the default. Where it moved or grew, leave the assistant out of that Deliverable for now.

That’s what a hybrid team looks like when it’s working. People and tools on the same Deliverable, effort visible for both, so you know what the whole thing took, even when AI is part of your delivery team. No slogan required.

FAQ

Isn’t two weeks too short to tell?

For a verdict on your whole agency, yes. For one Deliverable type you produce every week, two weeks gives you numbers to argue with, and that beats the nothing you have now. Repeat it each quarter and the picture sharpens.

What if the team resents logging review effort?

People resent logging effort that goes nowhere. Show them the comparison at the end of week two. When they see that their review time is the reason the “saved” time is a myth, they tend to want it counted.

Should we tell clients which Deliverables used AI?

That’s a separate call and it depends on your agreements. What this check gives you is the truth to base that conversation on, instead of a slogan.