🌟Vasilij’s note
This week something clicked for me that I keep explaining to clients: your website is not a brochure problem, it is a response-time problem. Thirty-nine percent of B2B conversations happen outside business hours. The average consultancy takes 42 hours to respond to an inbound enquiry. And the firms that respond within five minutes are 21 times more likely to qualify that lead. Meanwhile, ElevenLabs closed a $500M Series D at an $11B valuation, Anthropic shipped Managed Agents into public beta, and Meta unveiled its first model from its Superintelligence Lab. The tools are moving faster than most firms are thinking. The gap is not capability - it is configuration discipline.
In today's edition
This week in agents | What changed
Anthropic launches Claude Managed Agents into public beta.
A suite of composable APIs for building and deploying cloud-hosted AI agents at scale. Includes sandboxed code execution, checkpointing, scoped permissions, and persistent long-running sessions - designed for production-grade agent infrastructure without self-managed backends.
Meta's Superintelligence Lab releases Muse Spark, its first public model.
Proprietary model with multimodal reasoning, tool use, visual chain-of-thought, and multi-agent orchestration. Additional models in the Muse family will be open source. Performance comparable to leading frontier models in standard thinking mode, with acknowledged gaps in long-horizon agentic tasks.
Claude Mythos Preview autonomously identifies thousands of zero-day vulnerabilities.
Anthropic's Project Glasswing, built on Mythos, detected and fixed security vulnerabilities at scale across major operating systems and browsers in partnership with leading tech companies. The model was not trained specifically for cybersecurity — it is simply capable enough to do it.
Top moves | Signal → impact
Claude Managed Agents decouples the agent brain from the execution environment.
Anthropic's new composable API layer means developers no longer need to manage sandboxing, session persistence, or background execution themselves. Checkpointing and scoped permissions are built in. The system is described as designed for "programmes as yet unthought of" - built to accommodate future harnesses and sandboxes as models improve. → For consultancies building agentic workflows for clients, this changes the build vs buy calculation. Self-hosted agent infrastructure carries ongoing maintenance cost and security liability. Managed Agents shifts that burden to Anthropic's infrastructure. Evaluate whether your current agent builds would benefit from migration before they accumulate technical debt.Perplexity's shift from search to agents drove a 50% revenue increase.
Perplexity moved away from AI-powered search as its primary product and repositioned around AI agents. The revenue impact was reported at 50% growth. The company that most visibly built its brand on "better search" has concluded agents are the more valuable product. → This is not a product story - it is a positioning signal. If a search-native company concluded that agentic capability outperforms retrieval as a value proposition, the implication for knowledge-work consultancies is direct: the firms that reframe around what agents do for clients - not what AI can answer - will close faster.Enterprise AI adoption moves from experimentation into core workflows.
OpenAI reported rapid enterprise growth with AI now embedded in core business processes rather than sitting in pilot environments. Simultaneously, Ramp reported that 99.5% of its team is active on AI tools - achieved through deliberate internal adoption discipline, not mandate. → The adoption gap between early-moving firms and the rest is compounding. Consultancies that have not yet moved at least one workflow into production-grade AI are no longer in the "thoughtful early majority" - they are falling into the late majority. The measurement standard is shifting too: adoption metrics alone are being called vanity metrics; the bar is now ROI evidence.
Upskilling spotlight | Learn this week
ElevenLabs Conversational AI Documentation - Quickstart Guide
Step-by-step setup for voice agents: agent creation, knowledge base upload, voice selection, and website embed. Covers webhook integrations including Cal.com for calendar booking and Zapier for workflow triggers. Non-enterprise accounts are capped at 20MB or 300,000 characters for the knowledge base - more than sufficient for most consultancy service documentation.
Anthropic's Managed Agents Engineering Blog - Scaling Managed Agents
Technical breakdown of how Anthropic decoupled the agent "brain" from the "hands" in its Managed Agents architecture. Relevant reading for anyone advising clients on agent infrastructure decisions or evaluating whether to self-host versus use managed infrastructure. Thirteen-minute read; actionable for senior ops and technical leads.
Optifai Pipeline Study 2026 - Lead Response Time Benchmarks
Data from 939 B2B companies (Q1–Q3 2025) on how response time affects qualification and close rates. Key findings: only 23% of firms respond within five minutes; 42% take longer than 24 hours; leads contacted within five minutes close at 32% versus 12% after 24 hours. The channel breakdown is particularly useful - contact forms average 28 hours to first response, live chat averages eight. If you are making the internal case for a voice agent on your website, this is the research that builds the business case.
Maker note | What I built this week
This week I recorded a complete walkthrough of how I configured an ElevenLabs voice AI agent on the AiGentic Lab website - one that answers questions, qualifies leads, and books meetings around the clock without anyone in the office needing to be present.
Decision:
The gap at most consultancies is not capability - it is the contact form. A form that converts at 1% and gets checked the next morning is not a sales tool. It is a polite way of losing leads. The video below covers the full build: system prompt, knowledge base setup, voice selection, calendar integration via Cal.com, and a live stress test.
Operator’s picks | Tools to try
ElevenLabs Conversational AI
Use for voice agent deployment on consultancy websites. Embed in two code snippets (script tag and div tag); works on WordPress, Webflow, Framer, and Wix.
Standout: 5,000+ voice options including custom voice clones; 70+ languages with automatic detection. Pricing from $0.10/min, drops to $0.08 on annual Business plans.
Caveat: knowledge base quality determines output quality - thin content produces vague responses.
Cal.com
Use for calendar booking integration inside voice or text agent workflows. Connects via webhook to ElevenLabs agents, allowing visitors to book discovery calls during a conversation without human intervention.
Standout: open-source, self-hostable for firms with data residency requirements.
Pair with: ElevenLabs Conversational AI for end-to-end lead capture and qualification.
Poke
Use for lightweight automation accessible via text messaging apps. Pre-made "recipes" for scheduling, smart home control, and task management. Backed by $25M in funding; targets broad adoption without the complexity of full agent frameworks.
Caveat: designed for simplicity over customisation - not suited to complex multi-step professional services workflows, but a useful entry point for partners who want automation without configuration overhead.
Deep dive | The Consultancy Website Problem - and the Voice AI Fix
Most consultancy websites are architecturally identical to what they were in 2015. A contact form, a phone number, and an email address. The visitor experience has not changed. The AI available to replace it has changed dramatically. Thirty-nine percent of B2B conversations happen outside business hours. Contact forms convert at approximately 1%. Firms that respond to enquiries within five minutes are 21 times more likely to qualify the lead. The maths against the status quo is not close.
On paper
ElevenLabs Conversational AI launched in November 2024, rebranded to ElevenAgents in September 2025, and closed a $500M Series D at an $11B valuation in February 2026. The product lets you build a voice agent that answers questions, qualifies visitors, and books meetings - embedded on any website via two lines of code.
Pricing starts at $0.10 per minute, dropping to $0.08 on annual Business plans. For a consultancy receiving twenty conversations per day at three minutes average - roughly 1,800 minutes per month - the monthly cost sits at approximately £130–260 depending on plan. A single Google Ads lead in business services costs an average of £78 ($103.54 per WordStream's 2025 benchmarks). One qualified lead covers the monthly tool cost.
The knowledge base uses built-in RAG: upload service pages, case studies, FAQs, or a URL. The agent answers from your content rather than fabricating responses. Non-enterprise accounts cap at 20MB or 300,000 characters - sufficient for the content footprint of most 15–250 person consultancies.
In practice
The setup takes 30 minutes for a first installation. System prompt quality determines conversation quality - a vague prompt produces a vague agent. Voice selection matters: ElevenLabs offers 5,000+ options including custom voice clones. Calendar integration via Cal.com requires a webhook configuration that adds another 15–20 minutes but unlocks end-to-end booking without human involvement.
The agent handles four scenarios well: service enquiries, lead qualification, call booking, and graceful decline when a question falls outside its knowledge base. It handles adversarial questions - data concerns, scope challenges - with appropriate responses if the system prompt defines boundaries clearly.
After-hours capture is the clearest ROI case. Enquiries at 10pm, 6am, and weekends that would previously have hit an empty inbox are now qualified and booked before the partner opens their laptop the next morning.
Issues / backlash
Voice quality is excellent. Conversation intelligence varies by underlying language model - ElevenLabs lets you choose the model, better models cost more per minute. Knowledge base gaps produce either generic responses or hallucinations; the discipline of writing clear service documentation is the real upfront investment.
GDPR and data residency: conversations are processed through ElevenLabs servers. For most consultancies this is acceptable. For regulated industries - financial services, legal, healthcare - check the data processing agreements carefully. HIPAA compliance is available on higher tiers.
Cost scales with volume. Twenty conversations per day is manageable. Two hundred per day changes the monthly calculation materially. Model the numbers for your specific traffic before committing.
My take (what to do)
Startup (15–40 staff): Install and run it on your website this week. The setup cost is £0 to test. Document what the agent gets wrong in the first five conversations — that list is your knowledge base improvement brief. Measure whether post-hours enquiries increase within 30 days. If you get one qualified lead from it in the first month, it has paid for itself.
SMB (50–120 staff): Assign one ops team member as the agent owner — not an additional role, a 2-hour weekly responsibility. The owner maintains the knowledge base, reviews conversation logs monthly, and refines the system prompt based on what visitors are actually asking. Treat it like any other client-facing asset: it degrades without maintenance.
Enterprise (150–250 staff): Run a structured review of your current inbound process before deploying. Map where leads are lost: after-hours visits, slow follow-up, unqualified calls. Deploy the agent against the highest-loss stage first. If data residency is a concern, confirm ElevenLabs' DPA before going live. Pair the agent with HubSpot and n8n so qualified leads route automatically into your CRM without manual data entry.
How to try (15-minute path)
Go to elevenlabs.io, create a free account, and navigate to ElevenAgents → New Agent → Blank Agent. Write a system prompt that describes your firm's services, the three questions to ask a qualifying visitor (firm size, current challenge, timeline), and what the agent should do when it cannot answer - suggest booking a call. (5 min)
Upload your services page URL to the knowledge base. Select a voice. Copy the two embed snippets into your website's HTML. Open the widget and ask it one real question a prospect would ask. (5 min)
Write down what it got wrong or what it missed. That list is your refinement brief - update the knowledge base and system prompt accordingly. (5 min)
Success metric: an agent that correctly describes your services, asks at least two qualifying questions, and either books a call or provides a relevant next step - all before a human in your team knows the visitor existed.
"Managed Agents allowed us to build the integration in weeks and removed the operational overhead of maintaining agent infrastructure."
Spotlight tool | ElevenLabs Conversational AI
Purpose
Voice AI agent for consultancy websites that captures and qualifies leads outside business hours - without a human in the loop.
Edge
Custom voice cloning means the agent can sound like the founder or a named partner; 70+ language detection handles international visitors automatically; built-in RAG grounds every response in your firm's actual content rather than fabricated answers.
→ Two-snippet embed for any website platform
→ Cal.com and Zapier webhook integrations for booking and workflow triggers
→ 5,000+ voices including custom clones
→ Pricing from $0.10/min with annual plan discounts
→ HIPAA compliance available on enterprise tiers
What did you think of today's issue?
Did you find it useful? Or have questions? Please drop me a note., I respond to all emails. Simply reply to the newsletter or email [email protected].
This issue’s sponsor
n8n
An open‑source automation platform that lets you chain tools like DeepSeek, OpenAI, Gemini and your existing SaaS into real business workflows without paying per step. Ideal as the backbone for your first serious AI automations.
Refer and win
Share this newsletter for a chance to win!