AI in Construction: Forecasting Purchases and Stock by Build Phase

How can AI forecast construction material purchases?

By relating the bill of quantities, the programme, real progress, stock, past consumption and supplier lead times, and adding weather and price movements as external signals. The output is a forecast of what each site is likely to need, week by week — not an automatic purchase order.

It is not a system that predicts the site. It is a layer of operational intelligence built on data the company already has: purchase exports, invoices, stock by warehouse and by site, budgets, bills of quantities, the programme, valuations and percentage complete, foremen's requisitions, consumption history, prices and lead times, crew calendars and site reports.

Every site has its own calendar, but purchasing is shared

A builder can have, at the same time, a house in structure, a building starting its roof, a refurbishment at first fix electrics, apartments taking floor finishes, and another site already in painting. Each site needs different materials — but the suppliers, the warehouses, the crews, the vehicles and the cash are shared.

Without an integrated view, the predictable happens: finishes delivered far too early, no steel when the structure is ready to move, paint bought before the surfaces are prepared, tiles and flooring sitting in a warehouse for months, orders for material that already exists on another site, and a supplier whose lead time crept up without anyone noticing.

A useful forecast needs three calendars, not one

The site calendar gives the planned sequence: structure, roof, services, finishes, second fix. The real calendar gives actual progress, delays, variations and what is genuinely complete. The external calendar brings weather, seasons, public holidays, shutdowns, prices, availability and delivery lead times.

The value comes from comparing them. If the plan says painting starts in two weeks but the site reports show the walls are not ready, the system should question the purchase, not confirm it blindly. An out-of-date programme is the fastest way to produce wrong forecasts that look competent.

The critical item is rarely the expensive one

In the structural phase the model relates the bill of quantities to progress: if a slab is due in a given week, it checks the materials needed, quantities already on site, confirmed deliveries, lead times, storage capacity, site access and unfinished dependencies.

One small missing part can block the use of a lot of material that has already arrived. So the system does not only look at high-value items — it looks for the components that are critical to the sequence of works. The same logic applies to building services: service unfinished → wall does not close → finish cannot start → painting slips. Cheap components that block several crews deserve priority over expensive materials that can wait.

«It's summer, should we buy roof tiles?» is the wrong question

Dry spells favour roofing, waterproofing and facade work — but the calendar alone is not enough. The forecast has to combine the real phase of the site, the area, the tile type and reference, membranes, insulation, flashings, gutters, fixings, unloading capacity, delivery lead time and the rain and wind forecast.

The right question is: will the site be ready, can the supplier deliver, and is there a likely window to do the roof? If roofing cannot start yet, buying tiles because summer arrived creates months of storage and a breakage risk.

The same goes for paint, which depends on both the phase and the application conditions: surfaces complete, render and filler finished, drying time, areas, approved colours, coverage, interior or exterior, ambient conditions and existing stock. A useful alert does not say «painting was scheduled for this month» — it says the crew arrives in ten days, the surfaces have not been signed off, and rain is forecast.

Reinforcing stock does not always mean buying more

A storm has different consequences depending on the business. On an active site it may justify revisiting the schedule for external work, reinforcing temporary coverings, protecting sensitive materials, checking drainage and postponing deliveries that would be left exposed — that is, protecting the site and avoiding a delivery at the worst moment, not ordering more tiles.

In a company doing maintenance and repairs the effect is the opposite: after a storm, demand can rise for compatible tiles, membranes, flashings, fixings, gutters and waterproofing products. There, the history of requests following similar events may justify a controlled safety stock — depending on the region, the type of customer and storage capacity.

Portugal's IPMA publishes forecasts and warnings for rain, wind, thunder, heat and cold, which work as an external signal for planning — never as a guarantee of what will happen.

Before buying, check whether it already exists on another site

A company running several sites can have a surplus in one place and a shortage in another. Before proposing a purchase, the system checks central stock, stock allocated to each site, reserved material, confirmed surpluses, reference compatibility, transfer cost and lead time, and the future needs of the site the material would leave.

A site that has finished blockwork with mortar and blocks left over, and another starting that phase next week, is an internal transfer before it is a new order. The decision still belongs to the crews, because material recorded as available may already be committed or may not be fit to use.

Buy early only what genuinely has to be bought early

Two products needed in the same week can call for very different decisions: one arrives in 48 hours, the other takes six weeks. The model works out the recommended decision date from the expected date of use, the supplier's average lead time, the historical variation in that lead time, stock on hand, the safety margin, the risk of a design change and how important the material is to the critical path.

Prices come into it too, but carefully. Alongside internal history and quotes, Portugal's INE collects construction material prices for cost indices and contract price revision — indicators that help spot trends and decide whether to re-quote, bring forward a negotiation or fix a price. But the system should not recommend a purchase just because the price might rise: it first has to confirm there is a likely need and that the risk of change is acceptable.

Real progress is worth more than the original plan

Programmes change — weather, permits, client variations, crew availability, supplier failures, technical dependencies, additional works, corrections to an earlier phase. So the forecast has to be refreshed with recent information.

An agent can pull signals out of site reports — «structure complete», «awaiting approval», «work suspended, rain», «material partially delivered», «measurement not yet confirmed». These signals do not replace the formal programme; they help you see when reality started to drift away from it.

An alert has to say which site, and on what assumptions

Critical material at risk of arriving after the date needed, stock too low for the next phase, a purchase planned for an activity that has slipped, finishes ordered before final approval, a surplus on one site usable on another, a supplier whose lead time is above normal, rain incompatible with a delivery, exterior paint without the conditions to apply it, differences between budgeted, purchased and consumed.

Each one has to explain which material, on which site, when it is needed, what data supports the warning, what assumptions were used and what needs human confirmation. The AI consolidates, spots patterns, reads reports, compares plan against execution, forecasts, flags and drafts proposals; validating quantities, confirming measurements, choosing materials, approving suppliers, negotiating and authorising orders stays with the company — because a forecast can be technically correct and make no commercial sense.

An SME does not have to start with a platform

A first proof can use purchase exports, a stock file, the bill of quantities, the programme for two or three sites, the lead times of the main suppliers, weekly reports and weather data — to answer one narrow question: which materials might run short on these sites in the next four weeks?

If the answer is useful, you then add ERP integration, stock by site, automated reading of reports, price movements, alerts and transfer proposals. The technology grows after the value is demonstrated, not before — the same approach we took with stock forecasting in healthcare, which also began with two exported files.

In construction the question is not «where can we put artificial intelligence». It is which site can stop, which material will run out, and how long we have to act.