Procore Skills and the Race to Own the Construction Operating Layer

Procore Skills and the Race to Own the Construction Operating Layer

The most consequential item in Procore’s 23 July Digital Coworker announcement was not the agent count. It was Skills, a facility arriving across all packages in August that allows a contractor to encode its own processes, standards and best practices into the platform so that agents apply the company’s method rather than a generic one on every project. That reframes what is being sold.

A licence buys capability, and capability can be swapped for a competitor’s capability at renewal. An accumulating store of encoded company procedure is something else entirely, and the contractor that has spent two years teaching a platform its submittal thresholds, its safety protocols and its change-order escalation rules owns something considerably harder to move than a document archive.

Read that way, the three Digital Coworker packages, the expanded library of 20 agents and the Control Tower governance console are the delivery mechanism for a more ambitious position. Procore is bidding to become the layer on which construction work is executed and company method is enforced, rather than one more application competing on features.

Thiago da Costa, the company’s senior vice president of AI and data who arrived through January’s Datagrid acquisition, described the destination as “AI output that consistently reflects how the company operates, along with a way to preserve institutional knowledge as experienced workers retire, change roles, or leave the organization and new team members come on board.” Whether that constitutes an operating layer or an ambition to build one is the question buyers should be examining, and the answer is not yet settled.

Briefing

• Skills, rolling out across all Digital Coworker packages from August, lets organisations encode their own procedures using plain-language prompts or uploaded documents such as standard operating procedures and project standards.
• Three packages are generally available: Starter Pack with five agents, Pro with the full library of 20, and Enterprise adding Agent Studio for building custom agents.
• Control Tower ships with every tier and reports AI usage and credit consumption by agent, project or team member, with updated subscription terms published alongside.
• New agents cover site safety hazard identification from photographs, video and drawings, schedule sequencing and dependency risk, and change analysis across RFIs, drawings, specifications and project records.
• Haskell, an early adopter alongside Consigli Construction and Level 10 Construction, has compressed submittal reviews from seven days to ten minutes.

What Skills Actually Is, and What It Is Not

The language of teaching invites a misreading that matters commercially. Skills are created through plain-language prompts or by uploading company documents, which describes an instruction and retrieval layer sitting above the underlying model rather than any modification of the model itself. Nothing in the announcement suggests a contractor’s standard operating procedures are absorbed into model weights, and the distinction cuts both ways. It weakens the alarming version of the argument, in which a contractor’s hard-won method becomes inseparably baked into a vendor’s product. It also strengthens the buyer’s hand, because configuration is in principle portable in a way that trained weights are not.

That is precisely why exportability becomes the pressure point in any serious negotiation. If the commercial pitch is knowledge preservation, the rational customer response is a contractual guarantee that the preserved knowledge remains retrievable in a usable form at termination, along with the agent audit logs that record how it was applied. A contractor building 200 Skills over three years has created a structured, machine-readable description of how it delivers projects, which is an asset with standalone value irrespective of which platform executes it.

Very few construction firms are currently negotiating for that, and the window in which to do so is the renewal conversation now underway rather than the exit conversation several years hence. Procore has been more transparent than most enterprise vendors by shipping usage visibility as standard, and the reasonable next request is comparable clarity on what leaves with the customer.

The Knowledge Problem It Is Built to Solve

The demographic case for encoding method is genuine rather than manufactured. The 2025 AGC of America and NCCER workforce survey found 92 per cent of construction firms struggling to fill open positions, with 45 per cent reporting at least one project delayed by workforce shortages, ahead of material lead times and permitting.

The AGC and Sage 2026 outlook recorded 82 per cent of respondents having difficulty filling hourly craft roles and 80 per cent struggling with salaried positions, the highest proportions in three years, while Associated Builders and Contractors estimates a net requirement of roughly 349,000 additional workers during 2026. The people who know how a particular contractor sequences a hospital fit-out or manages temporary works approvals are retiring faster than the industry can replace them, and the traditional transmission mechanism of apprenticeship and shadowing depends on a continuity of staffing that no longer exists.

Against that, McKinsey’s finding that construction productivity grew at just 0.4 per cent annually between 2000 and 2022, compared with 2 per cent across the wider economy and 3 per cent in manufacturing, explains why the sector has become receptive to a proposition it would have dismissed a decade ago. Appetite is measurable: the AGC and Sage outlook found 61 per cent of firms using artificial intelligence or planning to increase investment in it, up from 44 per cent the previous year, concentrated in office, administrative, estimating and preconstruction functions.

Encoding process knowledge into software is not a substitute for experienced staff, and no credible reading of the evidence suggests otherwise. It is a hedge against the departure of the people who currently hold that knowledge in their heads, and it standardises delivery across a portfolio in a way that inconsistent human interpretation does not.

Where the Execution Layer Earns Its Keep

The claim to be an execution layer requires evidence that work is actually being executed, and the strongest datapoint Procore has offered is Haskell. The design-build firm deployed Procore AI across its project data and reduced submittal reviews from seven days to ten minutes.

Hamzah Shanbari, its director of innovation, framed the gain as decision latency rather than headcount reduction, noting that “It’s about getting to decisions faster with high confidence. The more we can reduce that cognitive load of trying to find the right information, the more efficient our teams become.” Consigli Construction and Level 10 Construction are working alongside Haskell as early adopters informing development.

Compressed by days across a portfolio, submittal turnaround stops being an administrative statistic and becomes a programme argument, because approvals for long-lead packages including switchgear, curtain walling, structural steel and mechanical plant sit directly on the critical path. The three newest agents push into territory where the commercial stakes are higher still.

Site Safety reads hazards from photographs, video and drawings and returns field-ready findings, Schedule Analyst examines sequencing, dependencies and delay risk, and Change Analysis traces changes, RFIs, drawings, specifications and project records to identify scope impact, cost exposure and required follow-up. Change and delay determine contractor margin more reliably than productivity does, and an agent that surfaces a compensable event in week three rather than week twelve is operating on commercial exposure rather than clerical workload.

The capability arrived quickly by industry standards, with Datagrid acquired on 20 January for undisclosed terms, the first agents shipped in May, a rebuilt common data environment in June and commercial packaging in July.

Lock-In First, Metering Second

The sequencing of Procore’s commercial moves rewards attention. Control Tower reports credit consumption by agent, project and team member, and updated subscription terms were published alongside the launch, which together establish the infrastructure for usage-linked charging in a category that has overwhelmingly been sold on seats and modules.

Consumption pricing is ordinarily a fragile model, because customers facing a variable bill shop the market. It becomes durable when the customer has accumulated something on the platform that cannot readily be reproduced elsewhere, which is exactly what a library of encoded Skills provides. The encoded method creates the retention, and the retention is what makes metered pricing hold at renewal rather than triggering a tender.

The practical consequences reach into project cost planning regardless of how the strategic argument resolves. A document-intensive fit-out on a hospital or data centre will consume agent capacity at a materially different rate from a linear highways scheme with fewer document cycles and more repetition, which pushes software cost towards being a project-level preliminary rather than a fixed corporate overhead

. Contractors on framework agreements or NEC target-cost arrangements will need to decide whether that consumption is recovered against the project or absorbed centrally, and joint ventures will need allocation agreements between partners before the first invoice rather than after. Procore has not published package pricing or the cost of a credit, so the mechanics remain to be tested.

Financial visibility across the wider business is solid enough to support the experiment, with first-quarter revenue of $359.3 million, up 16 per cent, gross revenue retention of 95 per cent and remaining performance obligations of $1.562 billion, ahead of second-quarter results due on 29 July.

Programmability and the Contest for the Layer

There is a tension in the operating-layer claim that deserves naming. Agent Studio, the facility for building custom agents, is available exclusively in the Enterprise package, which means the platform is programmable only at the top tier. Platform strategies historically push extensibility downwards, because value compounds when third parties and mid-market customers build on top and the ecosystem generates capability the vendor never wrote.

Confining custom agent construction to the largest accounts reads more like enterprise account defence than platform building. The signal worth watching over the next year is whether Skills become shareable, portable or tradeable between organisations, since a marketplace of encoded method would demonstrate genuine operating-layer intent in a way that a premium tier entitlement does not.

The competitive picture reinforces the point that this is a contested bid rather than an accomplished position. Financial truth in construction still sits in the ERP tier, whether that is Trimble Construction One, CMiC or Viewpoint, and design truth sits with Autodesk. Procore’s route into relevance across both is Datagrid’s connectivity into ERP systems, cloud storage and document repositories, which is an attempt to reason across layers it does not own. Rivals are assembling comparable positions on a similar timetable.

Autodesk completed its acquisition of field data specialist Rhumbix on 31 March, and Trimble moved in April for Document Crunch, whose contract analysis technology Mark Schwartz, senior vice president of AECO software, described as providing a contractual rule set to serve as intelligent DNA across the Trimble Construction One suite. Procore’s decision to include a Contract Review agent in its entry-level five-agent Starter Pack, when a competitor has just bought a company to do the same job, indicates how differently the two firms read where defensible value sits.

Governance, Provenance and the Regulatory Test

Encoding company method into software raises assurance questions that UK and European infrastructure clients will ask before their contractors do. Control Tower records which agent acted, on which project and under whose account, producing an audit trail that becomes evidence rather than administration. That matters under the Building Safety Act regime, where the golden thread of information must be accurate, accessible and traceable through design, construction and handover, and where an automated action that cannot be attributed or explained is a liability rather than an efficiency.

Procore addressed the data residency element in June by launching a purpose-built common data environment developed for European market requirements, released first in the UK and Ireland with a localised UK Data Zone and a dedicated EU Data Zone planned for the autumn.

European AI regulation has moved in a direction that extends preparation time without removing the obligation. The EU’s Digital Omnibus simplification package, endorsed by the European Parliament in June and cleared by the Council on 29 June, deferred obligations for most standalone high-risk systems under Annex III from 2 August 2026 to 2 December 2027, with product-embedded systems moving to 2 August 2028.

Construction administration tools of this kind are not automatically high-risk, though agents touching safety findings, payment determination or employment decisions warrant legal review rather than assumption. There is a governance risk specific to Skills that regulation does not address at all, which is that an encoded procedure built from an outdated standard will be applied across every project with perfect consistency. Human error is at least inconsistent, whereas codified error scales.

What Contractors Should Settle Before August

Three questions need answers before Skills arrives, and none of them is technical:

1. The first is ownership of standards, because someone inside the business must be accountable for deciding which version of a procedure gets encoded and how it is reviewed thereafter.
2. The second is exportability, which means negotiating now for the right to retrieve Skills definitions and agent audit logs in a usable format at termination, alongside clarity on the treatment of consumption data.
3. The third is measurement, which requires baselining current cycle times for submittals, RFIs and change orders before deployment rather than reconstructing a comparison afterwards from memory and optimism.

Underneath all three sits data quality, since agent output is bounded by input and a firm with fragmented document control, inconsistent naming conventions and significant project information held outside the platform will capture a fraction of the reported gains.

The strategic judgement is more finely balanced than the marketing suggests. Encoding institutional knowledge into a vendor platform is simultaneously the most valuable thing on offer and the most binding, and contractors should be clear-eyed that they are choosing to concentrate method, memory and execution in one commercial relationship.

That may well be the right decision given what the workforce numbers show, and firms that make it deliberately, with exportability secured and standards governed, will be in a considerably stronger position than those that arrive at the same place by accumulation.

Key Industry Questions

1. Does Skills mean a contractor’s intellectual property ends up inside Procore’s AI model? On the available description, no. Skills are created from plain-language prompts or uploaded documents such as standard operating procedures, which indicates an instruction and retrieval layer above the model rather than modification of the model itself. Company procedure functions as configuration that shapes agent behaviour, not as training data absorbed into weights. That distinction matters because configuration is in principle portable, whereas trained weights would not be. It also means the meaningful protection is contractual rather than technical, and contractors should seek explicit terms covering the export of Skills definitions in a usable format, retention of associated audit logs, and confirmation that encoded procedures are not used to improve services offered to other customers.
2. What exactly should be negotiated on exportability? Four things are worth specifying before signing. The first is the right to export Skills definitions at any time and on termination, in a documented, machine-readable format rather than a screenshot or PDF summary. The second is retention and export of Control Tower audit logs, which carry evidential value for golden-thread and assurance purposes independent of the platform. The third is a defined transition period after termination during which export remains available, since data access frequently lapses at the moment it is most needed. The fourth is clarity on ownership and permitted use of consumption data and the encoded procedures themselves. None of these are unusual requests in enterprise software, and they are considerably easier to secure at renewal than at exit.
3. How does credit-based consumption change the way software is budgeted? Per-seat licensing produced a predictable annual figure that could be allocated to corporate overhead. Metered consumption ties cost to activity, so a document-intensive project generates a higher software bill than a repetitive one, moving spend closer to a project preliminary. Contractors need to decide whether consumption is recovered against the project or held centrally, and joint ventures require allocation agreements between partners. Procore has not published package pricing or credit costs, so modelling should be done against representative project types once figures are available rather than against a corporate average. Control Tower’s reporting by agent, project and team member supplies the granularity to set internal thresholds and alerts before the first surprise.
4. Why does Agent Studio being Enterprise-only matter strategically? Because it determines who can extend the platform. Custom agent construction sits exclusively in the top package, which means mid-market contractors can configure behaviour through Skills but cannot build genuinely bespoke automation without a tier change. Platform strategies usually push extensibility downwards, since ecosystems generate capability the vendor never has to write and create value that compounds outside the product roadmap. Restricting programmability to the largest accounts is a defensible commercial choice but a narrower one. The indicator worth watching is whether Skills become shareable or tradeable between organisations, because a marketplace of encoded method would signal genuine operating-layer ambition rather than tiered feature gating.
5. Is Procore actually becoming the operating layer for construction? It is bidding for that position rather than occupying it. Financial truth still resides in the ERP tier, including Trimble Construction One, CMiC and Viewpoint, while design truth sits with Autodesk. Procore’s claim is to the execution layer, where work is coordinated and company method applied, and the Datagrid connectivity into ERP systems and cloud storage is an attempt to reason across layers it does not own. Competitors are pursuing comparable positions, with Autodesk completing the Rhumbix acquisition in March and Trimble moving for Document Crunch in April. The contest will be decided by which platform reasons most reliably across systems it does not control, not by agent counts.
6. What is the risk of encoding the wrong procedure? It is the most underdiscussed governance issue in the release. A Skill built from a superseded standard operating procedure will apply that error consistently across every project, without the variability that causes human inconsistency to surface problems early. Codified error scales in a way that individual error does not. The mitigations are organisational rather than technical: assign named ownership of each encoded standard, establish a review cycle tied to procedure updates, version Skills so changes are traceable, and maintain human review on outputs affecting safety, payment or contractual position. Control Tower’s audit trail supports investigation after the event, but it does not prevent a flawed standard from propagating in the first place.
7. What does this mean for UK clients operating under the Building Safety Act? The golden thread requires information to be accurate, accessible and traceable through design, construction and handover, which places an evidential burden on any automated action affecting the record. Control Tower’s logging of which agent acted, on which project and under whose account, supplies part of that evidence. Procore’s common data environment, launched in June for the UK and Ireland with a localised UK Data Zone and an EU Data Zone planned for the autumn, addresses data residency. Clients procuring long-duration frameworks should specify audit log retention periods, export rights and provenance evidence in contract terms rather than relying on platform defaults, and should ask suppliers how automated decisions affecting safety-critical information are reviewed.
8. How much of the Haskell result should other contractors expect? Haskell’s compression of submittal reviews from seven days to ten minutes reflects an early adopter working directly with Procore across well-organised project data. Outcomes elsewhere will track data quality closely, since agents reasoning over inconsistently named, incomplete or externally held documents produce correspondingly weaker results. Firms should baseline current turnaround times across a representative sample of packages before deployment, then measure the same workflows afterwards, rather than treating vendor benchmarks as a forecast. Starting with a high-volume, well-documented workflow such as submittals or RFIs is more informative than beginning with safety or schedule analysis, where the inputs are messier and the outputs harder to verify.

Strategic Takeaways

1. The commercial shift is from selling capability to accumulating encoded company method, which converts a replaceable licence into an asset that compounds on one platform and makes switching progressively harder.
2. Skills appears to be configuration rather than model training, so the protection contractors need is contractual rather than technical, and exportability of encoded procedures and audit logs should be settled at renewal rather than at exit.
3. Consumption pricing becomes durable only once retention exists, which explains why encoded method arrives alongside metering rather than after it.
4. Restricting Agent Studio to the Enterprise tier limits programmability to the largest accounts, and whether Skills eventually become shareable between organisations is the clearest available test of genuine operating-layer ambition.
5. Codified procedure scales error as efficiently as it scales good practice, making named ownership, versioning and review cycles for encoded standards a governance requirement rather than an administrative preference.