// FAQ

What we build, who it's for, and how engagements run.

If your question isn't here, write to us — we answer everything.

What does Craft Digital actually build?

Operational infrastructure — the systems behind revenue and delivery. CRMs that reflect how the business really sells, intake and qualification automations, AI agents wired into existing tools, dashboards that operating partners trust, and the integrations that hold it all together.

Who do you work with?

PE-backed and PE-curious mid-market businesses, $10M to $200M in revenue, where strategy is set and execution is the bottleneck. We work directly with operating partners and CEOs.

How do engagements typically start?

Most start with the Value Creation Assessment — a three-week diagnostic that produces a custom report and sequenced infrastructure plan. From there, we usually move into a phased buildout.

How long does a typical buildout take?

Twelve weeks across three phases. Phase 01 lays foundation and integrations. Phase 02 stands up demand capture and sales execution. Phase 03 delivers the operational backbone, dashboards, and handoff.

What does the tech stack usually include?

Owned wherever it matters: HubSpot or Salesforce as system of record, Make and n8n for orchestration, OpenAI and Anthropic for reasoning, Twilio and ElevenLabs for voice, Retool for internal tooling, Postgres and Supabase for data. We build to your stack, not ours.

Do you build AI agents?

Yes — voice agents for inbound qualification and outbound recall, support agents wired to your knowledge base, and operations agents that read your data and write back to your systems. Always with humans in the loop where it matters.

Do you do automation?

Constantly. Lead routing, speed-to-lead, contract handoffs, billing reconciliation, customer onboarding, reporting roll-ups. If a process happens more than once a week and a human touches it without thinking, it's a candidate.

Do you do data and reporting?

Yes. We unify data across the stack and build executive dashboards that match how the operating partner reads the business — not generic templates.

How are you different from a consultancy?

Consultancies write decks. We build systems. The deliverable on day one of a buildout is working software, not a slide.

How are you different from an agency?

Agencies execute campaigns. We architect the operating system underneath them.

Do you work on retainer?

After buildout, yes — most clients keep us on a managed-infrastructure basis to evolve the systems as the business changes.

Where are you based?

Fort Collins, Colorado and Los Angeles, California. We work with portfolio companies across North America.

What is an AI voice agent, in plain terms?

A software caller that answers and places phone calls using a real-sounding voice, follows a defined playbook, looks things up in your CRM, and writes the result back to your systems. It is not a phone tree, and it is not a chatbot bolted to a phone number — it is a structured conversation with tools attached.

Which AI voice agent is best for a $10M–$200M business?

There is no single answer, because the right platform depends on call volume, latency tolerance, your CRM, and how custom the playbook needs to be. As a buyer's framework: Synthflow is fastest to launch for simple inbound, Vapi and Bland win on latency for outbound, Retell sits in between, and a custom build on top of OpenAI Realtime or Anthropic plus Twilio wins when the workflow is unusual or the integrations are deep. We help operators choose by call type, not by brand. In our own practice we build primarily on Retell, because it lets us architect multi-node prompt graphs — separate reasoning nodes for qualification, objection handling, scheduling, and escalation — instead of one monolithic system prompt. That is where we have found the most success: multiple voice agents in production today, processing thousands of calls per month. We will still build on Vapi, Bland, Synthflow, or fully custom when the call type genuinely demands it.

What platform do you actually build voice agents in?

Retell, by default. We are platform-agnostic by capability — we have built on Vapi, Bland, Synthflow, and fully custom on OpenAI Realtime plus Twilio — but Retell is where the bulk of our production volume lives. The reason is technical: Retell supports multi-node prompt architecture, where each step of the call (greeting, qualification, objection handling, scheduling, escalation) is its own reasoning node with its own evaluation surface. That makes the agents easier to debug, easier to improve in isolation, and easier to operate at scale. We have multiple agents running in production today processing thousands of calls per month on this stack.

Are AI voice agents better than human agents?

On three dimensions, yes — speed to answer, consistency of script adherence, and 24/7 availability. On empathy, judgement, and handling unscripted situations, no. The right design uses the AI agent for the first sixty seconds (qualify, route, book) and a human for everything else.

How much does an AI receptionist cost in 2026?

Per-minute platforms run roughly $0.07–$0.25 per minute of conversation; flat-fee answering services start around $300–$1,200 per month for low volume. A custom-built agent typically costs $20K–$60K to build and a few hundred dollars per month to run on infrastructure. The honest comparison is total cost of ownership at your call volume — most mid-market operators find that flat-fee plans break above 1,500 minutes per month and per-minute platforms break above ownership of the playbook.

Can an AI receptionist integrate with our CRM?

Yes. Any serious deployment writes to the CRM at the end of every call: contact created or updated, call summary, intent tags, and a booked appointment if applicable. We integrate with HubSpot, Salesforce, GoHighLevel, ServiceTitan, Jobber, and most modern CRMs over native APIs.

What is speed to lead?

Speed to lead is the elapsed time between a prospect raising their hand — submitting a form, calling, texting — and the first meaningful response from your team. The widely cited Lead Connect study found that conversion rates drop ~80% when first response moves from five minutes to thirty.

How fast should we respond to an inbound lead?

Under five minutes for high-intent inbound (demo requests, pricing pages, quote forms). Under one minute is achievable with AI-assisted response and where appropriate, materially raises booked-meeting rates.

Does AI hurt close rates if it talks to leads first?

Only if it tries to sell. A well-designed first-touch agent qualifies, books, and hands off. Close rates typically rise because reps stop talking to unqualified leads and start their conversations after the prospect is already booked.

How do CRM integrations affect speed to lead?

They are the bottleneck nobody talks about. Most slow responses are not a willingness problem — they are a routing problem. Leads land in a form, sit in a queue, get assigned, get notified, and only then get called. We rebuild the routing layer so the first response happens in the same workflow as the lead capture.

Is hiring an AI automation agency worth it?

Worth it when three things are true: there is a measurable, repeated process that costs real money or time; you have a system of record clean enough to automate against; and you have an internal owner who will run the system after handoff. Skip it when the underlying process is broken — automating a broken process just breaks it faster.

How do we measure ROI on AI workflow automation?

Pick one of three units before you build: dollars per task removed, hours per week returned, or revenue per lead recovered. Baseline the current number, run for ninety days, and measure against that baseline. Anything else is a vibe.

Where do most AI automation projects fail?

Three places. Bad data — the system of record is messy, so the automation propagates the mess. No owner — nobody is responsible after handoff. Over-scoped phase one — teams try to automate ten workflows at once instead of getting one working end-to-end and earning the right to do the next.

How secure is AI in customer support and back-office workflows?

Secure by design when it is built right: SOC 2-eligible vendors, no customer data in training pipelines, scoped API keys, audit logs on every write, and PII redaction at the prompt boundary. Insecure when teams paste credentials into ChatGPT and call it automation. The technology is not the risk — the implementation is.

What is managed AI infrastructure?

It is the operating model where someone else owns the systems you depend on — uptime, model upgrades, prompt drift, integration breakage, vendor changes — so your team gets the output without staffing the engineering. It is the difference between buying a car and buying transportation.