BH TechnologyROI 2 / Field guide

01 The economics of AI

Make AI pay off where the business makes money.

Use AI to get more good work through a bottleneck or get a valuable system into use sooner. Then prove what changed in contribution margin, operating cost, or time to value.

See the case evidence Run the numbers
Where an AI productivity gain becomes a business return From a bottleneck, the upper path passes through review, increases accepted work, and reaches financial return. The lower path creates more output that fails review and generates rework. This is a conceptual diagram. THE RETURN PATH Bottleneck Review Accepted work Financial return Rework and errors Illustrative paths · no scale
When review catches errors early, higher output can become more work the business can actually use.
+15%issues resolved per hourLive customer support rollout −16.4%time on concept workP&G, AI-assisted individuals −19 ppaccuracy on one taskBCG, outside the tool’s capabilities

02 Case evidence

Where AI worked.
Where it went wrong.

ROI 1 starts with broad evidence. GenAI users in IDC’s Microsoft-sponsored 2025 survey estimated an average 2.8× return. These cases zoom in on individual workflows to show what improved, what broke, and what to measure in your own business. ROI 1 · IDC report ↗

01 Live rollout · Customer support

+15% issues resolved per hour

Help agents solve more cases.

Researchers tracked 5,172 support agents at a Fortune 500 software company as an AI assistant rolled out. Resolutions per hour rose on average, with larger gains among less experienced agents.

How to turn the gain into return
  1. Pick the queue. Start with high-volume requests with reliable answers and a clear handoff for exceptions.
  2. Count accepted resolutions. Track cases closed per paid hour alongside reopen rates, escalations, and customer satisfaction.
  3. Show where the hours went. Use capacity to reduce paid overtime, defer a planned hire, or absorb more demand. Record actual changes in cost or output separately.

Generative AI at Work ↗

02 Field experiment · Procter & Gamble

−16.4% time spent by individuals using AI

Shorten the path to a viable concept.

On real product innovation challenges at P&G, individuals using AI spent less time and produced work rated on par with two-person teams working without AI.

How to move a system into use sooner
  1. Compress one stage. Try speeding up the first concept, spec, or test plan. Keep the same review standard and approval owner.
  2. Track the whole cycle. Measure from approved scope to accepted production use, including review, integration, exceptions, and rollout.
  3. Value the earlier start. If a system goes live a month earlier and earns $4,000 in additional contribution that month, count that $4,000 and subtract any extra rollout cost. P&G measured concept work during an experiment.

The Cybernetic Teammate ↗

03 Field experiment · Boston Consulting Group

25.1% fasteron selected tasks suited to the tool
19 pp loweraccuracy on one task outside its capabilities

Same company. Opposite outcomes.

In an experiment with 758 BCG consultants, AI improved speed on a set of tasks the tool handled well. On a separate management problem, the AI groups finished roughly 18% to 30% sooner, depending on the group, and were less likely to reach the correct answer.

How to catch errors before they spread
  1. Decide at the task level. Test drafting, research, analysis, and final judgment separately. Each has different failure modes and consequences.
  2. Require the evidence. On recommendations, check the underlying records and calculation against a known answer or independent review.
  3. Track the cost of being wrong. Include revision time, downstream errors, and escalations. Keep human sign-off where a bad recommendation can carry a material cost.

Navigating the Jagged Technological Frontier ↗

04 Randomized study · Software development

+19% time to complete tasks

More help can mean slower delivery.

In one early-2025 test, 16 experienced open-source developers took longer with AI across 246 real tasks on repositories they already knew. A later METR study reported selection effects that made newer estimates hard to interpret.

How to measure delivery with AI
  1. Time the finished change. Include time spent prompting, reviewing diffs, correcting output, and getting the change accepted.
  2. Compare similar work. Separate familiar maintenance from new integrations and test scaffolding. A result from one category can hide a slowdown in another.
  3. Retest as tools and workflows change. Repeat the comparison with the models, codebases, and review standards your team actually uses.

Early-2025 study ↗ · 2026 update ↗

03 From a pilot to a return

Fix one bottleneck.
Compound the gains.

Start where another good unit of work can earn contribution, remove a real expense, or bring a valuable system online sooner. Pick one measure and give someone responsibility for proving it.

  1. 01

    Price the bottleneck.

    In sales, measure incremental contribution from closed deals. In delivery, count accepted units at their contribution margin. In support, track the cost per resolved case.

  2. 02

    Run a fair comparison.

    Establish the current time, volume, and quality. Test similar work with AI, include review and rework, and check that the next team can absorb the extra output.

  3. 03

    Reinvest what you free up.

    Send capacity to the next constraint. Reuse approved inputs, review rules, and integrations so the next rollout takes less work. Keep checking demand, adoption, and quality.

How gains can compound

A team uses AI to prepare quotes from approved prices. It clears the backlog, spends the freed hours following up on qualified deals, and reuses the pricing data to shorten onboarding. Measure quotes accepted, contribution won, and days to a billable start at the respective stages.

04 Work the numbers

What would one workflow have to earn?

Change the numbers to test a 12-month scenario. The starting values are illustrative. Capacity is tracked separately from dollars, and the estimate assumes a steady monthly run rate from month one.

Monthly workflow
Costs and financial gains (USD)

Illustrative scenario / 12 months

Net financial gain
$36,000
ROI
100%
Simple payback
3 months
Monthly capacity change
60 hours freed
What to include in the math
  • Benefits. Use cash expense actually removed and incremental contribution after variable costs. Count each benefit once.
  • Costs. Include implementation, training, licenses, oversight, and maintenance. Put each cost in the appropriate setup or monthly field.
  • Return. Divide 12-month net gain by total 12-month cost. Simple payback divides setup cost by positive monthly net gain.
  • Capacity. Apply minutes gained per unit only to the portion of volume using AI. Validate that work still meets the acceptance standard.

05 Start with one workflow

What would move your numbers?

Bring one bottleneck. In 30 minutes, we can map where AI might change the economics and what you’d measure to prove it.

Book a 30-minute consultation