AI Agents for AEC Sales and Proposals in 2026: A 5-Stage Playbook From Our Weekly Q&A

This all started with one question: where do we even begin with AI?

Every week, we run a live Q&A session for AEC professionals who are trying to figure out where artificial intelligence actually fits into their business. This week’s session was titled “How AEC Firms Can Use AI Agents for Sales and Proposals,” and it was hosted by our co-founder Ian Arden, with Roxy Zakharuk running the room and fielding questions from the chat.

The audience was a real mix — architects, BIM managers, engineers, some who’d never touched an AI tool before this call, some who’d already built their own review workflows on top of Claude. That range is exactly why the session is worth writing up. Ian didn’t just show demos. He laid out a mental model for AI adoption that applies whether you’re sending your first prompt tonight or already running agents in production.

Here’s the playbook, in the order he walked through it.

What is an AI agent, and why does it matter for AEC sales?

Before you automate anything, it helps to know what you’re actually building. An “agent” is a term borrowed from computer science: an entity that perceives its environment, decides what to do, and acts on the world to move it toward a desired state. In business terms, that means an AI system with three things — a trigger (a new email, a scheduled Monday run, a lead that’s gone quiet), a memory (it can write data down and read it back later), and a set of actions (it can send a message, update a deal, generate a document).

The important shift is where the human sits. Most people think in terms of “human in the loop” — someone has to click a button before anything happens. The agents Ian described put the human on the loop instead: supervising the system, not running it one click at a time.

Why now? Context windows — how much conversation an AI model can hold onto at once — have grown to roughly a million tokens on the more advanced models, which is somewhere around 750,000 words. Ian compared where we are to the automobile industry at the turn of the 20th century, when skeptics dismissed the first cars as an expensive way to replace two or three horses. In hindsight, that “expensive replacement” reshaped entire industries. The capability curve for AI agents is on the same trajectory — nascent right now, compounding fast.

The AI ascension path, at a glance

Ian framed the whole session around one progression. Each stage below builds on the last — most firms are further along than they think on the first one, and further behind than they’d like on the last.

• Stage 1 Advanced tool usage — get precise with the AI tools you already pay for: ChatGPT, Claude, Gemini, Copilot Studio.
• Stage 2 AI assistants & folders — turn one good prompt into a standing system instruction the AI never forgets.
• Stage 3 Automate one repetitive task — start with proposal generation, your highest-ROI, simplest win.
• Stage 4 Autonomous agents & automations — let the system react to events in your business without anyone clicking send.
• Stage 5 Holistic business autonomy — redesign the org chart itself around agents doing the repeatable work.

Stage 1: Get fluent with the AI tools you already have

This is the stage almost everyone starts at, and almost everyone underuses. The tools — ChatGPT, Claude, Gemini, Microsoft Copilot Studio — are more capable than most people’s prompts give them credit for.

The reason AI feels “uncontrollable” to newer users isn’t that it’s unpredictable. It’s that the prompt wasn’t specific enough. The fix is almost boringly simple.

Quick tips for taming AI with better prompts

• Spell out the role. Tell it exactly what it should act as — a systems strategist, a proposal reviewer, a co-founder — before it drafts anything.
• Structure the thinking, not just the output. Ask it to reason step by step, then format the answer a specific way, rather than just asking for “an answer.”
• Go longer, not shorter. The longer and more specific the instruction, the more predictable — and less “hallucinated” — the output.
• Remember every message drags the whole conversation with it. Every send carries the full prior thread into the model’s context. The longer the conversation runs, the more of that context window gets used before your model even sees your new question.
• Check your personalization settings. Most AI tools have a custom-instructions tab that acts as a standing system prompt. Ian’s own setup tells the model his working style — he described himself as ADHD, fast-moving, ideating in bursts — so every response adapts to how he actually processes information, plus it slips him one new Arabic word per reply while he’s living in the UAE.

Stage 2: Build AI assistants and folders

Once you’re comfortable steering a single conversation, the next stage is making that steering permanent instead of repeating it every time.

The simplest version: create a folder for a recurring type of work — Ian’s example was a folder for reviewing incoming drawings and documentation — and write the system instruction once in the folder’s settings. From then on, every new conversation in that folder inherits the instruction automatically. You just drop in the file.

The more powerful version is a custom GPT (or the equivalent “assistant” in whichever tool you use): a small, named worker built for one job. It gets a system instruction describing your firm, your services, and your tone; conversation-starter buttons for the tasks it handles; uploaded knowledge like your proposal templates and pricing; and, if you want it, tool access like web search or document generation.

Quick tips for building your first AI assistant

• Name it, describe it, and give it a face. A custom GPT with a name and a description behaves like a “worker,” not a generic chat window — that framing matters more than it sounds like it should.
• Write the system instruction like a job description. Cover the role, the products or services it should know, the behavior you expect, and what to emphasize (Ian’s example: always lead with ROI and benefits for the client).
• Upload the templates, not just the instructions. Proposal drafts, typical emails, service sheets, and corporate policy — the AI only knows what you give it.
• Turn on the capabilities you actually need. Web search for prospect research, document generation for output, a code interpreter if your proposals involve calculations.

One attendee asked whether this is basically the same idea as “skills” inside Claude. Ian’s answer was straightforward: yes. Different products use different names — skills, SOPs, job descriptions — but the underlying idea is identical. You decompose what a person would need to know into individual pieces, and a root file governs which pieces get loaded depending on the situation, instead of stuffing everything into every single prompt.

Stage 3: Automate one repetitive task — start with proposal generation

This is the stage where AI adoption stops being theoretical. Ian’s advice for picking your first real automation came down to two criteria: return on investment, and simplicity.

Concretely: pick something you do every day, not once a quarter. Daily, repetitive tasks pay back immediately because you feel the time savings the same week. Quarterly tasks don’t teach you anything fast enough to be worth automating first.

For AEC sales and marketing, proposal generation is the obvious starting point. It’s repetitive, it’s high-stakes enough to matter, and it’s simple enough to automate with the assistant you just built in Stage 2 — feed it your services, your templates, and your ROI framing, and let it produce the first draft so your team edits instead of starts from a blank page.

Quick tips for automating proposal generation

• Feed it the winning pattern, not just a template. Give the assistant your best past proposals, not a blank skeleton — it needs examples of what “good” looks like for your firm specifically.
• Bake in the ROI framing up front. Ian’s system instructions explicitly told the model to lead with return on investment and client benefit — don’t leave that to chance.
• Let it touch visuals too. The same AI tools that draft text can generate on-brand imagery once you’ve fed them your visual style, which matters for both proposals and marketing.
• Measure the time saved on the first task before adding a second. That’s your proof that the approach works before you scale it.

Stage 4: Move into autonomous agents that don’t wait for a click

Here’s where most firms haven’t gone yet — and where the actual leverage lives. The first three stages all share one limitation: you still have to be there, and you still have to hit the button.

This stage removes that. An autonomous agent reacts to events in your business instead of waiting for you to open a chat window. A new email arrives. A new WhatsApp message comes in. A lead hasn’t been touched in six months. Monday morning arrives and a report needs to land on someone’s desk without anyone remembering to ask for it.

The clearest example from the session was CRM reactivation. Most firms’ CRMs are, in Ian’s words, “majorly dormant” — leads six months old or older stop getting attention, and that’s a direct, measurable revenue leak. The fix Adaia built for its own sales process is a “lead nurturing director” agent: it’s connected to the CRM, it can read and update deal stages, research a lead using public information, draft outreach over email or WhatsApp, and reschedule the next follow-up — all on a recurring trigger, with no one manually writing to a client.

Quick tips for your first autonomous agent

• Consolidate before you automate. If your leads live across WhatsApp, email, and someone’s memory, structure that into one CRM first — the agent needs one source of truth to work from.
• Define the triggers explicitly. “Every half hour, check for new calls.” “When a new lead fills out a form.” “Every Monday, sweep leads with no planned next activity.” Vague triggers produce vague agents.
• Give it research access, not just a script. The most useful version of this agent looks up public information on a lead before it writes anything, so the outreach isn’t generic.
• Add a duplication check. A second agent that merges or deduplicates new contacts before they hit your CRM avoids a mess that gets worse the longer autonomy runs unsupervised.

Stage 5: Design for holistic business autonomy

The last stage is less about tools and more about how the org chart itself changes. Once several of your processes are running through agents instead of people, you start organizing work into a few repeatable patterns instead of ad hoc automations:

• Task dispatch — one coordinator agent splits work across several sub-agents, then assembles the results.
• Control loop — one agent tasks another and reviews what comes back before it goes further.
• Reflective loop — a third agent checks the quality of the work and improves the instructions for next time.
• Sequential — work moves stage by stage, each agent picking up where the last one left off.

The role of the humans in this picture doesn’t disappear — it moves up a level. Instead of executing the process, people supervise it, often with the help of their own AI supervision agents that track what the other agents are doing. During the session, Ian shared his own firm’s long-term bet: that a large majority of the routine work in the AEC industry will eventually run through AI, freeing the professionals in that industry to spend their time on the judgment calls only a human can make.

Book a session → Join next week’s Q&A

FAQ

Are custom GPTs the same thing as “skills” in Claude?

Yes, functionally. The name differs by platform, but the mechanism is the same: you break a job down into instructions, and the system loads only the relevant piece instead of stuffing everything into every prompt.

What’s the easiest first automation for an AEC sales team?

Proposal generation. It’s repetitive, high-stakes enough to matter, and simple enough to automate with a basic AI assistant and your own templates.

Why does AI still make mistakes if it’s this capable?

Mostly because the prompt wasn’t specific enough. The fewer assumptions you leave the model to guess at — role, structure, format, length — the more consistent the output.

Do I need to worry about data privacy if I connect AI to my CRM and files?

It’s a real concern for larger firms, and there’s a real answer: large language models can run locally, inside your own office environment or private cloud, with no connection to the outside world.

How do I decide what to automate next after my first win?

Score it on two axes: return on investment and simplicity. Take the simple, high-ROI, daily tasks first. Save the complex-but-high-ROI ones for after you’ve built confidence with the easy wins.

The takeaway

At the turn of the 20th century, the skeptics were right that the first cars only replaced two or three horses. They were wrong about everything that came after. The capability curve for AI agents in AEC is following the same shape — unimpressive-looking today, compounding fast from here.

You don’t need to build Stage 5 this week. Ian’s closing challenge to the group was simpler than that: open the AI tool you already pay for, pick the single task you repeat most often, and write it one real system instruction today. If that saves you even one hour tomorrow, the session already paid for itself.

This post recaps our weekly live Q&A, hosted by Ian Arden with Roxy Zakharuk. Want to join the next one or talk through your firm’s AI roadmap? Reach out to Ian or Nika, our account manager.