Your best leads arrive at 9 pm: an AI intake agent that answers, qualifies and routes them before morning

Monday, 8:40. Someone opens the shared inbox and finds the weekend in it. A quote request came through the website form on Friday at 21:12. Two emails went to an address that forwards to a sales manager on leave until Wednesday. A voicemail transcript sits unread. By the time anyone replies, one of those buyers may have booked a call with whoever answered on Saturday. Leads go cold because nobody owns the first touch outside office hours, and a shared inbox spreads ownership so thin it disappears.

The form emails one person. That person is on holiday. The phone goes to voicemail at six. Meanwhile the buyer filled in three forms the same evening, and the first company to answer frames the whole conversation.

The evidence is mostly vendor data. Digital Applied's benchmark page (updated 5 November 2026) is the most careful compilation we found: it states that none of its figures are first-party and lists caveats next to each.

• Blazeo surveyed 573 service businesses: 74% do not reply to a new lead within five minutes (23 September 2026), and 62.5% of those using automation met a 15-minute target against 39.1% of manual operators (24 February 2026). Self-reported, and Blazeo sells conversational AI.
• Entagl looked at 32,581 chat and DM conversations at 1,247 businesses (9 May 2026). Conversion was 35.1% with a first reply inside 60 seconds and 1.6% after a day. Chat only, observational; its own before/after estimate for businesses that switched on auto-reply is a smaller 1.94x lift.

The figure most blogs still quote comes from a 2011 Harvard Business Review audit. We leave it out. It is fifteen years old.

Six things, in this order, for every request:

1. Reads the message wherever it arrived: web form, email, chat, voicemail transcript.
2. Identifies the person and company and dedupes against the CRM, so a returning customer is not created as a new lead.
3. Classifies request type (new project, support, partnership, job application, spam) and urgency.
4. Asks one to three missing questions by reply, only when the answer changes routing: country, team size, deadline.
5. Scores the request against your criteria, written down by your sales lead.
6. Routes and records: creates or updates the CRM record and notifies the right human with a summary and a drafted reply.

The CRM record looks roughly like this (an illustration, not a client record):

{
  "channel": "web_form",
  "received_at": "2026-11-06T21:12:04+01:00",
  "contact": {"email": "ops@example.com", "company_domain": "example.com"},
  "crm_match": {"account_id": null, "possible_duplicate_of": null},
  "request_type": "new_project",
  "extracted": {"service": "ERP integration", "sites": 3, "deadline": "2027-02"},
  "missing": ["country"],
  "clarifying_question_sent": true,
  "score": {"value": 72, "rules_hit": ["size:50-200", "product:integration"]},
  "route_to": "sales_dach",
  "needs_human": true,
  "confidence": {"request_type": 0.91, "service": 0.78}
}

The confidence and rules_hit fields matter more than the score. They tell the salesperson why the agent decided what it did, and they are what you audit later.

Rules decide; the model reads. Routing by territory, company size and product line is a lookup table, so we write it as code, because a lookup table never has a creative day. The language model does what code cannot: turn a rambling email into fields. We make the same argument in workflow or agent, along with Gartner's forecast that over 40% of agentic AI projects will be cancelled by the end of 2027.

Our default auto-send line is narrow. The agent may send an acknowledgement ("we got it, a person will reply by 10:00 CET") and the clarifying questions from step 4. Everything else goes to a human queue with a draft attached:

• any price, discount or estimate;
• any commitment on dates or scope;
• legal text, contract terms, data processing questions;
• a rejection or "not a fit" reply.

That last one surprises people. A polite no still speaks for the company, and a model that misread the request will say no to the right prospect.

Extractions below a confidence threshold go to a human too. We start at 0.8 per field and move it after evals.

You test it on your own history before it touches a live lead, and you keep a record of every run. Before go-live the agent processes 200 to 500 past requests where you already know the answer: who should have got it, and whether it became a deal. If it routes your old requests worse than your team did, it does not ship.

In production every run leaves a trail: raw input, extracted fields, rules fired, message sent, who approved it. Application logs can be edited by whoever runs the database, which is fine for debugging and weak when a customer disputes what your company told them. For recorded runs we use SealedRun, an independent open-source tool that signs each step into a hash chain an outsider can verify. Each request also gets a cost cap, so one 40-message thread cannot burn a day's budget (cost control). And when the agent writes to the CRM through tool servers, those need vetting like any integration (MCP in production).

If most inbound is website chat, your CRM is HubSpot or Salesforce, and your qualification rules fit the vendor's settings screen, buy the vendor's agent. We would tell you that in discovery. HubSpot reports its Customer Agent resolves 65% of conversations across 8,000+ customers (13 April 2026); Intercom gives Fin a 76% average resolution rate across 12,000+ teams (12 March 2026). Both are vendor claims about support, not lead qualification. Intercom now also bills for "procedures with handoff", where Fin gathers context, writes to external systems and passes the case to a human. That is the intake pattern, sold per outcome.

Data readiness blocks every option below. In Salesforce's vendor survey of 3,075 service professionals (20 May 2026), 66% of service organisations use agentic AI and 72% of operations staff call data readiness a major blocker. An agent that routes by "industry" is only as good as the industry field in your CRM.

A custom build pays off in the cases in the right-hand column below.

Disclose that it is an AI, log it, offer a human. In the EU, Article 50 of the AI Act has applied since 2 August 2026: people must be told they are dealing with an AI system unless that is obvious from context. The Digital Omnibus moved the high-risk dates to December 2027 and August 2028 and left Article 50 alone (Gibson Dunn, 27 May 2026). If you rely on the "obvious" exception, write down why; Jones Walker (16 July 2026) advises keeping those facts in the compliance record. Fines are capped at EUR 15 million or 3% of worldwide turnover. Our AI Act article covers what to build for each date.

GDPR Article 22 protects people from decisions "based solely on automated processing" with legal or similarly significant effects, and where such decisions are allowed, requires a route to human intervention. A lead score rarely reaches that bar. An automatic rejection might, which is one more reason ours go to a human.

The US is a patchwork. California's AB 1609 applies from 1 January 2027 to businesses with more than $500 million in revenue: disclose the bot where a reasonable person could be misled, and make a good-faith effort to connect a customer who asks for a human within 15 minutes during business hours (Fox Rothschild, 30 September 2026). Utah's SB 226 (in force since 7 May 2025) requires disclosure when a consumer clearly asks, with a safe harbour for bots that disclose at the outset and throughout. Maine has its own bot disclosure law. Most 2026 state laws Orrick tracked target companion chatbots, which an intake agent is not.

Not legal advice. The rule we build to: the first message says it is an automated assistant, every message is logged, and "talk to a person" works in every channel.

This is our planning approach, not a promise. Discovery takes one week on your last 200 to 500 real requests: we classify them, write down the routing rules your team actually uses, and count how many could have been handled without a person. Then a pilot on one channel (usually the web form or the shared inbox), typically three to five weeks. Then shadow mode, where the agent drafts, your people send, and every edit is recorded; we stay there at least two weeks and until the override rate stops falling. Limited auto-send comes last.

We baseline three numbers before launch: median time to first response, share of requests routed to the right person, and the human override rate on drafts.

From you: sample requests (anonymised is fine), access to a CRM sandbox, the routing rules even if they live in someone's head, and one named owner who can say "that lead was ours".

Five questions you can probably answer today:

1. Where do requests arrive: which forms, inboxes, chat, phone?
2. Who replies now, and who covers evenings and holidays?
3. How long does a first reply take on a Tuesday, and on a Friday night?
4. What is a good lead, in one sentence your sales lead would sign?
5. Which CRM, and which of its fields does sales actually trust?

Send us your last 200 inbound requests (anonymised) and we will show you, in a one-week discovery, how many an intake agent could have answered, routed and qualified, and where a human must stay in the loop. We have spent 16 years and 500+ projects building in Python, JavaScript and PHP, with LLMs where they fit. Get in touch or see our services.