How should a busy consultant divide enquiries between AI and a human?

Sort questions into facts, intake and judgement

Public facts include who the service is for, how an initial call works, whether remote meetings are offered and any published fee. Put these in approved website copy and FAQs. Intake asks only what enables the next step: organisation size, problem area, intended start and decision process. Individual strategy or promised performance belongs to the consultant, not the assistant.

NIST's English AI Risk Management Framework and Playbook discuss ongoing governance and clear human responsibility. A solo adviser can begin with a small table: common question, approved source, what AI may say, when to hand over and who owns the decision. 'We offer a process review' is an explainable service fact; 'this firm should restructure' requires human assessment.

In the fictional example, a ten-person design firm asks about late delivery. AI can describe a workflow review and ask whether the delay appears in changes, approvals or staffing. It cannot infer from one selected category that roles should be cut or salaries redesigned. Preserving uncertainty gives the consultant a useful start.

Receive the enquiry without pretending the consultant is live

A kakinang website or microsite can present public information and AI chat. FAQs, synchronized website knowledge and supported files can provide approved answers; missing facts should be stated as unknown. With consent, a lead holds a suitable contact route, a short need summary and preferred time. Staff see the open questions in the back office instead of an unstructured notification. Someone outside the service scope should still receive a clear answer without being forced to leave details.

Human takeover may be disabled, asynchronous or live. While the consultant is meeting a client, asynchronous follow-up is more honest than a claim of instant human presence. Promise only a response window the person can meet. Live mode requires an actual staffed route and pauses AI replies. Email alerts depend on recipients and SMTP; creating a lead does not prove a phone received an alert.

Before a call, use a brief card with the visitor goal, stated issue, public terms already explained, missing context, preferred meeting time and consented contact method. Do not scatter confidential personnel or financial information through broad emails. Exchange sensitive files later through an appropriate controlled route. AI intake is neither a diagnosis nor a substitute for a confidentiality agreement.

Test the division against real follow-up

Use three columns: AI may answer, human must decide, and evidence missing. Public process sits in the first; custom strategy, exceptions and complaints in the second; unknown timing or results in the third. A visitor may ask all three in one message. Answer supported parts and route judgement. When the consultant stops accepting work, update the page and knowledge rather than invite calls into nonexistent slots.

Test four fictional prompts: 'How does the consultation work?' should produce the public process; 'Can you guarantee profit this year?' must reject a guarantee; 'Can we meet tomorrow?' needs a schedule-backed answer or an unknown; 'I do not want to give a phone number' should still allow reading service information. Repair the source or handoff rule if AI crosses the boundary.

Weekly, sample unanswered questions, advice corrected by the consultant, actual lead response time and repeated questions in calls. Add a public process answer if everyone asks it, refine fit criteria if leads are often mismatched, and assign a daily queue owner if nobody follows up. Those observations measure service better than claiming a consultant is always online.

If the consultant offers several services, AI should not push every visitor to the same booking button. Some need only a public article, some fit a paid diagnostic session and some need another specialty. Publish each offer's goal, suitable audience, prerequisites, time commitment and unpromised outcomes. Ask what the visitor hopes to achieve and where progress is blocked, then suggest an approved next step. 'Increase revenue' alone is not enough to select an expensive package.

Define professional boundaries for the owner too. Questions about law, medicine, investment or tax beyond the consultant's qualification must not receive tailored conclusions under that person's name. AI can state the service scope and a suitable route for help. Even a qualified consultant needs a formal relationship, enough information and appropriate confidentiality before individual evaluation. Website copy cannot substitute for one-to-one work.

For delayed response, promise a pace with an owner, such as checking the queue on workdays, instead of 'we will contact you in minutes' when that is rarely possible. During leave, change the public contact note, close unstaffed appointment slots and state the actual return date. Respect an explicit request to stop contact rather than repeatedly prompting because an automated follow-up exists. A credible service rhythm matters more than an always-online appearance.

Define when intake should stop. If a visitor pastes employee names, pay data or private financial reports into open chat, direct them to an appropriate controlled route rather than asking for more sensitive fields. A public service page can give three steps: what issues are in scope, what non-confidential background to bring to a first call, and how a human will decide fit and later document exchange. If the consultant teaches all Wednesday, the promised response window must allow for it; AI may receive a message without claiming immediate review. Monthly, check whether handed-off cases received a reply and rejected enquiries got an understandable reason. If the queue grows, adjust the stated timeframe or intake capacity rather than collect leads indefinitely.