Why AI Adoption Goes Wrong in Construction

Download the full Catha guide

Keep the complete branded version for internal review and implementation.

AI adoption in construction is not a technology problem

This paper is written specifically for contractors, general contractors, specialty trades and construction-focused operators. The principles are tested against construction realities: project-based work, fragmented software, field/office handoffs, estimating pressure, document volume, subcontractor coordination, customer communication and the need to protect contractual and safety-critical decisions.

It is an operating-model problem with technology inside it.

AI adoption in construction has moved from curiosity to competitive pressure. The problem is that access to powerful models is no longer the hard part. The hard part is turning those models into reliable construction-business capability that people will use, leaders can govern, and the company can measure.

The market is full of impressive demonstrations. Far fewer organizations have redesigned workflows, assigned clear authority, measured construction business outcomes, and built the controls required to scale. McKinsey’s 2025 global survey found that 78% of respondents said their organizations used AI in at least one business function, yet more than 80% said generative AI had not produced a tangible enterprise-level EBIT impact. The gap between usage and value is the adoption challenge.

This paper identifies 12 of the most common misconceptions we see in AI adoption and replaces them with a practical operating approach. The objective is not to slow AI down. It is to make AI useful enough, controlled enough, and measurable enough that a company can safely give it more responsibility over time.

THE 12 MISCONCEPTIONS AT A GLANCE

THE ADOPTION GAP

Usage is high. Value capture is still uneven.

The organizations making progress are redesigning work, not merely adding tools.

The strongest evidence now points in the same direction: value comes from organizational change. McKinsey reported that workflow redesign, defined adoption roadmaps, KPI tracking, senior governance, and role-based training are positively correlated with stronger bottom-line impact from generative AI. Yet fewer than one-third of respondents said their organizations were following most of those adoption practices, and fewer than one in five were tracking well-defined KPIs for generative AI solutions.

Deloitte’s State of Generative AI in the Enterprise likewise found a scaling problem: more than two-thirds of surveyed organizations expected 30% or fewer of their experiments to be fully scaled in the following three to six months. At the same time, nearly three-quarters reported that their most advanced initiative was meeting or exceeding ROI expectations. That combination matters: AI can create value, but success is concentrated where implementation has matured beyond experimentation.

Responsible AI is also becoming more operational. NIST’s AI Risk Management Framework and its Generative AI Profile emphasize governing, mapping, measuring, and managing AI risk across the lifecycle. That is a useful lens for business leaders because it treats AI risk as part of how the system is built and operated—not a legal review added at the end.

Treating AI as a software purchase

PRACTICAL CHECKPOINT

Can we name the construction business outcome, workflow owner, baseline, and target before discussing a model or vendor?

Starting with the model instead of the problem

PRACTICAL CHECKPOINT

Would this project still matter if the AI brand name were removed from the proposal?

Automating a bad process

PRACTICAL CHECKPOINT

Could a new employee follow the non-AI version of the process from a one-page workflow map?

Confusing a good demo with a production system

PRACTICAL CHECKPOINT

What is the system required to do when it cannot prove that an action succeeded?

Giving AI too much authority too soon

PRACTICAL CHECKPOINT

Which actions are reversible, which are consequential, and which require explicit approval today?

Assuming human review automatically makes AI safe

PRACTICAL CHECKPOINT

What exactly must a reviewer verify, and do they have the evidence and time to do it?

Skipping data and access discipline

PRACTICAL CHECKPOINT

Can we list every system, data source, credential, and external action this AI can access?

Measuring activity instead of value

PRACTICAL CHECKPOINT

Which business KPI would worsen if the AI disappeared tomorrow?

Treating governance as paperwork after the build

PRACTICAL CHECKPOINT

Where are authority, logs, exceptions, recovery, and change control defined in the actual workflow?

Ignoring the people who have to work with it

PRACTICAL CHECKPOINT

Have the people who perform the work helped design how AI fits into it?

Trying to scale before proving one lane

PRACTICAL CHECKPOINT

Can we prove one lane works before duplicating it across departments?

Believing AI is a one-time implementation

PRACTICAL CHECKPOINT

Who owns ongoing performance, incident review, model/vendor changes, and retirement decisions?

CATHA PRACTICAL ADOPTION MODEL

Build the business system around the AI

A seven-stage pattern for moving from idea to controlled value.

A PRACTICAL STARTING POINT

The first 90 days of responsible AI adoption

Enough structure to create evidence without turning the program into bureaucracy.

DAYS 1-30: SELECT & BASELINE

Choose one workflow with meaningful volume, visible pain, a clear owner, and data you can legally and operationally use.
Capture the current baseline: cycle time, labor time, cost, error/rework rate, conversion, backlog, or other business measure.
Map the process and remove obvious waste before adding AI.
Define the AI role and the actions it is not allowed to take.

DAYS 31-60: BUILD & PROVE

Build the narrowest solution capable of producing the outcome.
Add logging, permissions, thresholds, approval points, exception handling, and recovery behavior.
Test representative work plus edge cases and failure scenarios.
Run in shadow or approval mode until results are repeatable.

DAYS 61-90: MEASURE & DECIDE

Compare results to the baseline and calculate verified value.
Review incidents, false positives, missed cases, user friction, and manual rework.
Decide whether to improve, scale, narrow, pause, or retire the workflow.
If scaling, standardize the proven controls and training before adding more authority or departments.

LEADERSHIP SCORECARD

Ten questions before approving an AI initiative

A fast screen for whether an AI proposal is ready to become an operating system—not just a demo.

If several answers are vague, the project is not “not ready for AI.” It is not ready for implementation. Resolve the operating questions before expanding scope, authority, or spend.

Construction Applications

Apply these lessons first to information-heavy construction workflows such as estimating support, bid and lead research, RFIs and submittals, project administration, closeout documentation, customer and subcontractor communication, business development, M&A research and management reporting. Keep humans in control of contractual commitments, safety decisions, final pricing, change orders, payment authorization and other consequential actions.

CONCLUSION

Responsible AI is an accelerator—not a brake

The goal is not maximum autonomy. The goal is maximum useful authority the business can safely support.

The companies that win with AI will not necessarily be the ones that buy the most tools or deploy the most agents. They will be the ones that become good at converting business problems into controlled AI-enabled workflows, measuring the results, learning from failures, and expanding what works.

That requires a change in mindset. AI adoption is not a software installation. It is the design of a new operating capability. The model matters, but it is only one component. The workflow, authority structure, data access, evidence, human roles, recovery behavior, training, and metrics determine whether the capability creates durable value.

Good governance makes that easier. When the boundaries are clear and the evidence is available, leaders can safely delegate more. When failures are observable and recoverable, teams can move faster. When ROI is tied to construction business outcomes, the company can distinguish transformation from expensive experimentation.

CATHA CONSTRUCTION SERVICES

Practical AI adoption, governance architecture, workflow automation, and managed AI operations.

cathaconstructionservices.com/

RESEARCH NOTES

Selected sources

This paper is written for practical business use, not as legal, cybersecurity, or regulatory advice.

1. McKinsey & Company, “The state of AI: How organizations are rewiring to capture value,” March 12, 2025. Source

2. Deloitte, “The Path to Sustainable Generative AI Value Balances Passion, Pragmatism and Patience,” State of Generative AI in the Enterprise, January 21, 2025. Source

3. National Institute of Standards and Technology (NIST), AI Risk Management Framework (AI RMF). Source

4. NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024. Source

5. NIST AI Resource Center (AIRC), implementation resources for the AI RMF. Source

Statistics and survey findings in this paper are used to illustrate broad adoption patterns. Organizational outcomes vary significantly by use case, implementation quality, industry, data environment, operating model, and risk profile.

WHY AI ADOPTION GOES WRONG Common Mistakes, Dangerous Misconceptions, and a Practical Path to ROI for Contractors A practical white paper for business leaders implementing AI responsibly CATHA CONSTRUCTION SERVICES Practical AI. Responsible Control. Measurable Business Value.

THE CENTRAL IDEA AI does not create ROI because it is intelligent. AI creates ROI when it is placed inside a well-designed business process, given the right information and authority, measured against a real outcome, and surrounded by controls that make failure visible and recoverable.

1. Treating AI as a software purchase	7. Skipping data and access discipline
2. Starting with the model instead of the problem	8. Measuring activity instead of value
3. Automating a bad process	9. Treating governance as paperwork after the build
4. Confusing a good demo with a production system	10. Ignoring the people who have to work with it
5. Giving AI too much authority too soon	11. Trying to scale before proving one lane
6. Assuming human review automatically makes AI safe	12. Believing AI is a one-time implementation

WHAT THIS MEANS FOR BUSINESS LEADERS The competitive advantage is shifting away from simply having access to AI. The advantage is learning how to turn AI into dependable operating capacity faster than competitors—without losing control of quality, authority, data, customer trust, or accountability.

MISCONCEPTION	AI is often bought like a new app: pick a vendor, turn it on, and expect results.
REALITY	AI changes how work is performed. The value comes from redesigning the workflow around a specific construction business outcome, not from owning the model or subscription.
BETTER MOVE	Start with one measurable business problem, map the current workflow, and decide exactly where AI should assist, recommend, draft, classify, or act.

MISCONCEPTION	Teams ask, “What can we do with AI?” and then hunt for a use case.
REALITY	Technology-first projects produce impressive demos but weak adoption. ROI begins with a costly delay, repetitive workload, revenue leak, decision bottleneck, or service problem.
BETTER MOVE	Define the outcome first: hours saved, cycle time reduced, opportunities captured, errors prevented, or revenue enabled. Then select the lightest AI capability that can produce it.

MISCONCEPTION	If a process is slow, AI will make it fast.
REALITY	Automation amplifies whatever is already there. A broken approval path, duplicate data entry, unclear ownership, or poor handoff can become a faster and harder-to-see failure.
BETTER MOVE	Simplify before automating. Remove unnecessary steps, clarify ownership, standardize inputs, and establish the expected output before adding AI.

MISCONCEPTION	If the prototype works five times in a row, the system is ready.
REALITY	A demo proves possibility. Production requires repeatability, permissions, exception handling, monitoring, evidence, recovery, and a defined response when the model is wrong or unavailable.
BETTER MOVE	Test normal cases, edge cases, failure cases, and recovery. Decide what happens when confidence is low, data is missing, an API fails, or an action cannot be verified.

MISCONCEPTION	The fastest ROI comes from making the AI fully autonomous immediately.
REALITY	More authority increases both leverage and consequence. The right level of autonomy depends on reversibility, financial exposure, legal impact, customer impact, and the reliability of the surrounding controls.
BETTER MOVE	Earn autonomy in stages: observe → draft → recommend → act with approval → act within narrow standing authority. Expand only when evidence supports it.

MISCONCEPTION	A person in the loop means the risk is controlled.
REALITY	Human review can become rubber-stamping when volume is high, interfaces hide context, or reviewers are not told what to verify. A nominal approval step is not a control unless the reviewer can detect and stop the failure.
BETTER MOVE	Define what the reviewer is accountable for, provide the evidence needed to decide, and reserve human approval for decisions where it meaningfully changes the risk.

MISCONCEPTION	The AI only needs access to whatever helps it do the job.
REALITY	Every mailbox, file share, CRM, browser session, credential, and action endpoint expands the blast radius. Useful access is not the same as justified access.
BETTER MOVE	Use least privilege. Separate read, draft, send, approve, and administrative authority. Maintain an inventory of data sources, credentials, tools, and permitted actions.

MISCONCEPTION	More prompts, more agents, more generated content, and more automated tasks mean the AI program is working.
REALITY	AI can create enormous activity without improving the business. Token volume, task counts, and output volume are operating metrics, not proof of value.
BETTER MOVE	Track business KPIs: time to completion, conversion rate, error rate, rework, margin, response time, throughput, customer outcomes, or verified dollars created/saved.

MISCONCEPTION	Governance slows innovation, so it should be added after the system is working.
REALITY	Governance is the operating structure that makes delegation possible. Clear authority, logging, approval boundaries, evidence, recovery, and change control make systems easier to scale because everyone knows what the AI may and may not do.
BETTER MOVE	Design controls alongside the workflow. Governance should be proportional to consequence and embedded in how the system operates, not stored in a binder nobody uses.

MISCONCEPTION	If the AI is good enough, employees will naturally adopt it.
REALITY	People adopt tools that make their work easier, fit the actual workflow, and preserve accountability. Adoption fails when the system creates extra steps, threatens trust, or is imposed without role-specific training.
BETTER MOVE	Build with frontline users. Train by role and scenario. Explain what changes, what does not, what the AI is trusted to do, and how people escalate problems.

MISCONCEPTION	The bigger the launch, the faster the transformation.
REALITY	Scaling multiplies both value and defects. A narrow lane exposes integration, ownership, training, reliability, and control problems while they are still inexpensive to fix.
BETTER MOVE	Pilot one repeatable workflow with a baseline, owner, success threshold, rollback path, and short review cycle. Standardize what works before cloning it.

MISCONCEPTION	Once the system is deployed, the project is finished.
REALITY	Models change, vendors change, data changes, employee behavior changes, and business processes change. Without monitoring and maintenance, performance can drift while nobody notices.
BETTER MOVE	Operate AI as a managed capability: monitor outcomes, review failures, control changes, re-test high-consequence paths, and retire workflows that no longer create value.

1. OUTCOME	Define the business result and baseline. Choose a measurable target: time, cost, throughput, quality, revenue, conversion, response, rework, or risk.

2. WORKFLOW	Map the work before automating it. Identify inputs, decisions, handoffs, owners, exceptions, and where delays or waste occur.

3. AUTHORITY	Decide what the AI is allowed to do. Separate read, draft, recommend, approve, send, change, purchase, and administrative authority.

4. CONTROLS	Design failure handling before launch. Logging, approvals, thresholds, evidence, retry rules, rollback, escalation, and kill-switch behavior.

5. PILOT	Prove one narrow lane. Use representative work, including edge cases. Compare the result against the baseline and success threshold.

6. MEASURE	Verify business value and system reliability. Track both operating performance and construction business outcomes. A fast system that creates rework is not a win.

7. SCALE	Expand proven patterns deliberately. Reuse proven controls and training. Increase autonomy only when reliability justifies it.

A SIMPLE GO / NO-GO RULE Do not scale because the demo is exciting. Scale when the workflow is measurably better, the failure modes are understood, the people using it trust the process, and the organization can prove who or what is authorized to act.

#	QUESTION	READY WHEN…
1	What construction business outcome will change?	A measurable result and baseline exist.
2	Who owns the workflow?	A named business owner is accountable.
3	Why is AI needed?	AI is solving a real decision/content/automation problem, not adding novelty.
4	What authority will it have?	Read, draft, recommend, approve, send, change, and admin rights are explicit.
5	What data can it access?	Sources, permissions, retention, and sensitive data are understood.
6	How can it fail?	Normal, edge, integration, and recovery cases have been considered.
7	What happens when confidence is low?	Escalation and human review rules are explicit.
8	How will value be measured?	KPIs connect directly to business impact.
9	How will changes be controlled?	Prompts, models, tools, permissions, and workflows have owners and version control.
10	What is the rollback path?	The company can stop, isolate, or revert the system without losing control of the business process.

THE PRACTICAL STANDARD Start small enough to understand the system. Control it well enough to trust the system. Measure it clearly enough to justify the system. Then scale what is proven.

ABOUT CATHA Catha Construction Services combines real-world construction experience with data, relationships, and technology. Our AI work focuses on practical systems that solve operating problems, preserve accountable human control, and create measurable business value.

AI adoption in construction is not a technology problem

Usage is high. Value capture is still uneven.

Build the business system around the AI

The first 90 days of responsible AI adoption

Ten questions before approving an AI initiative

Construction Applications

Responsible AI is an accelerator—not a brake

Selected sources

Practical questions construction owners ask.

Keep reading

Use AI where it creates value. Govern it everywhere it acts.