Which Jobs at a 50-Person Texas Company Actually Get Faster With AI

A Fort Worth operations director asked me a good question in July. She had read that AI saves knowledge workers 5 hours a week, she had 52 people on payroll, and she wanted to know what 260 hours a week would look like on her P&L. Fair question. The math is also nonsense, because 31 of her 52 people run a warehouse. Forklifts don’t prompt.

So this post does the thing those headlines skip. It takes an actual Texas payroll shape, puts the measured research next to each role, and says which seats move. Some of the numbers are smaller than you’ve been told. One of them is negative.

What Does an AI Productivity Gain Actually Measure?

An AI productivity gain is the difference between how long a task takes with an assistant and how long the same task takes without one. Three different things get reported under that name. Self-reported hours saved, task completion time in a controlled trial, and blind-rated output quality. They rarely agree. The gap between them is the most useful part of the whole literature, because it tells you how much of a reported gain is real recovered time and how much is the pleasant feeling of watching a draft appear.

Self-reports run high. People remember the prompt that wrote the email in 40 seconds and forget the 20 minutes spent fixing the one that came back wrong. Controlled trials catch both. Stopwatches are rude that way. That’s why a trial can show a slowdown in the same week a survey shows a speedup, and neither side is lying about what they observed.

Keep the distinction in mind for the rest of this post. Every number below is labeled with what it measured. No exceptions.

What Do 50 Texas Jobs Actually Look Like?

Texas had 14,071,160 jobs in the May 2025 BLS Occupational Employment and Wage Statistics. Scale that mix down to 50 seats and you get a payroll shape no vendor deck ever shows you.

Add up the nine desk-based groups and you land on 21.9 seats. Round it. Twenty-two people out of 50 spend their day in documents, spreadsheets, tickets, email, code, drawings, or a CRM. That is the ceiling. Everyone else is somewhere a language model has nothing to grab, and no licence purchase changes that.

Your own shape will differ, and the direction it differs matters more than the exact count. A Houston accounting firm might run 40 desk seats out of 50. A Conroe fabrication shop might run 8. Count yours first. Before anyone quotes you a per-user rollout number, check it against this table, because a vendor pricing you at 50 users when 22 of them can use the thing is a rollout that fails on arithmetic before it ever gets a chance to fail on adoption.

Which Occupations Report Real Time Savings?

The Federal Reserve Bank of St. Louis has been tracking this since 2024 through the Real-Time Population Survey. Their November 2024 measurement put average time savings at 5.4% of work hours among people who use generative AI at work, which is a household survey rather than vendor telemetry, and that distinction matters because telemetry counts feature usage while this counts hours. Across all workers, users and non-users together, it was 1.4%. Aggregate productivity effect, 1.1%.

Two occupation figures came out of that same survey, and the spread between them is the point. Computer and mathematical occupations saved 2.5% of work hours. Personal service saved 0.4%. Six times the difference. By industry, information services led at 2.6% and leisure and hospitality trailed at 0.6%.

The distribution is more useful than the average. Among AI users, 33.0% reported saving an hour a week or less. Another 26.4% saved 2 hours, 20.1% saved 3 hours, and 20.5% saved 4 or more. So a third of the people already using this thing are getting back less than one lunch break a week. One lunch break.

Frequency explains most of the spread. Among people using AI every workday, 33.5% saved 4 or more hours a week. Among people using it one day a week, only 11.5% did. Dabbling pays nothing.

The trend is real but slow. The St. Louis Fed adoption tracker shows the share of work hours saved across all employed adults moving from 1.6% in Q3 2024 to 2.2% by Q2 2026. Nearly two years. Six tenths of a point.

Why the Biggest Gains Land on Your Newest People

Here is where the role-by-role picture stops being about job titles.

Brynjolfsson, Li, and Raymond studied 5,179 customer support agents through the staggered rollout of a generative AI assistant, published in the Quarterly Journal of Economics in 2025 after first circulating as an NBER working paper two years before that. Their headline result was a 14% increase in issues resolved per hour. Underneath the average, the newest and lowest-skilled agents improved 34%. The most experienced and highest-skilled agents barely moved. Barely at all.

Read that twice. The assistant wasn’t making support agents better. It was making new support agents perform more like the good ones, by surfacing patterns the veterans had already internalized across years of calls, and that is a very different business case from the one printed on the marketing page you were shown.

The same shape shows up in three other trials. It’s the most reliable finding in the whole body of work.

One rule explains every row. Just one. AI speeds up the work you would hand to a new employee, and it does not speed up the work you would hand to your best one.

That rule is also the practical test. Before you buy a seat for someone, ask whether the work in front of them is work they could delegate to a capable person on day 30. If yes, the licence probably pays. Buy the seat. But if the value of that role is 11 years of knowing which vendor always misses the delivery window and which client needs a call before the invoice goes out, no assistant is going to help, and you may be about to slow that person down.

Where AI Made People Slower

METR ran a randomized controlled trial in early 2025 with 16 experienced open-source developers working on repositories they had contributed to for years. Two hundred and forty-six real issues, randomly assigned to allow or forbid AI tools, mostly Cursor Pro with Claude 3.5 and 3.7 Sonnet. The developers forecast a 24% speedup going in. Afterward, they reported a 20% speedup. The stopwatch said 19% slower.

Sit with that for a second. Same people, same tasks, and a 39-point spread between what they felt happened and what the clock recorded. Thirty-nine points.

METR is careful about what this does and doesn’t prove. It’s a snapshot of early-2025 tools, on large mature codebases, with developers averaging years of prior experience on those specific repositories. They explicitly decline to claim it generalizes to other developers, other domains, or later models. Fair. But it’s the cleanest measurement anyone has of the exact condition most of your senior staff work under every day, which is deep familiarity with a system they already understand far better than any assistant does.

The BCG consulting experiment found the same cliff from the other side. Inside the model’s competence, big gains. Outside it, consultants using AI were 19 percentage points less likely to reach the correct answer than colleagues working alone. Confidently wrong, faster.

What Does This Add Up To on a 50-Person Payroll?

Take the 22 desk seats from the first table and price the recovered time. Texas annual mean wages come from the same BLS release, loaded at 1.43 times wages to cover benefits and payroll taxes, using the Q1 2026 Employer Costs for Employee Compensation ratio of $46.60 in total compensation to $32.60 in wages. Then apply the 5.4% user rate. That’s 112 hours a year.

Weight those by the seat counts from the first table and all 22 desk seats come to roughly $140,000 a year in recovered time. Don’t put that number in a board deck. Please don’t. It assumes 22 people use AI every working day at the rate reported by the heaviest users in the Fed survey, and there is no evidence any Texas company your size is doing that today.

The honest version is smaller and much easier to defend. Pick 5 seats. Two in office and administrative support, one in business and financial operations, one manager, one in sales. That’s $28,873 of recovered time against about $1,900 in annual licences. The ratio survives being wrong by half. It survives being wrong by three quarters. That’s the point.

Licence cost is its own subject, and the bundle arithmetic under 300 users is genuinely counterintuitive. We worked through it separately in the Microsoft Copilot cost guide for Texas businesses, including the break-even expressed in minutes per day.

How Many Texas Companies Your Size Are Even Using This?

Fewer than the noise suggests. The Census Bureau’s Business Trends and Outlook Survey has been asking every two weeks since 2023, and the May 2026 release put national AI use at 19.8% of firms.

• 37% of firms with 250 or more employees currently use AI
• 32% of firms with 100 to 249 employees
• Under 20% of firms with 4 or fewer employees, a figure that hasn’t moved significantly since December 2025
• Information sector, 39.7%. Retail trade, about 14%

Depth is thinner still. Census researchers found that 57% of firms using AI in any business function use it in only 1 to 3 of the 15 functions the survey tracks, a list that runs from marketing and customer service through to finance, human resources, IT, and research. Not 15. Not 8. One to three.

That reframes the competitive question entirely. If you run a 50-person company in DFW and you get 3 seats genuinely productive by the first quarter of next year, you aren’t behind anybody, and you got there by picking correctly rather than by buying widely. Picking beats buying.

How Do You Pick the First 3 Seats?

We run this as a short exercise with clients before anyone buys anything. One hour, with the right people in the room.

1. Count your actual desk seats. Not headcount. The people who spend 4 or more hours a day in documents, email, tickets, spreadsheets, or a CRM.
2. Inside that group, find the roles with the highest volume of repeatable written output. Quotes, service tickets, RFP responses, job descriptions, first-draft correspondence, meeting summaries.
3. Apply the new-hire test to each one. Could a capable person do this work on day 30 with good instructions? If yes, that seat is a candidate. If the work depends on judgment nobody has ever written down, skip it.
4. Bias toward your newer staff. That’s where the 34% showed up, and it’s the least intuitive part of this whole exercise.
5. Pick 3. Set a 60-day window and one measurable thing per seat, like quotes issued per week or median ticket response time.
6. Measure it against the prior 60 days. Self-reports don’t count, for the reason the METR developers demonstrated so clearly.

Three seats isn’t timidity. It’s the only size at which you can still tell whether the thing worked, because at 22 seats you’ll get an aggregate number that moved for six different reasons in the same quarter and you will never isolate which one was the licence.

If you’d rather not run this internally, our AI readiness assessment covers the same ground against your actual tenant data. The Copilot readiness checklist is the free version if you want to start on your own.

What Has to Be True Before the First Licence

Three things. None of them are technical.

You need a written rule about what data can go into which tool. Not a policy binder. One page, in plain language, naming the approved tool and the categories of data that never leave the tenant. We published a Texas AI acceptable use policy template for exactly this.

You need to know what’s already running. Almost every discovery pass we do turns up personal accounts nobody approved, which is the subject of what your Texas employees are already pasting into ChatGPT. Buying a sanctioned tool while the unsanctioned one keeps running gives you both problems and one bill.

And you need your permissions cleaned up before you switch on anything that reads your files, because a Copilot licence inherits whatever oversharing already exists in SharePoint, which is the whole subject of Copilot data oversharing. On the regulatory side, read the TRAIGA compliance guide and the broader AI governance and compliance page. For a 50-person company the binding constraints are usually contract terms and client data agreements rather than the statute itself.

An Honest Read on These Numbers

The 5.4% figure is self-reported. The St. Louis Fed says so plainly, and the METR result is the strongest available warning about what self-reports do to a number. Treat 5.4% as a planning ceiling. Prove your own figure in a 60-day window.

The 50-seat table is a composite. It’s the Texas job market scaled down, not a survey of 50-person companies, so it tells you what Texas work looks like rather than what your specific company looks like. Use it as a starting shape. Correct it with your own org chart.

The frontier moves. Every result above is tied to a specific model generation, and the METR authors are explicit that a later model in the same setting could produce a different sign on the same measurement. What has held steady across four studies and three years is the novice-versus-expert pattern, not any particular percentage.

We’re a managed IT provider and we sell AI enablement work, so here’s the threshold. If you have fewer than 10 desk seats, a clean tenant, and someone internal who already runs Microsoft 365 confidently, you can run the 3-seat pilot yourself from the checklist above and you don’t need us for it. Where we earn a fee is permissions remediation, tenant configuration, and the measurement discipline that keeps a pilot from quietly becoming 22 licences that nobody audits and nobody can show a number for when the renewal lands.

Common Questions About AI Time Savings by Role

How much time does AI actually save at work?

Which roles at a small business benefit most from AI?

Does AI help experienced employees less than new hires?

Can AI make someone slower?

Is Microsoft Copilot worth it for a 50-person company?

What share of Texas businesses are actually using AI?

Where should a 50-person company start?

Before you talk to anyone, it’s worth reading how we structure this work on our managed AI services page. If you’d rather compare providers first, we published the criteria and the scores in the best Microsoft Copilot deployment partners in Texas.