Watch the full walkthrough (59 min): the memory files, the permission model, the morning sync, and how to keep the whole thing low risk.
Every morning before the first meeting, a system reads the overnight email, combs the project management tool for slipping timelines and buried comments, scans team messages, reviews the day’s calendar, and produces a short brief: here is what changed, here is what needs you, here is where to focus.
That is not a product. It is a setup anyone can build, and it runs on three parts: context, permissions, and recurring jobs. For the wider landscape, see AI for Construction: The Complete Guide.
The recurring job is the one people skip, and it is the most important. You can build perfect context and wire up every permission, but if nothing runs on a schedule it never feels like an assistant. It just feels like another tab.
Why This Is Worth Building
The complaint from construction firms is remarkably consistent. Information is fragmented across systems. One person holds the login that everyone else needs answers from. And roughly 60 percent of the day goes to work about work: getting an email, then searching for the document, then researching the history, then assembling the answer, then finally replying.
That is the target. Not replacing judgment. Removing the search.
The other reason to build it is that a well-fed AI assistant sees things a person cannot. Not because it is smarter, but because it can read everything, every day, without getting tired. Comments buried three levels deep in a project management tool. A timeline slipping by two days. A support pattern forming across three tickets. An assistant reading all of it every morning surfaces the one that matters.
Part 1: Context, and Why It Lives in Markdown Files
Here is the thing most people get wrong. The desktop agents that actually do work do not have persistent memory the way a chat window does. Close the session and it is gone.
So the memory has to live somewhere, and that somewhere is plain markdown files on your computer. Simple text documents telling the AI who it is, what it knows, and how you work. Human-readable, not pretty, and not really written for you. The compounding version of this idea is in Build an AI System That Gets Better Over Time.
The fastest way to create them is to not write them. Ask the AI to interview you: tell it you want to build a personal assistant and have it quiz you. Answer the questions. It writes the files. Twenty questions gets you most of the way. Going deeper gets you more.
A working structure usually ends up with a few distinct pieces:
- A home file with universal rules. Who the assistant is, who you are, your role, your company, preferred output formats, and core objectives.
- Role and preferences. Your work style, priorities, and terminology. Construction runs on acronyms and no two firms speak identically, so let it quiz you on your language.
- Daily brief format. Be specific here. Given open access, it will produce pages of detail you stop reading by week two. Ask it to draft ten sample morning briefs, pick the best one, and lock that in.
- A projects folder. For a construction company this is the biggest lever. One file per project, brought current every morning from the sources you connected.
- Travel and logistics, if you want it handled.
- A decision log. More on this below.
One piece of advice that saves a lot of wasted effort: do not micromanage the file structure. Let the AI organize its own memory. Your brain’s filing system is not necessarily the one it retrieves from best, and there is no single correct approach. Some teams run one enormous markdown file and nothing else. If you never have to open these files, the structure is working.
The Decision Log Is the Real Unlock
This one takes months to discover, so take it for free.
AI is non-deterministic by nature. It follows workflows well. What it does not naturally do is decide. Send it an email saying “set up this project” and there are twenty small judgment calls between that sentence and a finished setup.
Memory files handle the workflows: here is who you are, here is the SOP, follow it. A decision log handles the judgment. It records key decisions, the dates, the stakeholders, and the reasoning behind them.
Once that history exists, the assistant stops being agreeable and starts being useful. It can look at a plan and say this timeline looks reasonable, but you pushed back in these three places the last time something like this came up. That is a team member, not a yes-man, and the difference comes entirely from having precedent to reason against.
Part 2: Permissions, Connected Deliberately
Both major AI platforms ship built-in connectors. Use them where they cover what you need.
They will not cover everything. Connecting meeting recording transcripts, team chat, shared drives, email, and a project management system all at once, with real permissions in each, usually means a connector service that sits between the AI and your tools and lets you choose exactly what each connection can see and change.
Two cautions. Do your own diligence on any connector service before you route company data through it, and do not install plugins from sources you cannot verify. That is the most common way people get burned with AI. Look for the same signals you would use anywhere else: real adoption, a real company, verifiable reputation.
The permission model itself has three levels, and they map cleanly onto trust:
Read-only. It can see and summarize. It cannot change anything.
Ask for permission. It prepares an action and waits for your click.
Always allow. It acts on its own.
Go read-heavy before write-heavy. Start everything at read-only and live with summaries, retrieval, and draft generation for a while. Then open specific write permissions with approval required. Reserve always-allow for low-stakes, high-repetition work.
Regardless of where you set the dial, keep human approval at the boundaries: any external email leaving your name, any payment, any final commitment. Those are the actions you cannot take back.
Two more controls worth turning on. Require source visibility, so answers come with citations or links back to where the information was found, which is how you catch hallucination on your own data. And use the platform’s data controls, including zero data retention where it is offered, so your documents are not retained or used for training.
Part 3: The Recurring Job
This is what turns it from a tool into an assistant.
In practice, it is a saved routine you invoke the same way every day, either on a schedule such as 7 a.m., or with a single command when you sit down. Some platforms call these skills. You already know them as SOPs.
Building one is a conversation. Tell it you want to create a morning sync, then walk it through the sequence: comb my email for everything since your last sync and flag what needs action today versus later; go into the project management system and summarize the last 24 hours as it relates to my role; check these specific project folders on the shared drive and flag anything that changed; update the project files with what you learned.
Note the last step. It is not just reporting. Every sync rewrites the memory, deleting what is stale and adding what is new, so the assistant is genuinely current rather than reconstructing from scratch.
The compounding effect is the point. Because it reads what you are working on every day, it starts anticipating. You agree to speak at a conference in an email, and it adds the travel booking to your list with three options, because it already knows your home airport, your drive time, and how you like a calendar event built, right down to putting the confirmation code in the event so you are not digging for it the night before check-in.
Keeping It Low Risk
Connect the minimum number of sources that make the assistant useful, not everything you have access to. Start read-only. Keep approval gates on anything external. Require citations.
For federal and regulated work, the architecture has a real advantage: markdown memory files live on your machine, not in a vendor’s cloud, and are accessed by an API call rather than uploaded. Teams under CMMC obligations have handled this with a dedicated machine kept separate from the compliance boundary, and CMMC for Construction Contractors covers what applies when. At the far end, you can run a model locally on the server you already have in the closet and never touch a cloud at all. Whatever route you take, involve your compliance consultant before, not after.
Your First Four Steps
- Pick one domain. A daily operations brief is the right first one for most people.
- Create the folder. One dedicated directory where the assistant keeps its memory files. It needs somewhere to write.
- Let it interview you. Say what you want it to do and what you plan to connect, then answer its questions and let it write its own files.
- Run the sync and refine. Run it, read the output, tell it what to add and what to stop reporting. It will be genuinely useful within a week.
Which platform to use comes down to one trade. The more consumer-friendly assistant is faster to set up and lighter on configuration. The more capable one takes longer to learn and gives you more control and more ceiling. Both will save you real time. Pick based on how much setup patience you have. If you are comparing surfaces, see Claude Chat, Cowork, or Code for the Back Office for Claude by Anthropic and ChatGPT Work for Construction Back Offices for OpenAI.
Frequently Asked Questions
What are markdown memory files and why does an AI assistant need them?
They are plain text documents on your computer that hold what the assistant knows: who you are, your role, your projects, your preferences, and its own accumulated notes. They are necessary because the desktop agents that do real work do not keep persistent memory between sessions the way a chat window does. The files are the memory. The fastest way to create them is to ask the AI to interview you and write them itself.
What is a decision log and why does it matter?
A record of key decisions, when they were made, who was involved, and the reasoning. It matters because AI follows workflows well but does not naturally make judgment calls. Given a history of how you have decided similar things before, the assistant can reason from precedent, push back on a plan, and handle situations its SOPs do not cover.
How do I connect an AI assistant to my email, drive, and project management system safely?
Use the platform’s built-in connectors where they cover what you need, and a reputable connector service where they do not. Connect only the minimum set of sources, start every connection read-only, and open write permissions one at a time with approval required. Keep hard approval gates on external email, payments, and final commitments. Verify any connector service before routing company data through it.
Can construction firms under CMMC use an AI assistant?
With care and with your compliance consultant involved. The helpful property is that markdown memory files live on local machines rather than in a vendor cloud and are reached by an API call. Some teams keep a dedicated machine outside the compliance boundary for this work. A firm that wants to avoid the cloud entirely can run a model locally on its own server. Do not make this call without your consultant.
How long does it take to build an AI assistant that is actually useful?
Reaching a genuinely useful morning brief typically takes about a week of daily refinement: run it, read the output, tell it what to add and what to stop reporting. Setup itself is short. The value comes from the iteration, and from the memory files getting richer every time it syncs.
The Back Office Version, Already Built
A personal assistant is a great first build. The contract and compliance workload is a harder problem, and Contracts Connected already runs it: an AI agent that works a shared inbox, collects documents from subcontractors, reviews certificates against your requirements, and escalates to your team when it needs a decision. If you are mapping out the specialists, start with The 3 AI Agents Every GC Should Be Running.
No pitch deck. Straight conversation about what you’re managing and whether we can help.