- Your CRM is losing you money right now, quietly. Leads older than six months almost never get a second look — until an agent is built to reactivate them on a recurring schedule.
- An AI agent isn’t a chatbot you talk to. It’s a system built from a “system instruction” that reads like a job description: a role, a set of triggers, and a standard operating procedure for every situation it might hit.
- The instruction is the entire game. The more detail you put into role, triggers, and edge cases, the less the agent “misses” — and there’s no shortcut around writing it out in full.
- Data privacy splits into two separate questions. Can this data leak, and will it train someone else’s model. Both have the same fallback answer: a local, self-hosted model.
- The fastest win available to you this week is small. One custom GPT trained on your own past proposals, built in an afternoon, can free up roughly an hour of your day.
Every week, we open a live Q&A for AEC professionals trying to figure out where artificial intelligence actually earns its keep in their business. This week’s session, “How AEC Firms Can Use AI Agents for Sales and Proposals,” was hosted by our co-founder Ian Arden, with Nika Zakharuk running the chat and fielding questions from the room.
The attendee list looked like a cross-section of the whole industry: a principal running civil and land-development engineering out of Toronto, the president of a 50-person architecture firm that specializes in historic preservation, the head of design at a UK-based MEP engineering firm working on parts of the HS2 high-speed rail line, and an architect finishing a second master’s degree who wanted to know where AI fit into his next career move. Different firms, different continents, same underlying question: where does this actually save me money?
Ian didn’t lead with theory. He led with an admission most sales leaders don’t say out loud: most CRMs are, in his words, “majorly dormant.” A lead goes cold after a few months, and nobody circles back — not because the lead is bad, but because nobody’s job is to remember it. That single observation became the spine of the entire session.
What an AI agent actually is (and why it isn’t a chatbot)
Before Ian showed a single screen, he defined the term precisely, because most of the confusion around “AI agents” starts with people picturing a fancier chat window. It isn’t that. Borrowed from computer science, an agent is an entity that perceives its environment, forms a plan using some form of intelligence, acts on the world to move it toward a goal, checks whether that goal was met, and then goes back to sleep until the next trigger wakes it up.
Translated into business terms, that means three ingredients: a trigger (a new WhatsApp message, a scheduled check every half hour, a lead that’s gone quiet for six months), a memory (nothing mystical — just a database or text file the agent reads from and writes back to), and a set of tools it can actually use to act, like updating a CRM record or sending an email.
The shift that matters most is where the human sits. Most people assume AI needs a “human in the loop” — someone approving every step. The agents Ian builds put the human on the loop instead: watching the system, correcting it when needed, but not clicking send on every message.
Case study: the agent that never lets a lead go cold
The clearest, most concrete example from the session was the one Ian’s own team runs internally — a “lead nurturing director” agent built to solve exactly the dormant-CRM problem he opened with.
Here’s what it’s actually plugged into. The agent connects to the firm’s CRM, which is organized as an Airtable instance, and it can both read and update deal records directly. It receives call recordings the moment a sales call ends and processes them for details. It picks up new WhatsApp messages and website chat the instant they arrive. It watches a shared inbox, so anything sent to the firm’s general contact address gets a response without a human ever opening it first. It even receives leads straight from advertising platforms the moment someone fills out a form on Meta or Google.
When a lead needs attention — either because something new just came in, or because a recurring sweep found a deal that’s gone quiet — the agent doesn’t just fire off a generic message. It hands the job to a sub-agent that researches the lead using publicly available information: who they are, what company they represent, their role, their general location. Only after that research comes back does it decide on an outreach strategy and draft the actual message, whether that’s an email or a WhatsApp reply.
Quick tips for building your own lead-reactivation agent
- Consolidate your lead data into one system before you automate anything. If your leads are scattered across someone’s inbox, a spreadsheet, and their memory, the agent has no single source of truth to read from or write back to.
- Give it a research step, not just a script. The version of this agent that actually converts looks the lead up before it writes anything — generic outreach reads as generic, and prospects notice.
- Define your triggers as specific rules, not vague intentions. “Sweep every deal with no planned activity every Monday” is a trigger an agent can run on. “Follow up sometimes” is not.
- Add a deduplication step. Ian’s setup includes a second agent whose only job is keeping new contacts from creating duplicate records — a small addition that keeps the rest of the automation from making a bigger mess the longer it runs unsupervised.
- Don’t throw away old leads. The leads sitting dormant in your pipeline already have brand trust built up, at zero acquisition cost. An automated re-engagement sweep is often the highest-return work an agent can do, because you’re not paying for the attention twice.
The real blueprint: what’s actually inside a system instruction
An attendee asked the question everyone in the room was clearly thinking: is there a template for this? Ian’s answer was direct — there’s no golden template, because every business process is different. But there is a consistent anatomy to follow, and he walked through it piece by piece.
Think of the system instruction as a full job description, not a prompt. It opens by naming the role — “you are the expert customer relationship management specialist” was his own example — and immediately follows with the primary objective and, just as important, the non-goals: what this agent is explicitly not supposed to decide on its own, so it knows when to leave something for a human.
From there, it lists every trigger the agent might receive and a standard operating procedure for each one: how to identify who sent a message, how to classify what kind of message it is, which tool to use to respond, and what to write back. Ian’s example was granular on purpose — an agent handling a new booking needs to know how to pull contact details, create the deal record, decide what communication channel to use, and even judge something as specific as whether it’s an appropriate hour to message someone, inferred from the time zone implied by their phone number. “That has to go really in detail,” he said, “because every aspect matters.”
The instruction closes by defining what the agent should write down after it acts — what goes into memory for next time — so the system gets more situationally aware with every interaction instead of starting from zero each time.
Quick tips for writing your own system instruction
- Write it like you’re training a new hire on their first day, not like you’re chatting with a search engine. Name the role, the objective, and what’s off-limits.
- Spell out every trigger you expect and a specific procedure for each one. Vague triggers produce vague, inconsistent behavior.
- Budget real time for iteration. Ian was candid that a good instruction takes “a few hours of writing” plus trial and error — you run the agent, see where it misses, and augment the instruction to close that specific gap.
- Decide what memory should hold before you launch it. A short, structured summary of what happened with each contact is enough — it doesn’t need to be a novel, just enough context for the next run to pick up smoothly.
The data privacy question, answered in two parts
No AI session in this industry gets through without someone asking about data privacy, and this one was no exception. Ian split the concern into two genuinely separate questions, because they have different answers.
Risk of a data breach
If your firm already runs on Microsoft, Google, or Autodesk’s cloud, connecting an AI tool doesn’t introduce a fundamentally new exposure — it’s the same category of risk you’re already carrying. For larger firms or sensitive projects, the real mitigation is a locally deployed model, so no data ever leaves your environment. Open-source models now benchmark close to the leading commercial ones.
Will it train on your data?
This depends on your provider’s terms of service. Larger AI providers are “very adamant,” in Ian’s words, about not training on enterprise customer data — but that protection typically lives in the enterprise tier, not the free plan. Local deployment solves this question too, if it matters enough to your firm.
How to decide what to automate first
With the mechanics covered, the practical question became: where do you even start? Ian’s advice was to resist the urge to automate everything you can imagine and instead work the process manually first, so you actually understand its shape before you hand any of it to an agent.
From there, build what he called a heat map of your own business processes — a simple exercise to spot where the most value is created, and separately, where the most time and money currently leak out. Those overlapping hot spots are where you automate first, because that’s where a single automation returns the most value for the least effort. Only after that do you move to designing the actual agent: what information it needs coming in, what tools it needs to act, and what it should hand back out.
That shift, done a process at a time, is what slowly turns a traditional AEC firm into what Ian called a more “AI-native” company — and he was upfront that it’s a change in how your team thinks about work, not just a new tool on the desktop.
The mindset shift nobody wants to say out loud
The most candid moment of the session came when Ian talked about his own discomfort with this shift. He described himself as a software developer by background who, these days, doesn’t write code by hand anymore — he governs and steers AI to make the right calls, and lets it execute.
“It’s kind of hard to admit and accept, because it removes part of the necessity in us, the humans.”
— Ian Arden, VitruAI co-founder
His broader point was that the human role isn’t disappearing — it’s moving up a level, from executing the work to supervising it. And that supervision is its own real job, not a formality: AI can now generate an immense volume of output, and reviewing thousands of lines of anything on a daily basis is its own form of cognitive overload. His answer was to build supervision agents for that too — feeding them your strategies, policies, and standard operating procedures so they can flag what a human actually needs to look at, rather than asking a person to read everything the system produces.
Want a copy of the system instruction Ian referenced, or help mapping your own sales process into an agent? Reach out to Ian or to Nika, our account manager — we’re glad to walk through it with you.
Book a session → Join next week’s Q&A
FAQ
What’s the single easiest automation for an AEC sales team to start with?
A custom GPT trained on your own past proposals. It’s small, it’s ROI-positive almost immediately, and Ian’s estimate was that it can free up roughly an hour of a business developer’s day.
Is my data actually safe if I connect AI to my CRM and documents?
It comes down to two separate questions — the risk of a breach, and whether a provider trains on your data. Both have the same fallback: a locally deployed model that never sends your data outside your own environment.
Why does an AI agent still make mistakes if the technology is this capable?
Almost always because the system instruction wasn’t specific enough. An agent without a template to follow, a defined trigger, and a clear procedure will guess — and guessing is where mistakes come from.
Do I need an IT team or a developer to build one of these agents?
No. Ian’s examples were all built using a folder’s system instructions or a custom GPT inside tools like ChatGPT — no code, no IT ticket required.
How long does it actually take to write a good system instruction?
Longer than a first prompt, and that’s the point. Expect a few hours of writing plus a handful of test runs where you notice what the agent missed and add exactly that detail back in.
The takeaway
The most memorable line of the session wasn’t about a tool at all — it was Ian pointing out that a CRM full of six-month-old leads isn’t a dead asset, it’s brand trust you’ve already paid to build and simply stopped using.
You don’t need to build a lead-nurturing director agent this week. Ian’s actual closing challenge was smaller: open the AI tool you already have, build one custom GPT around the most repetitive thing you do — for most people in sales, that’s proposal writing — and see what an hour back in your day is worth.
This post recaps our weekly live Q&A, hosted by Ian Arden with Nika Zakharuk. Want to join the next one or talk through your firm’s AI roadmap? Reach out to Ian or Nika, our account manager.