AI engineering is entering preconstruction: what the evidence says about the pain points it can actually solve

AI is no longer a side experiment at the edge of construction tech. It is showing up in the phase where contractors commit money before anything is built: preconstruction. That shift is visible in contractor surveys, consulting research, and market spend data. It is also incomplete. Most firms that "use AI" still apply it to office admin more than to estimating, and many pilots never leave a chat window.

This research synthesizes public surveys and industry analyses into a practical picture: why preconstruction is attracting AI engineering attention, which chronic pain points map to what current systems can do, and where the evidence still thins out. Where hard benchmarks are missing, we say so. For the hour-level picture of where precon teams spend time today, start with Where do preconstruction teams spend their time?. For the risk framing behind the estimate, see Estimating is risk management, not cost calculation.

What the surveys say right now

Contractor intent is rising faster than deep workflow adoption.

The AGC of America / Sage 2026 Construction Hiring and Business Outlook reports that 61% of responding firms use AI or plan to increase investment in it, up from 44% in the prior survey. Usage by function is uneven: 45% deploy AI for office and administrative work, 23% for estimating, and 20% for design or preconstruction. In other words, precon is on the adoption curve, but most AI hours still sit in easier office tasks, not in bid packages.

A global RICS 2025 survey of more than 2,200 professionals paints a similar caution: about 45% reported no AI implementation, and another 34% were in early pilots. Interest is real (investors surveyed for that report put AI at the top of planned tech increases), but operational use remains cautious.

McKinsey's work on AI in AEC, as summarized by Construction Dive covering the July report How AI is reshaping the future of the AEC industry, places bid/no-bid analysis, estimating, and proposal drafting in the near-term window (roughly the first 18 months of a deliberate program). Medium-term value, in that framing, comes from proprietary project data such as RFIs, drawings, specs, and close-out history. The same coverage notes McKinsey's claim that AI has potential to automate a large share of nonphysical construction work, with the practical caveat that advantage comes from redesigning end-to-end workflows, not from sprinkling chat on top of the same process.

Market analysts also see capital concentrating early in the lifecycle. Mordor Intelligence's AI-in-construction segmentation attributes about 37% of 2025 spend to the pre-construction segment, reflecting a bet that early optimization has outsized return.

Why preconstruction is the natural entry point

Preconstruction is where dense documents meet irreversible commitments. Teams turn drawings, specifications, addenda, and subcontractor proposals into a number the company will stand behind. That is exactly the class of work modern document models and orchestration systems target: high volume, cross-referenced, deadline-bound, and expensive when wrong.

Three structural reasons keep showing up across industry research:

1. Information density before physical work. Specs, drawings, and revisions arrive before labor and materials, so leverage from better reading and comparison compounds into every later phase.
2. Labor and capacity pressure. Persistent estimator shortages and thin pursuit capacity push firms toward tools that expand how many packages a team can evaluate without lowering the quality of review.
3. Digitization debt. McKinsey has long described engineering and construction as under-digitized relative to other large industries, and has argued that AI payoff depends on prior data and process investment. Precon tools that only chat over PDFs inherit that debt; tools that structure the bid set begin to pay it down.

None of that means AI replaces estimating judgment. It means the bottleneck is shifting from "can we open every PDF" to "can we turn the package into a defensible decision fast enough." That is the same tension described in why every construction estimate still starts too close to zero.

Pain points that map to AI engineering (and ones that do not)

Piper's own synthesis of precon time sinks puts a large share of estimator effort in table-stakes work: bid leveling, scope assembly, admin, and coverage, often on the order of 60-80% of hours on a pursuit, with manual leveling alone frequently cited in the 40-50 hour range on traditional workflows. Those ranges are triangulated, not a single national census. Treat them as planning scaffolds. See the time research for methodology limits.

What matters for AI is the shape of the work:

Chatbots help until the job becomes a decision process. Asking a general model to summarize a PDF is useful. Running a pursuit is different: decompose the package, reconcile conflicting sources, apply company standards, produce outputs someone can defend in review, and keep an audit trail. That is AI engineering: systems designed around estimator workflows, not a novelty Q&A layer.

McKinsey's AEC framing lines up with that distinction. Firms that treat AI as a superficial productivity tool and firms that redesign core tasks end up in different camps. AGC's educational track for estimating and preconstruction makes the same practical point: AI should support professional judgment, with validation, not replace it.

What "AI engineering" means in this phase

In preconstruction, useful systems tend to share a few engineering traits that generic chat lacks:

Document intelligence across modalities. Bid sets mix text PDFs, drawing sheets, schedules, and proposal tables. Pipelines that only embed paragraphs miss symbol-heavy sheets and structured bid forms.

Relationships, not only retrieval. Estimators care which addendum supersedes which detail, which spec section governs a scope line, and which proposal is silent on a required inclusion. Linking entities beats keyword search.

Provenance as a product requirement. In precon, fluency without a source is a liability. Findings that cite file, page, and clause are usable in review; findings that cannot be checked are not. Industry procurement guidance for construction AI increasingly stresses auditability for exactly this reason.

Company context. Unit rates, package boundaries, preferred exclusions handling, and go/no-go patterns are firm-specific. A model that starts blank every session cannot safely invent them. Systems that load standards and prior pursuit memory can assist without pretending to own the decision.

Human-in-the-loop controls. Suggestions, overrides, and recorded reasons matter more than autonomous "final answers." The carry number, the award recommendation, and the narrative still belong to the team.

For a concrete view of where automation stops and judgment starts inside estimating, see the AI section of why estimates start from zero and the leveling workflow in construction bid leveling.

Limits and open questions

Honest research has gaps:

• There is no single authoritative hour benchmark for "AI saved X hours per bid" that is independent of vendors. Measure your own leveling, scope, and admin buckets before trusting a savings claim.
• Adoption surveys disagree because they define AI differently. Prefer the AGC/Sage function split over any global "percent of construction using AI" headline.
• Accuracy claims for takeoff or risk extraction are often vendor-reported on selected packages. Demand tests on your closed jobs.
• Field AI and precon AI are different problems. Jobsite computer vision and estimating document systems share a marketing label more than a workflow.
• Process beats model choice. Standardizing scope structures and review checklists still precedes most of the AI upside. Tools amplify a defined process; they do not invent one.

Practical implications for GC and CM teams

If you are deciding where to put attention next:

1. Instrument the bottleneck. Track hours in leveling, scope packaging, and trades with thin coverage. That tells you whether AI should target comparison, document intake, or solicitation follow-up first.
2. Prefer source-linked pilots. Run any tool against a completed package where you already know the misses. Score citation quality as hard as speed.
3. Keep admin AI and estimating AI separate in your roadmap. Chat for emails is fine; it is not the same investment as precon decision infrastructure.
4. Redesign one end-to-end workflow (for example, addenda intake through estimate update, or proposal-to-leveling) rather than deploying five disconnected assistants.
5. Protect judgment time. The win is not fewer estimators. It is more of their hours spent on risk, strategy, and sub relationships. That is the point of attacking bid leveling time sinks and addenda churn.

Where Piper fits

Piper is built for the document-heavy middle of preconstruction: reading bid sets against your company's scope standards, surfacing exclusions and silent items with the clause or drawing behind them, and keeping estimators in control of status and adjustments. It is aimed at the assembly and comparison work that crowds out judgment, not at replacing the carry decision.

Sources

• AGC of America and Sage, 2026 Construction Hiring and Business Outlook (AI investment and function-level usage: admin, estimating, design/preconstruction)
• RICS, Artificial intelligence in construction report (2025 survey of 2,200+ professionals; adoption and pilot rates)
• McKinsey & Company, How AI is reshaping the future of the AEC industry (as reported by Construction Dive, July 2025 coverage: near-term workflows including bid/no-bid, estimating, proposal drafting; nonphysical work automation potential; workflow redesign)
• McKinsey & Company, Artificial intelligence: Construction technology's next frontier (under-digitization context; digitization as a prerequisite for AI ROI)
• Mordor Intelligence, Artificial Intelligence in Construction market analysis (pre-construction share of 2025 AI construction spend)
• Piper Resources cluster: precon time sinks, risk framing, starting from zero