Start free trial

What is AI-assisted engineering worth to your organization?

Estimate the engineering capacity an AI rollout creates, what AI tools and annual training cost to run, how long a cohort rollout takes, and when it pays back. Use it to set a target, test a business case, or check a training proposal.

Blue figures are inputs. Everything else is calculated. Negatives in tables are shown in (parentheses). Hover or focus a source number to preview it; select it to highlight the reference.
Scenario
Model
Value the gain as

Horizon
Team and AI spend
LocationEngineersLoaded cost / yr
$
$
$

Loaded cost is salary plus benefits, taxes, and overhead. BLS puts the median US developer wage at $136K, and benefits at 30% of compensation, so about $195K before overhead. [12]

Ongoing costs
Rollout + Investment
One-time costs

Net run-rate value

/yr

TL;DR

Why the values differ

The same productivity gain can create more output or reduce the payroll needed for today’s output.

Added output

Same team. More output.

Payroll freed Selected

Same output. Less payroll.

Both figures are before AI and training costs. Added output is capacity value. Payroll freed becomes cash savings when staffing or planned hiring decreases.

Run-rate bridge

One year with everyone at full proficiency, at today's prices. This is the steady state a rollout aims for.

Rollout

Show the schedule as a table

Cumulative net value

Every modeled benefit and cost by month, including engineers' time in the initial program, annual training spread evenly across months, and AI spend from the day each cohort starts. Payback is the month the curve crosses zero.

Underwriting view

Capacity is not cash. It reaches EBITDA only when some of it is taken as savings: a hiring freeze, attrition not backfilled, contractors rolled off. Capacity that is reinvested shows up as roadmap, not margin. AI spend hits EBITDA in every row.

Check a proposal

Enter the price and promised gain from an internal plan or a vendor pitch. The proposal gets the same team, AI spend, and annual training as this model; its price replaces the model's one-time cash costs.

    Where the gain comes from

    The whole job speeds up only as much as the parts AI actually touches. A 60% coding speedup lifts total throughput far less because coding is about a third of an engineer's week. Edit the split to match the team, then send the result to the model.

    Gain = 1 ÷ (1 − Σ share × (1 − 1 ÷ (1 + faster))) − 1. "+60% faster" means the same work takes 62.5% of the time.

    Evidence behind the defaults

    Published results vary widely. Tool rollouts with no process change land around 10–15% [1]. Teams that change how they plan, review, and test report 25–37% [1] [4]. The base case's 30% is a target for a structured program, not what licenses alone deliver.

    FigureWhat it measuresSource
    Productivity, org level
    +10–15%Typical gain from AI coding tool rollouts. 25–30% where teams redesign the end-to-end process around the tools.[1] Bain Technology ReportSep 2025
    +15–20%Average net gain across about 100,000 developers at 600+ companies, measured from repository data after rework. 30–40% on simple greenfield work, 0–10% on complex brownfield work.[2] Stanford, Denisov-Blanch2025 talk
    +26%Completed tasks in three field experiments with 4,867 developers at Microsoft, Accenture, and a Fortune 100 company. Code-completion era tools; junior developers gained more.[3] Cui, Demirer, Jaffe et al.Management Science, 2025
    +37%Median PRs per engineer per week across 500+ orgs (1.42 to 1.94). Heavy users save 6+ hours a week. Change confidence fell 6%.[4] DX, State of AI Impact Q2 2026Jul 2026
    No org gain10,000+ developers: 21% more tasks and 98% more PRs per person, but review time rose 91% and org-level DORA metrics did not improve. Review becomes the bottleneck.[5] Faros AIJul 2025
    −19%16 experienced open-source developers on their own repositories with early-2025 tools, who believed they were 20% faster. METR's 2026 update says developers are likely sped up with current tools.[6] METR RCTJul 2025; update Feb 2026
    +50%Self-reported by 132 Anthropic engineers, who use Claude in 59% of their work. Self-reports run high.[7] AnthropicDec 2025
    Productivity, single task (overstates org throughput)
    +21% / +56%Speed on one controlled task: Google enterprise RCT (about 96 engineers) and GitHub Copilot RCT (95 developers).[8] Google 2024 · GitHub 2023
    AI spend per engineer
    $150–250/moAverage Claude Code cost per developer across enterprise deployments, about $13 per active day. 90% of users stay under $30 per active day.[9] Claude Code docs2026
    $500–2,000/moUber's reported spend per engineer after Claude Code adoption rose from 32% to 84%. It used its 2026 AI budget by April.[10] CFO Dive2026
    $19–125/moSeat prices: Copilot Business $19 and Enterprise $39 (usage credits since June 2026), Cursor Teams $40, Claude Team premium seat $100–125.[11] GitHub · Cursor · Claude
    Engineer cost
    $136K → ~$195KMedian US software developer wage, and loaded cost before overhead given benefits at 30% of compensation.[12] BLS wages · BLS benefits2025–2026

    These are third-party results, not Binomial customer outcomes. Training prices and adoption rates are the weakest-sourced inputs: public workshop pricing runs $300–900 per seat, and a multi-week coached program costs more.

    Method and limitations

    • Gain is throughput per engineer, averaged across the whole team including slow adopters. It is not a task speedup.
    • Added output values the gain at loaded cost: what it would cost to hire the same capacity. Payroll freed values the payroll today's output would no longer need, which is gain ÷ (1 + gain).
    • Rollout is simulated week by week. Engineers lose a share of their time during the program, leave it with part of the gain, and reach the full gain after the ramp.
    • AI spend starts the week a cohort starts and grows each year. Teams use more as agents take on bigger tasks, and cheaper models get retired.
    • Training time is an opportunity cost, not cash. One-time costs are left out of run-rate EBITDA.
    • Not modeled: quality and rework effects, review bottlenecks, attrition caused by the change, and revenue from shipping faster.
    • Measure it. Baseline throughput, cycle time, and change failure rate before the pilot. Staggered cohorts give you a control group: compare trained cohorts against those not yet started.

    Replace assumptions with evidence

    Measure your own baseline before you fund the rollout.

    A scoped AI Engineering Readiness Assessment helps examine review burden, rework, and repository conditions with the team. Establishing AI usage and impact requires appropriate provider evidence and work context. Use the baseline to question this model’s assumptions, with confidence and limitations attached.

    Request an Assessment