Booking Q3 2026 partner slots
·
US · UK · EU · AU
AI agents are autonomous systems that hold conversations, qualify leads, run audits, or supervise workflows on behalf of an agency’s clients. We design and build five categories of agent — chatbots, voice agents, audit agents, lead-qualification bots, and workflow agents — under your agency’s brand. Every build ships EU AI Act compliant by default.
An agent makes decisions through reasoning. A chatbot follows a script. An automation follows a rule.
The distinction matters because the three categories have different unit economics, different failure modes, and different deployment constraints. A scripted chatbot from 2020 costs almost nothing to run but breaks the moment a customer asks something outside the script. A rule-based automation runs forever for cents per month but can’t handle ambiguity. An AI agent costs more per interaction but handles ambiguity, makes judgment calls in real-time, and can be trained against your client’s actual content rather than a brittle decision tree.
When we say “AI agent,” we mean a system that reads context, decides what to do next, and acts — usually through a combination of a language model, a retrieval layer pulling from your client’s data, and a small set of tools the agent is allowed to call. The agent is autonomous within its scope and explicitly blocked outside it. That scope-and-guardrails design is the difference between an agent that works in production and one that becomes a liability.
01 / Agent type
Conversational AI deployed on a website or inside a product. Trained on your client’s content library — podcast episodes, blog archive, courseware, support documentation — with recency filters that rank current information above older sources for time-sensitive topics.
02 / Agent type
Voice AI accessible via phone or web microphone. Handles intake calls, scheduling, after-hours support, qualification workflows. Identifies itself as AI on every call, logs every conversation, and escalates to a human when the scope is exceeded.
03 / Agent type
Scope-bounded agents that watch a system, read each new artifact, cross-check claims against current sources, and produce structured findings. Never auto-publish. Never rewrite content. Capability without overstep.
04 / Agent type
Agents that triage inbound leads — reading the website, evaluating fit against your client’s ICP, scoring priority, and routing high-fit leads to the sales team with research already attached.
05 / Agent type
Agents that supervise multi-step processes — running through a workflow, making decisions at each branch, calling tools when needed, and reporting completion or escalating when stuck.
A voice AI replacing an after-hours answering service at $130–$430/month. Platform fee runs at a fraction of the human service with richer call data. Fast ROI.
A public chatbot on a marketing consultancy’s site, trained on 250+ podcast episodes. Conservative estimate from the live build: 5–10 incremental qualified leads/month at meaningful seven-figure annual impact.
An AI lead enrichment agent researching every new lead in 10 seconds. Replaces 5–10 minutes of manual research per lead. Multi-hour weekly recovery that compounds across the year.
A scope-bounded audit agent that catches outdated stats in training materials before the trainer presents them. Hard to monetize until you’ve watched a credibility hit happen.
A lead-qualification bot built once for one client can be deployed against ten others. The work pattern shifts from “build once, bill once” to “build once, bill many times.” This bridges into AI Products & MVPs.
Your client sees only your agency. Every part of the deliverable is yours: the domain the agent lives on, the branding in the conversation, the documentation your client receives, the phone number the voice agent answers, the email address that surfaces from any human escalation. We never contact your client directly. We never appear in their systems.
For the team operating the agent post-launch, we provide complete runbooks under your brand: how to update training data, how to adjust scope, how to handle escalations, how to read the logs, what to do when the agent misbehaves. Most of our active partners introduce us internally as “our AI team” if they introduce us at all.
If your client asks who built it, the answer is your agency. That’s the actual point of the arrangement.
Build 01 · Voice agent
EU AI Act-compliant voice agent handling intake calls. Self-identifies as AI. Explicit scope guardrails. Logged conversations. Production voice clone of the firm’s principal. Shipped in 3 weeks under the partner agency’s brand.
Build 02 · Chatbot
250+ podcast episodes indexed with recency filtering. Citation links on every answer. Lead capture activates inside the conversation when intent signals fire. Generating an estimated $100K+/year in incremental qualified pipeline.
Build 03 · AI receptionist
Voice agent integrated with HaloPSA (caller validation, ticket creation), AlertOps (on-call lookup, warm transfer), and Twilio (telephony). Validates callers against the client database. Classifies emergency calls. Creates tickets for every call.
Build 04 · Workflow agent
Wraps a daily 6-minute podcast cadence (365+ episodes/year). Handles research aggregation, script drafting, transcription, SEO title generation, show notes, editor cues, and WordPress posting. Saves 50–70% of cumulative prep and admin time.
Every agent build in this category is delivered through our standard engagement: $1,500/month, including Discovery (month one), build capacity, deployment under your brand, source code handoff, and monthly strategy reviews.
| Build complexity | Time to production |
|---|---|
| Single-purpose chatbot on a content corpus | 2–3 weeks |
| Voice agent with telephony + 2 system integrations | 3–5 weeks |
| Audit agent with structured output to Sheets/Notion | 1–2 weeks |
| Lead-qualification bot with CRM integration | 2–4 weeks |
| Multi-step workflow agent (5+ tools) | 4–8 weeks |
Most engagements ship one agent in the first 4–6 weeks of the partnership, then continue building in the same category or expand to Automations or AI-Assisted Deliveries as the relationship matures.
Not every problem deserves an agent. We’d recommend against an AI agent in these situations:
01 · Rule-based problem
If the desired behaviour can be described in 5–10 deterministic if-then rules, an automation is faster to build, cheaper to run, and more reliable than an agent. Use agents for ambiguity, not for control flow.
02 · Catastrophic wrongness
Agents can be wrong. Their wrongness is usually subtle and intermittent. For workflows where a single error has irreversible consequences — financial transactions, medical advice, legal filings — the right answer is a human in the loop, not an agent acting autonomously.
03 · Low volume
An agent that runs 10 times a month rarely justifies the build investment. Agents make sense when the interaction volume is high enough that the per-interaction marginal cost approaching zero matters. Below ~100 interactions per month, automation or template-driven content usually wins on ROI.
If you’re not sure whether your use case fits, book a scoping call. Discovery exists specifically to surface which category fits which need.
Platforms ship a working chatbot in hours. Custom agents ship in weeks. The tradeoff: platforms train you and your client to fit their model. Custom agents fit your specific content corpus, your client’s exact compliance posture, and your agency’s branding without any platform tax. For high-volume, high-stakes deployments — where the chatbot represents your client’s brand to thousands of customers — custom usually wins on conversion quality and trust. For low-stakes internal use, platforms are fine.
Yes. We work with ElevenLabs’ production voice clone (PVC) feature to build voice agents that sound like a specific person — usually the firm’s principal, the consultancy’s founder, or a thought leader the brand is built around. Consent and rights are sorted before the build. The compliance disclosure (“I’m an AI version of [name]”) is non-negotiable.
Through a combination of retrieval-augmented generation (the agent searches the content at query time) and fine-tuning (for specific use cases where the agent needs to internalize tone or domain knowledge). Most builds use retrieval — it’s faster, cheaper, and updates automatically when the content library changes. Fine-tuning is reserved for cases where retrieval isn’t enough.
Three layers of defence: scope guardrails that block the agent from answering out-of-scope questions in the first place, citation requirements that show users where the answer came from (so humans can verify), and explicit “I don’t know — let me connect you to a human” handling for edge cases. Every wrong answer in production becomes a training case in the next iteration. We don’t ship agents that never make mistakes. We ship agents that make mistakes recoverable.
For agents replacing a clear existing line item (an answering service, a manual research workflow), payback is typically 1–3 months. For agents generating incremental pipeline (a content chatbot, a lead-qualification bot), payback is typically 3–9 months. For agents protecting credibility or unlocking new resale, payback is harder to measure but the agent justifies itself once the first real-world failure is averted or the first productized resale closes.
Yes. Most frontier LLMs handle major European languages, plus Hindi and Mandarin, with quality close to English. Voice agents work well in English, Spanish, French, German, Portuguese, and Italian. For other languages we’d test on a small scope first before scaling.
Yes, with the right architecture from the start. Multi-tenant deployment is straightforward — one codebase, many client-specific content corpora and configurations. This is where AI agents bridge into AI Products & MVPs. Most agencies don’t think about multi-tenant until build 2 or 3; by then it’s a re-architecture. We design for multi-tenant from build 1 when there’s a credible resale path.
It depends on the specific use case. We’ve shipped agents in legal services and managed services for regulated clients. The rule we follow: agents can intake, qualify, triage, and inform — they cannot give regulated advice. A voice agent can intake a legal inquiry and route it to an attorney. It cannot give the legal advice. The boundary is explicit in the system prompt and the scope guardrails. For most regulated deployments, this boundary is workable. For some (e.g., autonomous medical diagnosis), it isn’t, and the answer is don’t deploy.
A 30-minute scoping call to talk through your specific use case, your client base, and whether an AI agent is the right answer. We’ll be honest if it isn’t.
Book a 30-minute scoping callA clear sense of whether an agent fits the use case.
Rough scope and timeline if it does.
A pointer to a different service category if it doesn’t.
