AI Progress Prediction: What Forecasts Mean for Contractors

Nobody can tell you exactly how fast AI will improve, but the evidence is clear enough to act on. Capability is rising quickly, forecasters have mostly underestimated it, and the gains arrive unevenly from one task to the next. For a contractor, the sensible response is not to wait for a forecast to settle. It is to put your project management, accounting and field data into one system now, so every improvement lands on clean, connected data. That is what our AI-native construction ERP is built for.
TL;DR
- Forecasts disagree on timing, but expert panels, superforecasters and the public all lean toward a “moderate” progress scenario by 2030.
- Past forecasters badly underestimated AI benchmark progress, so “slower than predicted” is not a safe planning assumption.
- AI gains are jagged: big on some tasks, small on others. Pick specific workflows instead of betting on general claims.
- McKinsey sizes the U.S. architecture, engineering and construction opportunity at about $228 billion a year by 2030.
- Your best hedge is a single connected system with a short rollout, so you capture gains as they arrive.
Table of Contents
- How researchers measure AI progress
- What current forecasts say
- Forecasters have missed in both directions
- What the studies show about time savings
- What this means for construction
- How to plan when the forecast is uncertain
- Where our ERP fits
- Our point of view
- FAQ
- Sources
How researchers measure AI progress
“AI progress” sounds like one number. It is not. Researchers measure it in several different ways, and the choice of yardstick changes the answer.
The OECD’s AI Capability Indicators, published in June 2025, score AI on five-level scales for each human-like capability. Level 1 covers problems that are already solved, such as search retrieval. Level 5 means the system matches every aspect of the human ability. The OECD ranks AI at the highest level where it performs robustly and reliably, and it uses expert judgment where no formal test exists yet.
Google DeepMind researchers take another route. Their “Levels of AGI” proposal rates systems on two axes: how well they perform and how broadly they perform. Artificial Analysis publishes a Progress Index that puts language models released years apart on one scale, by estimating both model capability and task difficulty together.
Why this matters to you: when a headline says AI “doubled” in a year, ask what was measured. A benchmark score, a task success rate and a business result are three different things. The one you care about is the last.
What current forecasts say
The most useful recent data point comes from a lab measuring its own tools. In a November 18, 2026 report, OpenAI said autonomous agents’ success rate on two-to-four-hour research tasks rose from 23% in March to 46% by October 2026. On four-to-eight-hour tasks it rose from 13% to 35%.

The same report shows how uneven the gains are. Success reached as high as 79% for monitoring training runs, but only 25% for writing inference and serving code. OpenAI’s own framing is that uplift is “jagged” and humans remain central.
For longer-range views, the LEAP forecasting panel asked experts, superforecasters and the public what general AI progress will look like by 2030. All three groups most often picked the “moderate” scenario:
| Group | Moderate scenario | Rapid scenario |
|---|---|---|
| Experts | 49% | 22.96% |
| Superforecasters | 48.88% | 13.89% |
| General public | 45% | 26.26% |
The Forecasting Research Institute’s November 10, 2026 review adds a trend line. It reports that the frontier of the Epoch Capabilities Index rose about 6 points a year before reasoning models and about 14 points a year after September 2024. That is more than double the pace.
Read together, the picture is steady and substantial improvement, with a real chance of faster. Nobody credible is forecasting a stall.
Forecasters have missed in both directions
Is AI overhyped? The track record gives a more careful answer than either camp. The Forecasting Research Institute reviewed earlier predictions and found that forecasters dramatically underestimated AI progress on benchmarks and often underestimated adoption. They overestimated some real-world outcomes, including autonomous-vehicle use by the end of 2027. The institute found too little evidence to judge forecasts of large economic and social effects.
The lesson for a business owner is practical:
- Technical capability has tended to arrive faster than forecast.
- Physical, regulated deployment has tended to arrive slower.
- Economy-wide effects are still unproven either way.
Construction sits where both meet. Software that reads drawings, schedules work and flags risk is a capability problem, and those are improving fast. Pouring concrete is a physical one. Plan for rapid gains in the paperwork and decision layer, which is where most of your overhead lives.
What the studies show about time savings
Forecasts are one thing. Measured results are another.
A Stanford and MIT experiment gave 453 college-educated professionals midlevel writing tasks. Those using ChatGPT showed a 40% reduction in average task time and an 18% increase in output quality. Differences between workers narrowed too.
A June 2026 Stanford study estimates that current language models could save workers substantial time on at least half of their tasks in 46.6% of U.S. occupations. Yet 79.4% of those high-savings tasks show minimal usage of one major model. The authors point to privacy constraints and proprietary systems as the main bottlenecks.
That second finding is the one for contractors to read twice. The time savings exist, but they sit locked behind data that lives in spreadsheets, inboxes and disconnected tools. AI cannot help with a job-cost report it cannot see.
What this means for construction
Construction has the most room to gain and the weakest recent track record. McKinsey reports that U.S. construction productivity improved 10% from 2000 to 2022, or 0.4% a year, against 90% in manufacturing. Its analysis of more than 150 workflows across 25 AEC domains concludes that AI and other automation could unlock roughly $228 billion in annual value in the U.S. by 2030. It estimates AI could automate 39% of nonphysical construction work and 50% of nonphysical architecture and engineering work.
Nonphysical work is your back office and your project office: estimating, takeoff, bid leveling, scheduling, change orders, compliance paperwork, job costing and reporting. Here is where progress forecasts translate into daily work for each type of firm:
- General and commercial contractors: risk flags and schedule updates that previously depended on a project manager noticing a pattern. Our guide to how AI flags construction risks early shows what this looks like in practice.
- MEP and specialty contractors: faster quantity takeoff and estimating from drawings, which trims design and preconstruction hours. See how AI estimating reduces manual work.
- Engineering firms: less time on administrative chores, with more hours on billable design. Our notes on engineering team productivity cover the routines.
- Owners and construction managers: real-time visibility across jobs instead of a monthly rollup.
The same “jagged” pattern applies. AI will be strong at some of these tasks before others. That is why a staged approach beats a big-bang bet.
How to plan when the forecast is uncertain
You do not need to know whether the rapid or the moderate scenario wins. A good plan works in both. Here is a routine that does.
1. Get your data into one place
Every study above points to the same constraint: AI is limited by the data it can reach. If estimates live in Excel, costs in an accounting package and field notes in text messages, no model can connect them. Our platform combines project management, accounting and field operations in one system, including general ledger, AR/AP, certified payroll, job costing by cost code and WIP. That gives AI one connected record to work from.
2. Start with one workflow, not ten
OpenAI’s own numbers show success rates ranging from 25% to 79% depending on the task. Treat your own firm the same way. Pick one process with clear volume and a clear before-and-after measure, such as takeoff, resource allocation or schedule updates. The step-by-step logic in our guide to running one AI pilot before scaling applies here, as does the one on AI resource allocation.
3. Measure hours, not impressions
The Stanford experiment measured actual task time and output quality. Do the same. Track hours of manual data entry per week, days from bid invite to bid submission, and how often a schedule slip is caught before it hits the critical path. Our platform’s reported result is a 70% reduction in manual data entry, and you can compare your own baseline against that.
4. Keep the rollout short
If a forecast says capability could jump within a year or two, an 18-month implementation is a bad fit. Our implementation takes 2 to 4 weeks, which lets you start capturing gains while the technology is still improving instead of finishing the project after it has moved on.
5. Keep people in the loop
OpenAI’s report is explicit that humans remain central. Use AI for prediction and drafting, and keep your estimators, project managers and superintendents making the calls. A tool your team actually uses beats a smarter one they ignore. We report a 98% user adoption rate, driven by mobile and voice data entry in the field.
6. Revisit quarterly
Because trends can shift, review results every quarter. OpenAI itself cautions that historical trends and metrics that involve heavy human-AI collaboration may not predict future progress, after seeing sudden breakthroughs in mathematics. Set a calendar reminder, look at your numbers, and widen or narrow the pilot.
Where our ERP fits

We built DesignFlow Build as an AI-native ERP for contractors, engineering firms and construction managers. It brings together AI-driven project management (risk prediction and resource allocation), estimating and AI blueprint takeoff, bid management with vendor bid leveling, integrated scheduling, compliance automation for certificates of insurance and lien waivers, a customer portal, a mobile field app with voice data entry, and an automation builder for workflow rules.
That is one system where you might otherwise run spreadsheets, an accounting package and a separate project tool. Pricing is by seat: a free tier covers core features, Pro is $100 per month per office user, and Field seats are $25 per month per user. Essentials is also available. See the pricing page for what each plan includes.

If you want the whole picture first, start with our overview of the construction ERP, or go straight to project management or scheduling. You can also start free on DesignFlow Build and bring in your first project.
Our point of view
We do not know whether AI progress over the next three years will be moderate or rapid, and we do not think any contractor needs to know. What the research shows is simpler: capability keeps improving, the gains are uneven, and the biggest barrier to using them is data that is scattered across tools.
That is why we built one system instead of another add-on. When estimating, scheduling, job costing, payroll and field reports share a single record, every improvement in the models shows up as less typing, fewer surprises and faster decisions. Nothing has to be re-integrated when the technology moves.
We also built it to be adopted. A 2 to 4 week implementation, a mobile app with voice entry, and no army of consultants mean your crew gets value in weeks. Waiting for a clearer forecast costs you the weeks in which the savings would have accrued.
Start with one workflow, measure the hours, and let the results, not the headlines, decide how far you go.
Keith
FAQ
How fast is AI progressing right now?
Quickly, and unevenly. OpenAI reported that agent success on two-to-four-hour research tasks went from 23% to 46% between March and October 2026. The Forecasting Research Institute found that the pace of its capabilities index frontier more than doubled after September 2024. Task by task, results still range widely.
Is AI progress overhyped?
The evidence says both. Forecasters have underestimated benchmark progress and often adoption, but they overestimated some physical-world outcomes such as autonomous-vehicle use. For a contractor, the more reliable gains are in office and project workflows, not in replacing field labor.
What will AI do for construction by 2030?
McKinsey estimates AI and automation could unlock roughly $228 billion a year in the U.S. AEC industry by 2030. It sees potential to automate 39% of nonphysical construction work and 50% of nonphysical architecture and engineering work. Estimating, scheduling, cost tracking and compliance paperwork fall in that category.
Should I wait for AI to mature before adopting it?
No. Waiting means a longer period of manual entry and disconnected data, and you still have to do the data work later. Starting with a connected system and one pilot workflow lets you capture gains as they arrive. Our free tier and 2 to 4 week implementation make it a low-risk start.
Can AI predict project progress and risk on my jobs?
Yes, when it has your project data. Our AI-driven project management predicts risks and supports resource allocation using the information in your schedules, costs and field reports. The more of that work you run in one system, the more the predictions have to draw on.
Sources
- OECD: Constructing a framework to measure AI capabilities
- Google DeepMind: Levels of AGI for Operationalizing Progress
- Artificial Analysis: Progress Index v1 Methodology
- OpenAI: Research acceleration: The jagged frontier
- LEAP: General AI Progress Scenarios
- Forecasting Research Institute: How Close is Recursive Self-Improvement?
- Forecasting Research Institute: How Accurate Have AI Progress Forecasts Been So Far?
- Stanford SCALE: Experimental evidence on the productivity effects of generative AI
- Economic Evaluations of Language Models (arXiv, 2026)
- McKinsey: How agentic AI is transforming the AEC industry
