Build the AI business around accepted outcomes
AI usage and faster drafts do not establish customer value. A practical venture-building framework measures delivery cost, review effort, accepted outcomes, and paid continuation.
An AI product can generate more work without making the customer’s business more productive. A faster draft still needs to become something useful: a resolved request, an accepted deliverable, or a decision that someone can act on.
For founders, that distinction belongs in the commercial model. Our recommendation is to track the cost of an accepted outcome alongside usage and speed. This October 12, 2026 analysis draws on dated research, rather than treating earlier findings as breaking news. The operating framework below is TaylorVentureLab’s proposal, not a report of results from our ventures.
Productivity depends on the work being measured
A September 21, 2023 Harvard research summary describes an experiment involving 758 Boston Consulting Group consultants. For tasks within the tested AI system’s capabilities, assistance improved speed, task completion, and human-rated performance. The researchers emphasized that usefulness varied across tasks and arrangements of human and AI work. [1]
A different setting produced a different result. METR’s July 10, 2025 study randomly assigned AI access across 246 tasks completed by 16 experienced open-source developers working in familiar repositories. With the early-2025 tools studied, tasks took 19% longer. Participants nevertheless believed afterward that AI had helped them work faster. METR explicitly cautioned against generalizing the finding to all developers or future systems. [2]
These studies do not establish an average productivity effect for an AI startup. They concern different populations, tools, tasks, and success criteria. Their practical value is to make a founder’s question more precise: which customer, doing which job, benefits under which conditions?
Update the evidence when the tools change
The 19% figure should not become a permanent verdict on coding assistants. On February 24, 2026, METR reported that its follow-up experiment had serious interpretation problems. Some developers declined participation or withheld tasks because they did not want to work without AI. Concurrent agent use also complicated time measurement. METR believed developers were probably receiving greater benefits, while describing its data as weak evidence for the size of that improvement. [3]
In a separate survey published May 11, 2026, METR examined responses from 349 technical workers collected during February–April. Depending on the question, median self-reported gains in work value ranged from 1.4 to 2 times; reported speed gains were higher. The researchers highlighted selection bias and reasons to doubt the magnitude of respondents’ estimates. These were perceptions, not experimentally demonstrated business returns. [4]
Our conclusion is methodological: keep customer enthusiasm, observed workflow performance, and commercial results separate. Each can inform a decision. None automatically substitutes for the others.
Define the unit the customer accepts
Start with a job that has a recognizable end. For a research product, that might be a brief accepted against an agreed standard. For a support product, it might be a resolved case that does not immediately reopen. Avoid counting every generated response as a completed unit.
Before measuring, agree on three things with the customer:
- What makes the result acceptable?
- Who determines whether it meets that standard?
- How long will you watch for corrections or reopened work?
Hold those definitions steady during the comparison. A team that relaxes quality requirements halfway through can appear more productive without delivering the same service.
Our proposed operating metric is simple: total delivery cost divided by accepted outcomes. Include the costs attributable to failed attempts in the numerator. If nothing is accepted, report that directly instead of presenting an artificially attractive unit cost.
Include the human work after generation
For an initial commercial test, count model and tool charges, delivery labor, review, correction, and support. Show setup and integration costs separately, with an explicit assumption about how many future outcomes might absorb them. This keeps a promising marginal cost from hiding an expensive onboarding process.
Consider a hypothetical comparison. A manual workflow costs $600 to deliver 20 accepted briefs: $30 each. An AI-assisted workflow costs $400 in total, including review and failed attempts, but delivers only 10 accepted briefs: $40 each. The lower total spend has not produced the lower cost per accepted brief.
Those figures are illustrative, not observed results. They show why the denominator matters. Also record elapsed turnaround time and human effort separately: a process can finish sooner while requiring more supervision.
Released time creates financial value only through a credible mechanism. Ask whether it enables additional paid work, avoids an actual expense, improves service, or gives staff useful capacity. Describe that mechanism before assigning a dollar value.
Test whether the customer will buy again
A commercially useful experiment should reach the budget owner. Ask that person to define a paid continuation decision before the trial ends: the acceptable outcome, the service level, and the price at which continued use makes sense.
Then examine repeat behavior. Does the customer submit another real job? Does the workflow still work when the founder is not personally correcting every output? Can another team member deliver the service at a comparable cost?
Treat these as proposed diligence questions, not universal pass marks. A high-consequence workflow may justify substantial review; an early product may require costly learning. The decision is whether the economics can improve for identifiable reasons.
This week, choose one customer workflow and reconstruct its last set of completed jobs. Count accepted outcomes, total delivery costs, review effort, and corrections. Put that evidence beside the customer’s paid continuation decision. That gives the next product investment a concrete purpose: improve the result customers accept and the economics of delivering it.