
Lessons from Amelia LeRutte
Amelia LeRutte is SaaStr’s Chief AI Officer and co-host of its firsthand series on operating more than 20 AI agents. Her work spans deployment, training and daily maintenance across sales, marketing and events. — The Agents #001.
Part 1: The Build vs. Buy Equation
- Buy Most, Build the Missing Workflow: SaaStr describes a 90/10 policy: buy most AI capabilities and build the few workflows available products cannot meet. — SaaStr 90/10 Build Story.
- Choose for the Work, Not the Wrapper: SaaStr tried available marketing agents before building 10K because those tools generated content but did not orchestrate the campaigns it needed. — SaaStr 90/10 Build Story.
- Make the CRM Useful to the Agent: Amelia connected 10K to Salesforce so it could read pipeline and revenue and later write back, making integration part of the workflow rather than an afterthought. — SaaStr AI VP of Marketing Playbook.
- A Wrapper Must Solve a Workflow: The 10K case illustrates a concrete buy-versus-build test: packaged agents that only drafted content did not solve SaaStr’s marketing-orchestration problem. — SaaStr 90/10 Build Story.
- Context Before Model Tweaks: SaaStr’s AI SDR guidance emphasizes tested messaging, segmented audiences and current customer context before expecting good agent output. — SaaStr AI SDR Playbook.
Part 2: The Reality of AI Maintenance
- Production Is the Starting Line: SaaStr’s agent deployment exposed outages, subtle wrong-year errors and later model regressions; production use required ongoing checks and fixes. — The Agents #001.
- Budget Human Maintenance Time: SaaStr reported roughly 15–20 weekly hours of human management for its AI SDR deployment, with performance worsening when attention dropped. — SaaStr Six-Month AI SDR Report.
- Watch for Model Regressions: SaaStr’s pitch-deck analyzer began inventing revenue figures after a likely model update even though the application code had not changed. — The Agents #001.
- Read Agent Output Daily: SaaStr recommends reviewing live interactions every day; an agent can keep running while its answers or data quality quietly deteriorate. — The Agents #001.
- Retest After Changes: SaaStr’s analyzer example shows why a previously stable agent must be retested as models and workflows change. — The Agents #001.
- Silent Failure Needs Detection: SaaStr describes an AI adviser that failed to ingest new content for months without a clear alarm, and an internal agent that quietly omitted a high-interest event session. — The Agents #001.
- Keep Humans in the Loop: The team’s results depended on people who knew the business reviewing conversations, revising context and catching errors; these were not zero-headcount deployments. — SaaStr Six-Month AI SDR Report.
- Record Agent Corrections: In Amelia’s 10K build, a project memory file stores goals, voice rules, contacts, send-domain rules and corrections so the agent can use them later. — SaaStr AI VP of Marketing Playbook.
- Diagnose Before Blaming an Integration: When a SaaStr preview environment lost database connectivity, the agent blamed unrelated integrations; the team had to identify the actual failure rather than accept its explanation. — The Agents #001.
Part 3: Treating Agents as Teammates
- Onboard with Real Context: SaaStr treats a new AI SDR like a hire that needs proven messaging, examples, documents and correction before it works well. — SaaStr AI SDR Playbook.
- Review the First Outputs: SaaStr advises reading every message and response in the first 30 days, then shifting to spot checks once behavior has earned trust. — SaaStr AI SDR Playbook.
- Set Autonomy Boundaries: Amelia’s marketing-agent playbook lets 10K pull data and propose ideas on its own but requires approval before it emails the whole database. — SaaStr AI VP of Marketing Playbook.
- Give Specific Context: SaaStr found generic AI SDR inputs produced generic results; separate segments need different histories, proof points and messaging. — SaaStr AI SDR Playbook.
- Experiment with Agent-Centered Roles: SaaStr publicly described a marketing hire who would work with Amelia and 10K and, in its wording, report to the AI agent. This is one company’s experiment, not a general future org chart. — SaaStr Human–Agent Org Experiment.
- Check the Input Before the Model: SaaStr’s guidance is to inspect the segment, examples and business context when an AI SDR underperforms before assuming a vendor-model defect. — SaaStr AI SDR Playbook.
- Earn Trust in Stages: SaaStr first deployed agents where human coverage was weak, then moved toward more critical inbound interactions after learning how to train and check them. — SaaStr Agent Deployment Playbook.
- Give Each Agent a Clear Goal: Amelia’s 10K playbook recommends one goal and one agent brain per function, rather than asking a single agent to own marketing, customer success and events. — SaaStr AI VP of Marketing Playbook.
Part 4: Go-to-Market and Hyper-Personalization
- Personalize with Segments: SaaStr’s AI SDR guidance says different audiences require distinct context and proof points instead of one generic outbound campaign. — SaaStr AI SDR Playbook.
- Use Real Account Context: SaaStr’s outbound agent used company signals and prior relationship history to tailor outreach; the useful work is choosing the right segment and context. — SaaStr Six-Month AI SDR Report.
- Use More Than an Inbox: SaaStr combined website interactions, CRM history and outbound tools so agents could respond using the prospect’s known context. — SaaStr Six-Month AI SDR Report.
- Fix the Data First: SaaStr says weak CRM and RevOps data limit what agents can do; it cleaned up processes and loaded specific context before scaling AI SDRs. — SaaStr Six-Month AI SDR Report.
- Test One Variable at a Time: SaaStr reports testing subject lines and calls to action while changing one variable daily during agent-training optimization. — SaaStr AI SDR Training Report.
- Measure Outcomes, Not Just Sends: SaaStr reported response rates, positive replies, meetings and revenue alongside message counts rather than judging its AI SDRs only by volume. — SaaStr Six-Month AI SDR Report.
Part 5: The "No Lead Left Behind" Philosophy
- Reach the Leads Humans Miss: The SaaStr team’s no-lead-left-behind approach used agents to contact prospects and prior leads that a small human team did not have time to work. — The Agents #001.
- Put Agents on B-Leads: SaaStr reported assigning Artisan to promising B-leads that human sellers had not worked, while retaining human attention for the largest opportunities. — The Agents #006.
- Respond While Interest Is Live: SaaStr’s inbound agent engaged visitors in real time rather than leaving them at a form awaiting a later human follow-up; the source does not verify a universal three-second response. — SaaStr Six-Month AI SDR Report.
- Cover Off-Hours Demand: SaaStr’s inbound deployment handled prospects overnight across time zones when its Pacific-based humans were unavailable. — SaaStr Agent Deployment Playbook.
- Use Cadenced Follow-Up: SaaStr deployed separate agents for active nurture and ghosted-lead recovery, putting follow-up where its human team had left gaps; it does not advocate limitless messaging. — SaaStr Six-Month AI SDR Report.
- Qualify Before the Handoff: For higher-value sponsorship deals, SaaStr’s agents qualify and book meetings before handing the prospect to a human seller. — SaaStr Six-Month AI SDR Report.
- Revisit Prior Interest: SaaStr’s lapsed-customer and active-nurture agents re-engage contacts using previous relationship and engagement data. — SaaStr Six-Month AI SDR Report.
- Consistency Needs Training: SaaStr values agents’ consistent application of proven messaging, while stressing that good output depends on segmentation and continuing review. — SaaStr AI SDR Playbook.
Part 6: Navigating "Vibe-Coding" and Deployment
- Begin with a Small Vibe-Coded Tool: Amelia began 10K as a dashboard to avoid weekly copy-pasting and expanded it through repeated development into a marketing co-pilot. — SaaStr AI VP of Marketing Playbook.
- Write a Detailed Spec: Amelia’s 10K playbook starts with one measurable goal and a detailed spec; vague instructions produce less useful outputs. — SaaStr AI VP of Marketing Playbook.
- Build Workflows Sequentially: The 10K team added dashboard, ideas, campaigns and newsletter functions one at a time across months instead of launching a complete agent in one prompt. — SaaStr AI VP of Marketing Playbook.
- Guard the High-Impact Actions: Amelia’s playbook separates autonomous data work from outbound sends and checks AI-produced numbers against real data before messages go out. — SaaStr AI VP of Marketing Playbook.
- Keep Shared Operating Instructions: SaaStr uses a memory file with its agent goal, voice and send-domain rules, contacts and corrections; the source does not call a system prompt an operating system. — SaaStr AI VP of Marketing Playbook.
- Escalate Uncertain Outputs: SaaStr added numeric checks and manual approval before campaign sends after an agent invented a plausible contact count. — SaaStr AI VP of Marketing Playbook.
Part 7: Scaling from 1 to 20+ Agents
- Start with One Workflow: The SaaStr deployment playbook recommends a small first agent where the work is not already covered, then adding harder use cases after the team learns to train it. — SaaStr Agent Deployment Playbook.
- Track Agent Economics: SaaStr compared agent platform costs and staffing trade-offs against meetings and revenue; the evidence does not substantiate a specific token-cost rule. — The Agents #006.
- Do Not Skip Human Review: SaaStr’s fast agents increased the need to review high-blast-radius actions, including an irreversible 1,500-person email sent from a banned address. — The Agents #006.
- Report Results with Scope: SaaStr counted 2.25 million website sessions, 402,000 chats, and 614 qualified meetings for its inbound agent; these are company-reported activity figures, not a general ROI guarantee. — The Agents #006.
Part 8: AI Leadership and the Future of Operations
- Treat AI Leadership as Operations: Amelia’s SaaStr role centers on selecting, integrating, training and maintaining production agents tied to business workflows. — SaaStr Agent Deployment Playbook.
- Value Operators Who Can Deploy: SaaStr describes Amelia as an operator who became able to build integrations and agent workflows through AI tools without a conventional developer background. — The Agents #002.
- Proprietary Context Matters: The SaaStr 10K account attributes its usefulness to historical company data, live integrations and an accumulating script library, not a proprietary base model. — SaaStr AI VP of Marketing Playbook.
- Pair Agents with Human Judgment: SaaStr gives agents repetitive execution and analysis while keeping humans responsible for training, approvals and relationship work; its own playbook does not claim humans disappear. — SaaStr AI VP of Marketing Playbook.