Last month, I watched a consulting firm owner spend four hours writing a single proposal. Four hours — for a document that follows roughly the same structure every time.
That moment crystallized something I'd been thinking about for months: consulting firms are sitting on a goldmine of repetitive, high-value work that AI agents can handle in seconds.
So I built AgentDesk. And what happened next surprised even me.
The Problem Nobody Talks About
Consulting is a margins game. You sell expertise, but a shocking amount of your billable hours get eaten by operational work:
- Client intake: Reading inquiry emails, qualifying leads, drafting responses. 30-45 minutes per lead.
- Proposals: Translating discovery calls into polished proposals. 2-6 hours each.
- Status reports: Compiling project data into executive-friendly reports. 1-2 hours weekly per client.
For a firm handling 10 active clients and 20 inbound leads per month, that's easily 40+ hours of non-billable work. Every month. That's an entire employee's week, gone.
What I Built (and How)
AgentDesk is a SaaS platform with three pre-built AI agents, each designed for a specific consulting workflow:
1. Intake Agent
Reads incoming inquiries from any source — email, form submissions, chat. It qualifies the lead on a 1-10 scale, drafts a personalized response in your firm's voice, and recommends a next action (book a call, request more info, or politely decline). Cold leads get filtered automatically so your team only sees qualified prospects.
2. Proposal Agent
Feed it discovery call notes or a meeting transcript and it generates a complete, professional proposal: executive summary, understanding of the client's situation, proposed solution with specific deliverables, timeline with milestones, investment breakdown, and next steps. The output is structured markdown ready for PDF conversion.
3. Report Agent
Give it your project metrics, milestones, and notes. It produces an executive-friendly status report with key metrics, progress indicators, risk flags, and recommendations. Written for the busy executive who has two minutes to read — data-first, not fluff.
The Technical Approach
Under the hood, AgentDesk runs on a stack I chose for speed and reliability:
Claude Sonnet 4 API via the @anthropic-ai/sdk — each agent has a tuned system prompt and uses Claude's tool-use capability. The Intake Agent runs at temperature 0.3 for consistent qualification scoring. The Proposal Agent runs at 0.4 with a 4,096-token output window to generate comprehensive proposals without truncation.
Next.js 16 + React 19 — the dashboard is server-rendered for fast initial loads, with client-side streaming for real-time agent output. Users see the proposal being written token by token, which turns a 30-second wait into something that actually feels fast.
Stripe integration via the Stripe Node SDK for subscription billing. Three tiers: Starter ($99/month), Professional ($349/month), and Agency ($799/month). The free trial requires no signup and no credit card — you can run all three agents right now.
TypeScript end-to-end — fully typed from the API routes to the agent engine to the frontend. The agent engine (agent-engine.ts) defines typed interfaces for AgentConfig, AgentTask, and AgentAction, so every agent interaction is predictable and debuggable.
Here's a simplified version of the core agent config:
interface AgentConfig {
id: string;
name: string;
systemPrompt: string;
tools: Anthropic.Messages.Tool[];
maxTokens: number;
temperature: number;
}
Each agent is a config object with a system prompt and tool definitions. The engine handles the Claude API interaction, tool-use loops, and structured output parsing. Clean separation means adding a new agent is just adding a new config — no engine changes needed.
The 30-Second Number
When I say proposals generate in 30 seconds, that's the real number. Claude Sonnet 4 with streaming typically returns a full 3-page proposal in 25-35 seconds. The Intake Agent is faster — lead qualification and response drafting usually completes in 10-15 seconds.
For context, the consulting firm owner I mentioned earlier was averaging 4 hours per proposal. Even if you cut that to 1 hour for a fast writer, you're looking at a 120x speed improvement. And the output is consistent every time — no more "Friday afternoon proposals" that are half the quality of your Monday work.
What I Learned Building This
1. System prompts are everything. The difference between a generic AI response and a genuinely useful proposal is entirely in the system prompt. I went through 15+ iterations on the Proposal Agent's prompt before the output consistently matched what a real consulting firm would send.
2. Tool use makes agents real. Without Claude's tool-use capability, these would just be fancy chatbots. The structured tool definitions force the model to produce typed, predictable output — qualify_lead returns a score, qualification level, suggested response, and recommended action. Every time.
3. Consulting firms are underserved. Most AI SaaS targets developers or marketers. Professional services firms — consultants, agencies, fractional executives — are still doing most of their operational work manually. They're ready for this.
4. No-signup trials convert. Removing the signup wall for the free trial was the best decision I made. People want to see what the agents produce before they commit to anything, even an email address.
Try It Yourself
The free trial is live with no signup required. You can run all three agents right now:
Try AgentDesk Free -- agentdesk-inky.vercel.app/dashboard
Paste in a sample client inquiry and watch the Intake Agent qualify it. Drop in some discovery call notes and get a full proposal back in 30 seconds. If you run a consulting firm or build software for one, I'd genuinely love your feedback.
The full product page is at agentdesk-inky.vercel.app.
Building in public. If you're working on AI agents for professional services — or thinking about it — I'd love to connect in the comments.
Top comments (0)