The limits of AI in construction: what still needs human judgment

The construction technology industry sells certainty. AI will catch every error. Automation will keep your project on time and on budget. The dashboard will tell you everything you need to know. Anyone who has managed a construction project in Kenya knows this is not true.

Construction is a physical, relational, and unpredictable process. Drawings get interpreted differently on site. Subcontractors overcommit and underdeliver. Ground conditions turn out differently from what the soil investigation suggested. A contractor submits a variation claim that is partly legitimate, partly inflated, and entirely urgent.

AI and automation can help with all of these problems — but they cannot solve them alone. The most dangerous thing a developer can do is to trust a dashboard when the site needs an experienced professional to walk it, inspect the work, and make a judgment call. This article is an honest accounting of where AI stops and human judgment begins in construction, written for developers who want to use technology without being misled by it.

Variations are the single largest source of cost overrun on Kenyan construction projects. A change in scope, an unforeseen condition, or a client instruction generates a variation claim from the contractor. The claim typically includes a priced breakdown of additional work, a programme extension request, and sometimes a claim for associated preliminaries and overheads.

AI can assist with variation assessment by cross-referencing the claimed items against the original bill of quantities, flagging rates that exceed the contract schedule, and checking whether the claimed scope was in fact instructed. These are valuable automation tasks that save QS time and catch arithmetic errors.

What AI cannot do is assess whether the variation is fairly priced for the type of change. A variation for additional excavation in coral limestone on the Kenyan coast is not the same as additional excavation in red soil in Nairobi. The subcontractor market for that work, the availability of suitable plant, and the actual conditions on site all affect what a reasonable price should be. That assessment requires a QS who knows the local subcontractor market, has visited the site, and can form a professional opinion on whether the claimed rate reflects market conditions or opportunistic pricing.

AI also cannot assess whether a variation claim is strategically timed. Contractors sometimes submit inflated claims knowing the developer is under pressure to maintain programme and will negotiate rather than dispute. Recognising that dynamic requires experience with contractor behaviour patterns that no algorithm captures.

Construction sites in Kenya present conditions that no amount of pre-construction data can fully predict. Coral formations that were not visible in the soil investigation. Groundwater at a different level than the borehole logs suggested. An existing service line that does not appear on any utility map. A retaining detail that works on paper but looks fragile once the excavation is open.

AI tools can process site data — progress photos, drone surveys, sensor readings — and flag deviations from the plan. They can track whether the number of columns poured this week matches the programme. They cannot walk the site at 7am, observe the quality of the concrete placement, notice that the formwork is showing signs of deflection, or decide that the ground conditions require a different foundation solution than the one specified.

The structural engineer and the clerk of works are the professionals who make these judgments. Their value is not in processing data faster but in seeing things that data does not capture: the quality of workmanship, the condition of materials, the competence of the site team, the signs that a subcontractor is cutting corners. These are human observations that require professional training and site experience. No sensor array replaces them.

On coastal sites, humidity, salt exposure, and aggressive soils add layers of complexity that particularly demand experienced inspection. A specification that works in Nairobi may fail in Mombasa. The engineer on site is the person who knows the difference.

Construction contracts in Kenya — whether based on the Joint Building Council (JBC) forms, FIDIC conditions, or bespoke agreements — contain clauses that interact in complex ways. A delay event may trigger an extension of time clause, which may interact with a liquidated damages clause, which may be affected by a concurrent delay clause. The interpretation of these interactions determines who pays for the delay.

AI can assist by retrieving relevant clauses from the contract, comparing the language of a claim against the contractual notice requirements, and flagging missing documentation. This is useful for contract administration efficiency but does not constitute legal interpretation.

The determination of contractual entitlement — whether a contractor is entitled to an extension of time, whether a variation is within the scope of the original contract, whether a termination notice is valid — requires legal training and professional judgment. In Kenya, this is typically the domain of the contract administrator or quantity surveyor operating within their professional remit, supported by legal counsel where disputes escalate.

AI also cannot assess the commercial context of a contractual dispute. A strict reading of the contract may support the developer's position, but enforcing that position may damage the contractor relationship to the point where programme recovery becomes impossible. Knowing when to stand on the contract and when to negotiate a commercial settlement is a judgment call that comes from experience, not from a clause-matching algorithm.

A construction project in Kenya involves multiple stakeholders with competing interests: the developer, the financier, the design team, the main contractor, subcontractors, the county planning department, NEMA where triggered, the local community, and sometimes the political leadership of the area. Each stakeholder has their own priorities, timelines, and tolerance for delay.

Managing these relationships is fundamentally human work. It requires reading a room, understanding what each party actually needs versus what they are asking for, knowing when to escalate an issue and when to let it resolve informally, and building the trust that keeps a project moving when things go wrong.

AI tools can track correspondence, log decisions, and maintain an audit trail of who agreed to what and when. These are valuable support functions, particularly for projects where multiple parties are communicating across email, WhatsApp, and formal correspondence. But the tool is a record-keeper, not a negotiator. It does not know that the county planner is under pressure to deliver revenue this quarter and might approve the application faster if the fee is paid now rather than next month. It does not know that the community elder whose support is critical responds better to a site visit than a formal letter.

The project manager, architect, or developer who manages these relationships effectively is performing work that AI cannot replicate. The technology supports their record-keeping; it does not replace their diplomacy.

AI cost estimation tools produce preliminary estimates from documented benchmarks — construction rates by building type, specification level, and location. This is genuinely useful at the feasibility stage, where a developer needs an order-of-magnitude cost to decide whether to proceed.

For complex or unusual elements, benchmarks become unreliable. A specialist facade system that has only been installed on two buildings in Kenya. A marine piling solution for a waterfront structure. A heritage restoration component with unknown conditions behind the existing fabric. For these items, there is no benchmark to reference. The cost must be built up from first principles: material procurement, specialist subcontractor availability, plant requirements, and programme constraints.

This type of estimating requires a QS or estimator with deep knowledge of the relevant specialist market. They need to know which subcontractors can do the work, whether those subcontractors are currently available, what the logistics of importing specialist materials look like, and what contingencies are appropriate for the level of uncertainty. This is not data processing. It is professional judgment informed by market relationships and construction experience.

In the Kenyan coastal market, specialist trades — curtain walling, building management systems, marine works — often require mainland or international subcontractors. The cost and programme implications of bringing those subcontractors to site introduce variables that no benchmark database captures.

The conversation about AI in construction should not be framed as 'does it work or not.' AI works for specific, well-defined tasks: preliminary cost estimation from documented benchmarks, bill of quantities error flagging, programme tracking against baseline, correspondence logging and retrieval, and progress photo comparison against the BIM model.

The more important question is: where does AI stop adding value, and what are the consequences of trusting it beyond that point? The answer, for construction, is that AI stops at the boundary of professional judgment. It cannot assess the fairness of a variation. It cannot inspect the quality of workmanship. It cannot interpret contractual entitlement. It cannot negotiate with stakeholders. It cannot estimate what has never been built before.

For developers, the practical implication is clear. Use AI to handle the data-intensive, repeatable tasks that consume professional time without adding professional value. Free your QS, engineer, project manager, and architect to focus on the site inspections, the variation assessments, the contract interpretations, and the stakeholder conversations where their judgment makes the difference between a project that holds its budget and one that does not.

The REDM approach reflects this: automation handles the cost benchmarks, the programme tracking, the document management, and the audit trail. The professionals handle the judgment calls. Neither replaces the other.