People expect AI to do everything. The actual best AI implementations do almost nothing—surrounded by a much larger system that does almost everything.
That's the inversion most teams get wrong on day one.
The Story Most Teams Walk In With
A model is going to read the RFP, summarize the risks, draft the clarifications, route the document, file the pricing, and update the tracker. The model will do all of it.
Then they deploy. The model summarizes the risks well, sometimes. It hallucinates a section reference, occasionally. It drafts clarifications in slightly different formats every time. It "routes" the document by mentioning routing in its output. The pricing never gets filed because the model doesn't have permissions to write to the system. The tracker never gets updated because the model doesn't know what the tracker looks like this week.
The team is frustrated. The PE who tested it for two weeks goes back to doing it manually. The director writes off AI for the year.
The story they came in with was wrong.
The Story That Actually Works
A deterministic flow does almost everything. Triggers, file moves, naming, validation, routing, notifications, logging. Boring, predictable, repeatable. The same input produces the same output. Always.
The model is dropped into the middle of the flow to do one specific judgment task. Read this RFP and pull out the risk language. That's it. That's all the model does. The output goes into a structured field. The deterministic flow takes it from there—formats it into a report, routes it to the right person, logs it in the tracker, archives the source document.
The model isn't running the operation. It's a senior estimator brought in for a 90-second consult. The rest of the system runs without it.
This is the version that holds up at scale. Not because the AI is more limited. Because the surrounding system is more rigorous.
Why This Matters On Your Jobs
Construction work has a thousand small repetitive steps surrounding a few moments of judgment. The mistake is treating it as one big AI task.
Take bid-day intake. The judgment moment is interpreting the RFP. Everything around it—logging the email, downloading attachments, naming files, writing to the tracker, alerting BD, scheduling the precon meeting on the calendar, creating the SharePoint folder structure—is deterministic. Every one of those steps should run on rules. The model is only used to read and interpret the document.
If you let the model do the deterministic steps, two things happen. It runs slower. It also gets some of them wrong, eventually. The rule-based parts of the operation should not be subject to the model's vibes on a Wednesday.
The Inverse, Where Teams Get the Other Side Wrong
The other failure mode: trying to do everything deterministically.
You can't write a rule that reads a 200-page spec and tells you the risk language is non-standard. You can't write a rule that interprets messy meeting notes into a clear summary. You can't write a rule that drafts a clarification that's both technically accurate and politically appropriate to a particular GC.
Those are judgment tasks. They need a model. The team that refuses to bring AI into the workflow ends up doing those tasks by hand forever, while the competitor next door runs them in 30 seconds.
The point isn't deterministic versus AI. It's deterministic with AI, in the right places.
How to Decide What Goes Where
Three questions on every step in a workflow.
Does the same input always produce the same correct output? If yes, deterministic. The bid date in the email always goes into the bid date column. No judgment required.
Does the step require interpretation, summarization, or generation of language? If yes, AI. Reading the RFP. Drafting the clarifications. Summarizing the OAC.
Does the step involve a binding commitment—price, scope, schedule, safety, contractual language? If yes, AI does the prep. A human signs.
That's the entire decision tree. There isn't a fourth question.
What This Looks Like in the Next Year
Bid intake. AI extracts GC name, bid date, scope hints, and risks. Deterministic logic logs to SharePoint, names files, alerts BD, creates folder structure, schedules calendar items. The AI runs once per email. The deterministic flow runs constantly.
Estimating handoff. AI highlights scope gaps in proposals. Deterministic rules ensure the gaps get into the leveling matrix and the leveling matrix gets to the buyout meeting. Nothing is lost in translation because the translation isn't manual.
Document processing. AI summarizes specs and flags unusual sections. Deterministic flows file the summaries, version them, and tie them to the project record.
The pattern keeps repeating. AI does the thinking that needs interpretation. Rules do the work that needs reliability. Together, you get systems that are both intelligent and predictable.
The Working Test
If you can write a clear rule for the step, write the rule. Don't make a model do it.
If you cannot write a clear rule because the step requires interpretation, use a model. Surround it with rules.
The teams that internalize this stop chasing magic. They start building plumbing. The plumbing is what produces the results everyone else is still pitching as a vision.