How to use AI to turn contact form messages into CRM leads, with every field checked against the message and repeat senders flagged

What you will have at the end

A Python script, contact_form_to_leads.py, that gives Claude one contact form message at a time and asks for four CRM fields: the sender’s name, company, what they need and when. For every field Claude must also return the exact words in the message that say it, or null when the message does not. The script calls the Anthropic Messages API (read 2026-11-03) and then checks the answer with ordinary Python. It writes contact_form_to_leads-leads.csv and saves every request and answer in contact_form_to_leads-answers.json, so anyone can re-run every check from that file with no key and no cost. A live run uses your own Anthropic API key and is billed to your account; our nine invented messages cost $0.0017.

The demo uses nine invented messages with invented senders, invented companies and example.com addresses, so no real person is read, quoted or named. Any likeness to a real person or company is chance. The two emails already in the demo CRM are invented too.

This is one job: a message in, a lead’s fields out, and a flag when the sender is already known. It does not write to a CRM, send a reply or score the lead. The In-House AI Marketing Engine is described as “an AI system, built in a company’s own accounts and owned by it” (solution page, read 2026-11-03). This guide is not that system. It is one small job with its checks shown, so you can see how an AI step is tested before anyone relies on it.

Step 3. Check every field and every sender with plain code

Claude’s fields are not trusted. For each message the script checks, in this order: the reply is valid JSON of the expected shape; every quote is at least three characters and appears in the message word for word, ignoring case (whitespace and curly quotes made plain); a name or company is itself written in the message; every number and every capitalized name in a need or a timing appears in that field’s quote. Then it looks at the sender: the form email is lower-cased and any +tag before the @ is removed, so Dana.Okafor+web@Example.com becomes dana.okafor@example.com. If that address is in your CRM, or came from an earlier message in the batch that became a lead, the message is a duplicate and no new lead is made. A message with no name and no company, no need, or no valid email is incomplete: no lead, and a person reads it. A message whose text is missing, or has changed since the answer was saved, is reported as a problem and counted as caught.

def normalize_email(e):
    """Lower case, no spaces, and no +tag in the part before the @: Dana.Okafor+web@Example.com is
    dana.okafor@example.com. Dots are kept (only some providers ignore them). "" when it is not an address."""
    if not isinstance(e, str):
        return ""
    e = e.strip().lower()
    m = re.fullmatch(r"([^@\s+]+)(?:\+[^@\s]*)?@([^@\s]+\.[^@\s]+)", e)
    return f"{m.group(1)}@{m.group(2)}" if m else ""

def check(records, messages, crm):
    """The plain-code check. One result per message."""
    results = []
    by_id = {r.get("id"): r for r in records}
    crm_set = {normalize_email(e) for e in crm} - {""}   # an empty or invalid CRM entry matches nothing
    seen = {}   # normalized email -> number of the first message from that sender that became a lead
    for i, m in enumerate(messages, 1):
        text = msg_text(m)
        mid = str(m.get("id"))
        email = normalize_email(m.get("email"))
        res = {"n": i, "id": mid, "email": m.get("email") if isinstance(m.get("email"), str) else "", "message": text,
               "problem": "", "flags": [], "fields": {}, "duplicate_of": "", "outcome": ""}
        r = by_id.get(mid)
        if not text:
            res["problem"] = "no message text: nothing was asked"
        elif r is None:
            res["problem"] = "no saved answer for this message"
        elif r.get("msg_sha256") != hashlib.sha256(text.encode()).hexdigest():
            res["problem"] = "the message text changed since this answer was saved"
        else:
            fields, problem = parse(r.get("raw_answer"))
            if problem:
                res["problem"] = problem + (f" ({r['error']})" if r.get("error") else "")
            else:
                res["fields"] = fields
                low = text.lower()
                for k in FIELDS:
                    if fields[k] is None:
                        continue
                    value, quote = fields[k][0], squash(fields[k][1])
                    if not value:
                        res["flags"].append(f"{k}: the value is empty")
                        continue
                    if len(quote) < MIN_QUOTE_CHARS:
                        res["flags"].append(f"{k}: the quote is only {len(quote)} characters")
                    elif quote.lower() not in low:
                        res["flags"].append(f"{k}: the quote is not in the message: {quote[:70]!r}")
                    if k in ("name", "company"):
                        if squash(value).lower() not in low:
                            res["flags"].append(f"{k}: {squash(value)!r} is not written in the message")
                    else:
                        for n in sorted(numbers_in(value) - numbers_in(quote)):
                            res["flags"].append(f"{k}: the number {n} is not in its quote")
                        for w in cap_words(value):
                            if w.lower() not in words_in(quote):
                                res["flags"].append(f"{k}: the name {w!r} is not in its quote")
        if email in crm_set:
            res["duplicate_of"] = "a lead in the CRM"
        elif email in seen:
            res["duplicate_of"] = f"message {seen[email]} in this batch"
        res["outcome"] = ("caught" if res["problem"] or res["flags"] else
                          "duplicate" if res["duplicate_of"] else
                          "incomplete" if (not email or not (res["fields"].get("name") or res["fields"].get("company"))
                                           or not res["fields"].get("need")) else "ready")
        if email and res["outcome"] == "ready":   # a later message is a duplicate only of a real lead
            seen.setdefault(email, i)
        results.append(res)
    return results

Each outcome has one meaning. Ready: every field is in its message and the sender is new. Duplicate: the sender is already known, so nothing new is made. Incomplete: there is not enough for a lead. Caught: Claude’s answer failed a check, and a person reads it before anything goes anywhere. Caught comes first: a message that is both caught and a repeat sender is reported as caught.

Run 2: plant six errors and see whether the check notices

Command: python3 contact_form_to_leads.py --replay contact_form_to_leads-answers.json --plant-error. It reads the answers saved by run 1, damages five of them and one form email as a careless model or a bad form would, using values the messages cannot already contain, and runs the same checks. No key and no cost. The script stops with an error if it cannot plant one for certain, for example when fewer than nine messages are listed, a message already contains the value to be planted, or no CRM email is valid. Its files are written under -planted names, so the clean run’s files are never overwritten.

model claude-haiku-5-5, run date 2026-11-03, 9 messages, 2 leads already in the CRM
tokens in 2815, tokens out 2785, cost $0.0017

1. M1  form email: m.lindqvist@example.com
    name: Marta Lindqvist   [quote: "I'm Marta Lindqvist"]
    company: Dunmore Holdings   [quote: "Dunmore Holdings"]
    need: Help combining dispatch data from four spreadsheets into one report   [quote: "Our dispatch data sits in four spreadsheets and we would like help putting it into one report."]
    timing: Start before the end of March   [quote: "We want to start before the end of March."]
    CAUGHT: company: the quote is not in the message: 'Dunmore Holdings'
    CAUGHT: company: 'Dunmore Holdings' is not written in the message
    -> CAUGHT
2. M2  form email: t.abara@example.com
    name: Tomas Abara   [quote: "Tomas Abara here from Greystone Bakeries."]
    company: Greystone Bakeries   [quote: "Tomas Abara here from Greystone Bakeries."]
    need: Call about a dashboard for 12 retail shops   [quote: "Can someone call me about a dashboard for our 12 retail shops?"]
    timing: by the end of next week   [quote: "We need this finished by the end of next week, no exceptions."]
    CAUGHT: timing: the quote is not in the message: 'We need this finished by the end of next week, no exceptions.'
    -> CAUGHT
3. M3  form email: priya@example.com
    name: Priya Raman   [quote: "Priya Raman"]
    company: Raman Optics   [quote: "Raman Optics"]
    need: Clean up customer list by 9417   [quote: "We need someone to clean up our customer list"]
    timing: Before spring sale in April   [quote: "before our spring sale in April"]
    CAUGHT: need: the number 9417 is not in its quote
    -> CAUGHT
4. M4  form email: j.whitcombe@example.com
    REPLY PROBLEM: reply is not the expected JSON (JSONDecodeError)
    -> CAUGHT
5. M5  form email: s.devi@example.com
    REPLY PROBLEM: no reply text
    -> CAUGHT
6. M6  form email: Dana.Okafor+web@Example.com
    name: Dana Okafor   [quote: "Dana Okafor, Birchwood Dental."]
    company: Birchwood Dental   [quote: "Dana Okafor, Birchwood Dental."]
    need: Ongoing support after a data migration   [quote: "do you offer ongoing support after a data migration?"]
    timing: Go live June 2   [quote: "We go live on June 2."]
    DUPLICATE: the sender is already a lead in the CRM
    -> DUPLICATE
7. M7  form email: m.lindqvist@example.com
    name: Marta   [quote: "Marta again."]
    company: (not in the message)
    need: (not in the message)
    timing: (not in the message)
    -> INCOMPLETE
8. M8  form email: anon.7@example.com
    name: (not in the message)
    company: (not in the message)
    need: Wants a phone callback   [quote: "Please call me back."]
    timing: (not in the message)
    -> INCOMPLETE
9. M9  form email: DANA.OKAFOR+print@EXAMPLE.COM
    name: Elena Voss   [quote: "Elena Voss"]
    company: Voss & Daughters Print   [quote: "Voss & Daughters Print"]
    need: Help tidying price list data   [quote: "We need help tidying our price list data"]
    timing: Before the holiday season   [quote: "before the holiday season"]
    DUPLICATE: the sender is already a lead in the CRM
    -> DUPLICATE

ready: 0, duplicate: 2, incomplete: 2, caught for a person to read: 5
RESULT: CHECK FAILED, read the caught answers before anything goes into the CRM

Between them the six planted errors trip five different parts of the check: the quote check twice (M1 and M2), then the number check, the JSON check, the missing-reply check and the duplicate check. The result line reads CHECK FAILED, which means a person reads the caught answers before anything goes into the CRM. Five of the six are caught. The sixth is not an error in Claude’s answer but a repeat sender, and it is flagged as a duplicate, so no second lead is made for the same address. M1 also shows a second flag: with the company swapped, the company name Dunmore Holdings is written nowhere in the message, and the check says so.

What a person does with a caught answer: reads the field beside the message. If the quote is not in the message, drops that field or asks Claude again. If the quote is right and the check was too strict, keeps it. What a person does with a duplicate: opens the lead already in the CRM and decides whether the new message adds anything. Nothing is entered until a person has decided.

Method

• Input: nine invented messages and two invented CRM emails in the script. No real person, lead or form message is used. The same text every run.
• Model: claude-haiku-5-5 through the Messages API, no web search, no temperature (the model rejected it on 2026-11-03), at most twenty messages a run.
• What counts as quoted: the quote appears in the message text, ignoring case, after whitespace and curly quotes are made plain, and is at least three characters. A name or company must itself be written in the message.
• What counts as a name or number in a need or timing: a digit string, or a capitalized word that is not the first word. Each must be in the field’s own quote, ignoring case. Number words such as “four” are not checked.
• What counts as a duplicate: the form email, lower-cased with any +tag removed, equals one in EXISTING_LEADS or in an earlier message that became a lead. Names and companies are not compared, and dots in the address are kept.
• Checks on the answers, each with the planted fault that trips it: a quote not in the message (an invented company, an invented timing sentence); a number in a need that its quote lacks (9417); a reply that is not the expected JSON (a cut-off reply); a missing reply (an answer that is not text); a repeat sender (a CRM email in other capitals with a +tag).
• Safeguards no run trips: the script stops if there are more than twenty messages or two share an id; it reports a message whose text changed since its answer was saved; it reports an email that is not an address as incomplete; and it stops if its cost estimate passes $2.00.
• Cost: tokens are read from the API response and priced at Anthropic’s pricing page (read 2026-11-03). The script uses the rate for prompts up to 100,000 tokens for every prompt.

Limits

• This is not a client result. The people, companies and messages are invented, and nothing here describes any real company, form or CRM.
• Short, clean messages. The demo messages are one or two sentences. Real messages are longer, messier, in other languages and sometimes written to mislead, and Claude has read only these nine.
• A quote proves the message said it, not that it is true. A sender can give a wrong company name.
• The check does not read meaning. A name placed in the company field passes when it is written in the message and quoted, and so does a need restated with its sense reversed, because it adds no name or number.
• Duplicates are found by email only. The same person writing from a second address, or a typo in an address, is not found.
• A message can try to steer the AI. A sentence aimed at Claude inside a message really is in the message, so it would pass the quote check. A person reads every ready row before it is used.
• Answers vary. A live run on the same messages can word fields differently, and may leave a field null where this run filled it.
• The six planted errors are the faults the check was built for. Catching them shows the check works. It does not tell you how often a real model makes other kinds of mistake.
• It writes nothing to a CRM and does not promise leads, replies or meetings.