In 2024, Constrafor partnered with Forethought AI to tackle a problem that every construction finance team knows well: the data exists, but getting to it costs too much time. Tommy Rosenkranz, VP of Product at Constrafor, joined Forethought's Implementation Lead Rowen Witt to walk through how the conversational layer was designed, tuned, and launched. This post captures the key ideas from that session and explains why this architecture is now the foundation of Cru's AI agents.
Most subcontractors running $10M–$100M in revenue have more financial data than they can act on. Job cost reports, AP aging, retainage schedules, COI status, cash forecasts — it all exists somewhere. The issue is that accessing it requires navigating different tools, running the right reports, and knowing where to look in the first place.
A project manager who wants to know whether a job is trending over budget has to: open the accounting system, pull a job cost report, filter to the right job, and compare actuals to estimates by cost code. That's a five-minute task on a good day. Multiply it across a portfolio of active jobs, and it becomes a half-day exercise that happens monthly rather than daily.
Conversational AI collapses that process to a single question: "What's the cost-to-complete on the Riverside project?" The system understands the intent, retrieves the right data, and returns an answer in seconds — no navigation required.
The architecture Constrafor built with Forethought centers on three components that work together.
When a user types a question in plain language, the system identifies what they're actually asking for. "Are we over budget on Job 412?" maps to a job cost query. "What's our cash position for the next 30 days?" maps to a cash forecasting model. Intent detection handles the translation between natural language and the underlying data queries — so the system understands construction context, not just generic finance terminology.
This is harder than it sounds. Construction has its own vocabulary: retainage, pay apps, schedule of values, cost codes, WIP. A system trained on generic accounting questions will misinterpret construction-specific queries. Constrafor's model was tuned specifically on construction financial language to handle this.
The conversational layer maintains context across a session. If a user asks about Job 412's cost status, then follows up with "and what about the labor budget?", the system understands that the second question is still about Job 412 without requiring the user to restate the context.
This matters because real financial questions aren't isolated — they're part of a diagnostic chain. A project manager investigating a cost overrun will ask three or four follow-up questions before they understand the root cause. Context retention keeps the conversation coherent rather than forcing a fresh query every time.
The most powerful component is orchestration: the ability to call multiple underlying systems in response to a single question. A query about a job's financial health might require pulling data from the accounting system (cost actuals), the billing system (progress billings and retainage), and the project management platform (percent complete) — then synthesizing the results into a coherent answer.
Constrafor's conversational layer does this automatically. The user asks one question; the system calls the right tools, retrieves the right data, and returns a synthesized response. The same mechanism that Forethought uses to orchestrate customer support workflows maps cleanly onto construction finance operations.
The practical impact shows up in three areas where construction finance teams spend the most time.
Instead of pulling monthly cost reports, a project manager can ask in real time: "Which of my active jobs have labor costs running more than 10% over budget?" The system returns a list ranked by variance, so attention goes to the highest-risk jobs first. Problems that would have been discovered at month-end get surfaced in the middle of the month when there's still time to act.
Cru's Job Costing Agent uses this same pattern — it monitors every active project continuously and alerts teams to overruns before they compound, rather than reporting them after the fact. For a deeper look at how this works, see Job Cost Accounting for Subcontractors: A Step-by-Step Guide.
Cash position questions that previously required pulling an AR aging report, a payment schedule, and a cash forecast model can now be answered with: "What's my expected cash position in 45 days?" The AI synthesizes outstanding receivables, upcoming payables, projected billings, and retainage release schedules into a single forward-looking answer.
This is particularly valuable for subs managing multiple GC relationships with different payment terms. The system tracks each relationship and flags cash gaps before they become crises. Learn more about how AI approaches this problem in our guide on how to improve cash flow in construction.
COI management and lien waiver tracking generate constant questions: "Does sub-tier contractor Martinez have current coverage on the downtown project?" Answering that manually means checking the compliance tracker, verifying the expiration date, and confirming the coverage limits — a task that takes 5–10 minutes per question and happens dozens of times per project.
A conversational layer connected to the compliance database answers it instantly and accurately, every time. Cru's COI Agent extends this capability by proactively alerting teams to expiring certificates before they create project risk. For more on construction compliance automation, read Construction Compliance Software: Automate Lien Waivers and COI Tracking.
The session Tommy and Rowen ran in 2024 was both a technical walkthrough and a proof point: conversational AI applied to construction finance data isn't a research concept, it's an architecture that ships. The design patterns from customer support AI — intent detection, context retention, tool orchestration — translate directly into construction financial operations because the underlying problem is the same: understanding what a user needs, finding the right data, and returning a useful answer.
Cru's current suite of AI agents is built on this foundation. The Cash Forecasting Agent, Collections Agent, Materials Agent, Job Costing Agent, COI Agent, and Bookkeeping Agent each handle a specific domain — and all of them are accessible through the same conversational interface that lets subcontractor teams ask questions in plain language rather than navigating reports.
If you're still spending time hunting through dashboards for answers your data already contains, Cru gives you a better way. Book a demo to see the conversational layer in action →
Conversational AI in construction finance is a natural-language interface that lets project managers and finance teams query their financial data by asking questions in plain English, rather than navigating dashboards or running reports. Instead of pulling a job cost report, you ask "Is Job 412 over budget?" and get an immediate, synthesized answer drawn from your actual data.
Cru's AI agents — covering Cash Forecasting, Collections, Materials, Job Costing, COI compliance, and Bookkeeping — all operate through a conversational interface. Users can ask questions about their financial data in plain language, and the agents retrieve, synthesize, and return contextual answers without requiring manual report navigation. The agents also proactively alert teams to issues like cost overruns or expiring COIs.
A conversational data layer sits between users and their underlying data sources — accounting systems, project management platforms, compliance databases — and translates natural-language queries into data retrievals. It handles intent detection (understanding what the user is asking), context retention (remembering the conversation thread), and tool orchestration (calling the right systems and synthesizing the results).
Constrafor partnered with Forethought AI to build a conversational layer tuned specifically for construction financial language. The system uses intent detection trained on construction finance vocabulary, context retention to maintain coherent conversations, and tool orchestration to pull data from multiple underlying systems in response to a single question. The architecture was designed by Tommy Rosenkranz, VP of Product at Constrafor, and is now the foundation of Cru's AI agent suite.
