The Brain: 2.2 Times a Consultant's Output in 24 Workdays

What Is the Brain?

The brain is the AI system we install for a client. You ask it for work in plain language, and it builds the piece: a workbook, an email draft, a deck, a proposal or a PDF. It files what it makes and records finished work on the client and project notes. You keep the big calls.

The brain ran about 63 steps on average for each request, and the consultant sent about six requests in a working session.

We sell the brain, so read this as a vendor's report. We tried to make it a fair one. We recounted each figure in a separate pass, and a skeptic tried to explain each change away with the season, holidays, team edits and busy client weeks before a number went into the report. We put the costs and weak spots near the end, next to outside studies with results unlike ours.

How Did We Measure?

We counted output from the consultant's Outlook, synced SharePoint files, Teams and calendar, plus the brain's own records. We compared the 24 working days after the install with their 12-month normal, scaled to 24 weekdays, and with the same weeks last year. We give email and documents equal weight in the output measure.

Output has two halves of equal weight. For the email half we count the messages the consultant sent and the words they wrote. For the document half we count the files they saved and the new ones they started. We set a normal stretch from the 12 months before the install at 1.0, so 2.0 means twice the usual work.

Each measure comes in two versions. The total covers everything the consultant shipped. For the by-hand version we take the brain's work out, which leaves what they made themselves. The gap between the two is the brain's share.

We left requests to the brain out of the working-day figures, so a 6 a.m. request doesn't count as the start of the day. First and last activity mean the first and last email sent, file saved or Teams message posted on a weekday.

We measured quality a different way. A scorer rated eight pairs of documents, one from before the brain and one made with it, on design, polish, fit for the reader, clarity and substance. The same scorer rated six strategy pieces from before against six from after. The scorer helped make the after versions, so read the scores as an informed opinion.

How Much More Did They Ship?

Over 24 working days the consultant shipped 2.2 times their normal output, about 52 days of work. They made 1.1 of that themselves, a little above their normal, and the brain made the other 1.1. With an adjustment for file dating the figure is 1.9 times, and against the same weeks last year it is 2.5.

The top row of the square chart holds the 24 working days since the install. Below it, at the consultant's old pace, they did about 26 days of work themselves and the brain did about 26 more.

A single month could be a fluke, so we set it beside the history in the next chart. Each dot is a four-week block. Before the install, the consultant's blocks ranged from 0.56 to 1.38 times normal. The first block after the install comes to 2.16 times, well above any block since November 2024.

The headline number depends on the base. Against the 12-month normal it is 2.2 times. A file the consultant saves again moves from its original week into a recent one, so we ran a version that allows for this, and it comes to 1.9 times. Against the same weeks last year, which ran below normal at 0.9, it is 2.5 times. We quote 2.2 and show the range beside it.

Could the Season Explain It?

Some of the rise likely came from the season. The rest of the consultant's team rose about 75% over the same weeks, so this autumn ran busy for the whole team. The consultant's total rose 116% in the latest four weeks while their by-hand output rose 9%. Last year the same weeks ran below normal, at 0.9.

Each point on the line chart is a four-week average, from November 2024 on. Before the install, the consultant's line moved between 48% below normal and 38% above it. The same weeks of 2025, shaded on the chart, sat below normal. After the install the line rises past its earlier peak.

With the team up 75%, a large share of the extra work likely came with the season. Over the same weeks the consultant's by-hand output rose 9%, far less than the team's. One reading fits both numbers: in a busy month, the brain took on work the consultant would have done by hand. If that reading holds, a part of the brain's 1.1 replaced their work instead of adding to it, and the true gain over a brain-free autumn sits below 2.2. We can't settle it from the counts.

Where Did the Extra Output Come From?

Most of the jump came in documents, and most of that came from the brain. Decks rose 680% in total and 1% by hand. New documents rose 224% in total and 3% by hand. Email moved less: words written rose 30% in total, and the number of emails sent fell 1%.

By hand, the consultant's work grew most in spreadsheets: their spreadsheet saves rose 55%, against 39% for files of all kinds. The brain's workbooks may have given them more to work on in Excel. We can see the rise in the counts, but we can't see the cause.

Email moved little. The consultant sent 1% fewer emails than normal, wrote 30% more words in them and attached a file 36% more often. Email drafts made up about a fifth of the brain's finished pieces, yet the number of messages leaving the outbox held flat. The brain changed what went into the email more than how much email went out.

Meetings rose 28% over the same weeks and hit a 12-month high. With more meetings the consultant had fewer hours to make things, so the output rise came in a month with less room for it.

Did It Give Time Back?

No, the consultant's day got longer. Their median first activity moved from 8:30 to 7:20, and their last stayed near 6:44. Active hours rose from 8.4 to 9.1. Off-hours work fell from 9.3% to 6.7% of activity, and the median reply took 43 minutes against 36.

You might expect a tool that does work for you to shorten your day. The consultant started about 70 minutes earlier and finished at the same time as in a normal stretch. In the 34 days before the install they finished at 6:18, so against that stretch the day ran longer at both ends.

In the same weeks last year the consultant started at 8:56, later than normal, so we can't pin the early start on the season alone. From the logs we can't tell whether they started early to work with the brain or to keep up with clients.

A smaller share of activity fell outside working hours: 6.7%, against 9.3% in a normal stretch and 11.3% in the same weeks last year. Replies slowed by seven minutes on the median, though 82% still went out within four hours. The consultant worked longer days and fewer evenings.

What Did They Use It For?

The consultant used the brain on every weekday but one, at about six requests per working session. Workbooks and email drafts led the finished pieces at 19% each, ahead of decks at 15% and PDFs at 13%. At least one piece in seven went to a client or a partner.

The consultant sent most requests on weekdays, and about 9% of the period's requests on the busiest day, September 15. Monday, September 21 was the one weekday without a request. They also sent some on Saturday, October 3, so they used it outside the work week too.

We also counted the share that reached clients, since we learn little from a deck that stays in a folder. At least one finished piece in seven went to a client or a partner,.

Is the Work Better?

On eight before-and-after pairs, document scores rose from 4.6 to 7.8 out of 10, a 70% gain. Design and polish gained most, at 103% and 96%. Substance gained least, at 19%. The scorer helped make the after versions, so read these as an informed opinion on a small sample.

Fit for the reader rose 88% and clarity 69%. A client feels those two, because people read a document that fits them. Substance rose from 6.0 to 7.1, the smallest move from the highest starting point. The consultant's documents already had content, and the brain improved how it looked and read.

I read the gap between design and substance as the main risk in this data. A reader who sees a polished deck may assume the analysis behind it had the same care, even when the thinking is thin. The brain raised the floor on looks for this consultant, and the substance stayed their job.

Did Strategy Improve?

Strategy scores rose from 3.6 to 6.3 out of 10, on six pieces from before against six from after. Evidence gained most, from 2.7 to 6.7. Originality gained least, from 5.5 to 6.3. The consultant kept the big calls, and the best pieces came where the brain had rich input.

In our reading, the biggest gains landed on the parts of a strategy piece that take legwork. Evidence rose 150%, the recommendation 119% and actionability 116%. The smallest landed on the parts that take judgment, with options up 39% and originality up 15%. The brain helped the consultant back a call and spell it out. It added less to the range of calls on the table.

The best pieces came where the brain had rich input to work from, and with a thin brief it has less to build on. A client who spends time feeding it context gets better work back.

What Did It Cost?

The costs fell on time and on substance. The consultant's day started 70 minutes earlier, active hours rose by about 40 minutes, and replies slowed by seven minutes. Substance and originality gained far less than design and polish. A part of the brain's output may have replaced work the consultant would have done anyway.

• Longer days. First activity moved to 7:20 from 8:30, and active hours rose to 9.1 from 8.4.
• Slower replies. The median reply took 43 minutes against 36.
• Polish ahead of substance. Design rose 103% and substance 19%. Originality in strategy rose 15%.
• Flat email. The consultant sent 1% fewer emails, though email drafts made up about a fifth of the brain's pieces.
• Possible displacement. By-hand output rose 9% while the rest of the team rose about 75%.
• Weekend use. The consultant sent requests on a Saturday.
• Unmeasured review time. We didn't count the hours the consultant spent checking the brain's work. Those hours sit inside the longer day.

What Does Other Research Say?

Researchers have found both gains and harms. Consultants with GPT-4 scored more than 40% higher on tasks it handled well and were 19% less likely to get a problem right outside them. The most skilled support agents gained little, and Danish workers' earnings held flat. None of these studies tested the brain.

We read the studies that support results like ours and the ones that don't. For each we give the link, what the study measured and how it differs from one consultant using the brain for 24 working days.

Consultants inside and outside the frontier (Harvard and BCG)

Fabrizio Dell'Acqua, Ethan Mollick and colleagues ran a pre-registered experiment with 758 BCG consultants on realistic tasks, with and without GPT-4. On tasks the tool handled well, consultants with GPT-4 finished 12% more tasks and worked 25% faster, and their work scored more than 40% higher. Read the study summary.

On a problem the researchers picked to sit outside the tool's strengths, consultants with GPT-4 came out 19% less likely to reach the correct answer, according to the 2026 paper in Organization Science. BCG reports that many took the tool's answer without challenging it, and that ideas from the GPT-4 groups came out 41% less varied.

This study sits closest to ours, with consultants doing consulting work, but it tested a general chatbot on set tasks with no client files. We set no task outside the brain's strengths, and the 41% drop in idea variety fits our weakest strategy scores, options and originality.

Customer support agents (NBER, 2023)

Erik Brynjolfsson, Danielle Li and Lindsey Raymond tracked 5,179 support agents at one firm as it rolled out an AI assistant that suggested replies. Agents resolved 14% more issues per hour on average, and novice and low-skilled agents 34% more. The most skilled agents gained little, and their resolution rates and customer satisfaction dipped. Read the working paper.

We count this one against us, since our consultant is experienced. The agents got suggested lines in a chat while the brain builds whole documents from the consultant's files, and we can't weigh one person's gain against a result across thousands.

Experienced developers on their own code (METR, 2025 and 2026)

METR ran a randomized trial with 16 experienced open-source developers on projects they knew well. They took 19% longer when they could use AI tools, and afterward they still believed the tools had made them 20% faster. Read the 2025 study.

In a 2026 follow-up with 57 developers and more than 800 tasks, returning developers took about 18% less time with AI and new recruits about 4% less, and neither result reached statistical significance. METR calls the data an unreliable signal: some developers wouldn't sign up to work without AI, and 30% to 50% held back tasks. Read the update.

The gap between felt and measured speed is the reason we counted from the consultant's systems instead of asking them. METR timed each coding task, and we counted documents and email, so we can't say whether each piece went faster.

Earnings and hours across Denmark (NBER, 2025)

Anders Humlum and Emilie Vestergaard linked a survey of about 25,000 Danish workers in 11 occupations exposed to AI chatbots to national records of pay and hours. Two years after ChatGPT launched, they found no effect on either and could rule out effects above 2%. Users reported saving about 3% of their hours, and most moved that time to other work. Read the working paper.

Humlum and Vestergaard measured pay across a country, and we measured one person's output without looking at pay. A worker can ship more without earning more, so 2.2 times the output doesn't mean 2.2 times the value.

Workslop (BetterUp Labs and Stanford, 2025)

BetterUp Labs and the Stanford Social Media Lab surveyed 1,150 US desk workers about workslop, their term for AI-made work that looks good and lacks substance. Of the workers, 40% had received some in the past month, and they estimated each case took about two hours to deal with. Read the HBR article or the BetterUp summary.

The survey rests on self-reports and covers no single tool. Its term fits the risk in our scores, with design up 103% and substance up 19%. We didn't ask the consultant's readers how long they spent fixing the work.

Critical thinking (Carnegie Mellon and Microsoft Research, 2025)

Hank Lee and colleagues surveyed 319 knowledge workers who use generative AI each week and collected 936 real examples of their work. More confidence in the AI went with less critical thinking, and more confidence in their own skill went with more. Read the paper.

The study rests on self-reports, with no test of skill. It lines up with our strategy scores, where evidence and actionability rose far more than options and originality, though we didn't measure the consultant's own thinking.

Skills that fade (The Lancet and PNAS, 2025)

Researchers in two fields outside office work found skills slipping after people worked with AI. At four endoscopy centers in Poland, doctors found precancerous growths in 28.4% of the colonoscopies they did without AI in the three months before the centers adopted it, and in 22.4% in the three months after. Read the study. Among about 1,000 high school students, a plain GPT-4 tutor raised practice grades 48%. Once it went away, those students scored 17% below peers who had worked without it, and students whose tutor had safeguards avoided most of that drop. Read the paper.

Neither study involved consultants or documents, and the colonoscopy researchers looked back at records instead of running a trial. We can't test their worry in 24 working days: the consultant's by-hand decks rose 1% while total decks rose 680%, so they now practice that skill less.

We read the outside studies this way: AI lifts output on the right tasks and puts judgment and skill at risk. We found the first half for one consultant, and we'd need a longer count to test the second.

How Sure Are We?

We're sure of the counts and less sure of the cause. We recounted each figure in a separate pass, and the output rise held under each base, from 1.9 to 2.5 times. The quality scores rest on eight pairs and six against six, from a scorer who helped make the after versions. One consultant can't stand for a firm.

• One person, one firm, 24 working days. The result holds for this consultant and may not carry to others.
• No control group. The rest of the team rose about 75% in the same weeks, and we can't separate the season from the brain with certainty.
• Small quality samples: eight document pairs, and six strategy pieces against six.
• A scorer with a stake. The scorer helped make the after versions.
• Volume over value. We count a deck the same whether a client acts on it or leaves it unread.
• A first month. People often use a new tool more at first, and this report covers the first 24 working days alone.
• A vendor's report. We sell the brain.

What Do We Tell New Clients?

We show them this report, costs and all. The brain made one consultant's month look like two and made their documents look better. It didn't shorten their day, and it added less to substance than to polish. Zach and I think clients should expect more output and should keep the thinking theirs.

Zach and I think the honest pitch is narrower than 2.2 times. The brain makes more of the work you already do and makes it look finished. You keep the judgment and the job of checking that the substance holds.

• Feed it context. The best strategy pieces came where the brain had rich input.
• Check substance before polish. You can put a weak argument in a clean deck.
• Watch your hours. Count your first and last activity before and after, the way we did here.
• Measure before you install. We needed a 12-month baseline from the client's own systems to write this report.

If you want us to run this count on your own work, start with a bottleneck analysis: we find the step where work waits, build the fix and time it before and after.