AI saves hours at a 5 to 15 person agency. Whether it saves margin depends on how you bill

In a field study of 758 consultants at Boston Consulting Group, those using AI finished some tasks about 25% faster and got a task outside the AI's abilities wrong more often. For an agency of 5 to 15 people, those saved hours land in one of two places: your margin or your client's invoice. Which one depends on how you bill, and nobody checks it by default.

How much time does AI save on agency-type work?

On tasks AI handles well, the measured saving was about 25% of the time. On a task it handles badly, it lowered the share of correct answers by 19 percentage points. The range matters more than either number.

25%less time on tasks inside the frontier, where AI does well.

Source: Dell'Acqua et al., Navigating the Jagged Technological Frontier, Harvard Business School Working Paper 24-013, 385 BCG consultants on this task, 2023 (strategy consultants, not agencies). See all benchmarks

19 pointsfewer correct solutions on a single task outside the frontier, built for AI to get wrong.

Source: Dell'Acqua et al., Navigating the Jagged Technological Frontier, Harvard Business School Working Paper 24-013, 373 BCG consultants on this task, 2023 (strategy consultants, not agencies). See all benchmarks

Task in the studyWith AI vs without
Inside the frontier (tasks AI does well)12.2% more tasks completed, 25.1% faster, more than 40% higher quality
Outside the frontier (a single task built for AI to get wrong)19 percentage points less likely to reach a correct answer

Source: Dell'Acqua, Mollick and coauthors, Harvard Business School Working Paper 24-013, September 2023. A field experiment with 758 Boston Consulting Group consultants. Strategy consultants, not agency staff. 25.1% is the average of the study's two AI conditions. See all benchmarks

The researchers call the boundary between the two a "jagged frontier": it doesn't follow how hard a task feels. The first test asked consultants to develop a footwear idea for a niche market, from a list of ideas to market segmentation and market entry (pages 12 and 13 of the paper). The second test was a single task, built from BCG's case interviews and designed so that GPT-4 would get it wrong: the spreadsheet looked complete, and only the interview notes led to the right answer (page 17). The study used GPT-4 in 2023. Tools have changed since, and the boundary moves with them.

One more finding from the same study shapes the decision. Consultants below the average performance threshold improved by 43%; those above it improved by 17%.

Diagnose where your saved hours went

Answer three questions about your largest client before deciding anything. Each one has a yes or no answer.

  1. Do you bill this client by the hour? If yes, every hour AI saves is an hour you no longer invoice.
  2. Is most of this client's work drafting, summarizing or producing first versions? That is the work most likely to sit inside the frontier. Diagnosis from messy data, strategy calls and anything where a wrong answer costs the account sit closer to the edge.
  3. Has the client asked for a lower fee because of AI? If so, the saving is already being negotiated away.

Karl Sakas, an agency consultant, puts the client side bluntly:

Generative AI offers huge productivity gains. But it's often really ‘cover’ for cost-cutting.

Source: Sakas & Company, 19 predictions for agencies in 2030, May 4, 2026.

A prediction, not a measurement.

The math: same saved hours, two billing models

On a fixed retainer, saved hours raise the margin. Billed by the hour, the same saved hours lower revenue and leave the margin where it was. The example below uses planning assumptions; only the 25.1% comes from the study.

Assumptions

Client workPlanning assumption
60 hours a month
Delivery costPlanning assumption
$50 an hour
Share of those hours inside the frontierPlanning assumption; nobody has measured this for agencies
40%
Time AI saves inside the frontierDell'Acqua et al., 2023; strategy consultants, not agencies
25.1%

Result

Hours saved: 60 × 40% × 25.1%
about 6 hours
New total
54 hours
Planning example
Before AIAfter AI, fixed retainer of $6,000After AI, billed at $100 an hour
Revenue$6,000$6,000$5,400
Delivery cost$3,000$2,700$2,700
Delivery profit$3,000$3,300$2,700
Delivery margin50%55%50%

Billed by the hour, the agency keeps its margin percentage and loses 10% of its profit on the account. The client gets the saving.

On a retainer, the extra $300 a month is room to choose: keep it, or share part of it with the client before the client asks. Sharing $150 a month still leaves $3,150, more than before AI.

What to use to measure it

The only tool this needs is a time log with a client field, used for a month before and a month after a change in how the team uses AI. Without hours by client, the saving in the table above is a guess.

The client hours calculator turns a month of logged hours into margin by client. If the team doesn't log hours yet, start with how to ask the team to log hours: three lines to send, and the rule that the numbers measure the account, never the person. If the saved hours are the reason you think you can delay a hire, run the numbers in before you hire: the capacity math first.

When this math won't help

Skip the exercise, or treat its result as a ceiling, in these cases:

  • Less than about a fifth of a client's hours are drafting, summarizing or first versions. At 20% inside the frontier, the saving is 20% × 25.1%, about 5% of hours. On a 60-hour client that is 3 hours a month, inside the noise of most time logs.
  • The work is mostly outside the frontier. Diagnosis from incomplete data and strategy recommendations are where the study found more wrong answers with AI. Saved hours there can cost the account.
  • The client is billed by the hour and won't move to a fixed fee. The saving then belongs to the client by contract. That can still be the right call; it just isn't margin.
  • Nobody logs hours by client yet. There is no "before" to compare against. Log one month first.

The one-afternoon check

Pick your largest fixed-fee client and estimate, in one afternoon, the most AI could be worth on that account.

  1. List last month's tasks for that client with their hours.
  2. Mark each task "inside" (drafting, summarizing, first versions, reformatting) or "outside" (diagnosis, strategy, anything a wrong answer would cost).
  3. Multiply the "inside" hours by 25%. That is the ceiling of the monthly saving, not the expected value.
  4. Multiply the ceiling by your delivery cost per hour. Write the number down before deciding to keep it, share it, or reprice.

More on agency operations for 5 to 15 people, with every sourced number in the benchmarks behind this site.

Sources