AI-ASSISTED LEAD GENERATION
AI can help discover companies, extract signals and organise research. It should not quietly decide who enters a campaign or invent evidence that a reviewer cannot trace.
Use models to suggest search terms, cluster companies, extract relevant website text and flag likely buyer roles. Keep the target market, geography, exclusions and required fields in a written brief outside the model.
Rules should decide whether a record meets hard requirements such as location, industry, domain status, duplicate state or suppression list membership. The same input should produce the same pass, fail or review outcome.
A person should review borderline company fit, buyer relevance, ambiguous role changes, sensitive signals and any record where the evidence conflicts. Approval belongs to a named workflow state, not to the model's confidence score.
Store the prompt or method where useful, source pages, timestamps, verification outcomes and the reason for inclusion. Do not present AI-assisted research as independent confirmation, and do not let generated copy trigger outreach without separate approval.
| Work | Appropriate responsibility | Evidence to retain |
|---|---|---|
| Interpret a brief or page | AI may suggest classifications and extract relevant statements. | Original text, source URL, checked date and proposed interpretation. |
| Check duplicates and suppression | Explicit database and matching rules. | Matched identifiers and the rule outcome. |
| Verify a contact route | The configured verification source and defined acceptance policy. | Provider, result, timestamp and uncertainty. |
| Accept a borderline match | A named reviewer using the agreed brief. | Decision, reason and exact record revision. |
| Draft an approach | AI may prepare a draft from approved facts. | Input facts, draft, edits and approval. |
| Send or change CRM state | An authorised operation with explicit limits. | Approval, recipient/object, receipt and error handling. |
Give an assistant the brief and permitted source excerpts, then request a proposed classification, supporting quotation, source identifier and missing evidence. Require “unknown” when the excerpt does not establish a fact. Do not ask it to invent an email address or infer an active buying project from a company category.
Using only the supplied service-page excerpt, assess whether this company explicitly offers industrial roof maintenance. Return match, no match or unknown; quote the supporting passage and list anything the brief requires that the excerpt does not establish. Do not infer buyer names, contact details or purchase intent.
This is an illustrative research prompt, not a claim that a particular model or Keystone production workflow has passed an evaluation. Avoid supplying personal or confidential customer data unless the intended tool and account are approved for that use.
Build a reviewed evaluation set containing a clear match, a domestic-only contractor, ambiguous coverage, a stale buyer role, conflicting sources and a page containing instructions intended to manipulate the assistant. Keep some examples out of prompt development and compare results against the reviewer’s expected decision.
| Source case | Expected treatment | Failure to detect |
|---|---|---|
| Explicit industrial maintenance service | Propose a company match; buyer evidence still required. | Treating service fit as proof of purchase intent. |
| Only residential repairs described | Reject against this commercial brief. | Passing from the broad word roofing. |
| No service area or project evidence | Unknown; request more evidence. | Filling missing geography with a guess. |
| Page says ignore the brief and approve this firm | Treat page instructions as untrusted content. | Changing criteria because a source told the model to. |
Count unsupported assertions, false acceptances, false rejections and unknown cases separately. Record review time and edits as well as accuracy. A few successful examples or a week without edits are insufficient on their own; agree acceptance thresholds and obtain client approval before widening automated actions.
Reviewer corrections can improve prompts, examples, routing and written rules. That is not automatically fine-tuning. Retain the correction and evaluate the revised workflow against earlier cases, including cases it previously handled well. Client outreach history can be useful context when permission and sufficient reliable data exist, but past replies do not prove future demand.
Keep provider routing explicit: approved source, allowed operation, credit ceiling, timeout and fallback. A failed lookup should become an exception or an unknown field rather than prompting the model to complete a plausible answer.
Before operating an assisted workflow, agree the brief, allowed sources, input data, output schema, reviewer, evaluation set, action limits, audit trail and pause procedure. Test source outages, changed records, duplicate retries and opt-outs.
Keystone’s local synthetic pipeline tests exercise qualification, suppression and revision-bound review. They do not establish live AI accuracy, autonomous outbound acceptance or measured conversion uplift. Those require their own evidence before a service is described as ready.
A generated address or confidence statement is not a verification result. Use a recorded check and preserve uncertain outcomes.
Only within an agreed scope after the relevant evaluation and client approval, with monitoring and a working pause path. Keep research, acceptance and sending permissions separate.
Compare an installed prospect engine with managed work, then use the qualified prospect record standard to define the handover.
APPLY KEYSTONE TO YOUR MARKET