Module 5 — After and Around the Raise · Lesson 5.4
An Agent-Assisted Fundraising Operation, Governed
What to delegate, what to approve, and what the law still holds you to
What you'll learn
- Divide fundraising work into what an agent may do alone, with approval, or not at all
- Apply the CAN-SPAM requirements to automated commercial outreach
- Explain why regulators treat AI-assisted communications like any other
- Configure approvals, suppression and fact-checking so automation cannot outrun governance
The work of a raise divides cleanly. Some of it is judgement — which investors to trust, what price to accept, what to promise. Much of it is labour — researching firms, keeping records current, drafting the fifth follow-up, noticing which deal went quiet. An agent is good at the second kind and should not be trusted with the first. The question is not whether to use one; it is how to draw the line and enforce it, so that the speed is real and the risk is not.
Three lanes of work
A useful division — Kavanah's recommendation, not a regulatory standard — puts every fundraising task in one of three lanes.
Alone: reading and recording. Summarising the pipeline, finding stale deals, ranking investors against a profile, extracting what a data room says, logging a meeting, drafting a message for a person to review. None of this reaches anyone outside the organisation, and all of it can be checked.
With approval: anything that leaves the building or commits the company. Sending an email, enrolling investors in a sequence, sending a document for signature, starting a campaign. The agent may prepare all of it; a person decides that it goes.
Not at all: judgement that carries legal or fiduciary weight. Deciding whether an investor is accredited, whether an offering may be publicly promoted, whether an issuer's claims are true enough to recommend, what terms to accept. The agent can gather the evidence for each; a responsible person, and often counsel or compliance, decides.
The lanes matter most for unattended work — schedules, automations, background campaigns — because nobody is watching when they run. Those should default to the first lane, with anything in the second parked for a person.
The law applies to the message, not the sender
Automated outreach is still outreach. The FTC's CAN-SPAM compliance guide makes three points that bite on any automated sequence of commercial email. The law makes no exception for business-to-business email. Opt-out requests must be honoured within 10 business days. And hiring someone else to send your email does not let you contract away your legal responsibility — an agent or a sending platform is no different. Each separate email in violation is subject to penalties of up to 53,088 dollars, the figure as updated for inflation in January 2024.
For broker-dealers, FINRA said the same thing about generative AI directly. Regulatory Notice 24-09, published in June 2024, reminded members that FINRA's rules are intended to be technology neutral and that the rules, and the securities laws generally, continue to apply when firms use generative AI just as they apply to any other technology. Firms using such tools in their supervisory systems — reviewing correspondence, for instance — should address technology governance, including model risk, in their policies.
And the securities-law lines from earlier in the course do not move because a machine drafted the message. A 506(b) raise may not use general solicitation (Lesson 1.4); a mass sequence to strangers can drift across that line faster than a person typing could. Decide the exemption and the outreach rules before the agent starts, and write them where the agent can read them.
Controls that cannot be skipped
A control that depends on the agent remembering to apply it is not a control. The ones worth having are enforced by the system, at the point where the risk occurs.
Approval at the wire: outbound messages the agent prepares wait for a person to approve them, and the default for an unconfigured category is to park, not to send.
Suppression at both ends: an address that has opted out or bounced must be checked when a sequence step is generated and again when the message is about to leave, because the opt-out that matters is often the one that arrived after the step was scheduled.
Stop on reply: a sequence must end when the person replies, and a meeting on the calendar should stop it too.
Verified facts only: personalisation drawn from the web must be checked against its source before use, because a confident sentence about the wrong person is a misrepresentation.
A complete record: every action the agent takes, executed, parked or blocked, is logged, and what leaves the building is archived where the rules require it.
Measuring whether governance works
Governance that is never measured decays into ritual. Three questions, asked weekly, keep it honest. How many agent-prepared messages were approved, edited or rejected — and are the rejections teaching you something about the drafts? Did anything reach a person who had opted out? Did any personalised claim turn out to be wrong?
The aim is not zero automation or zero approvals. It is a steady state where the agent does more of the labour each month, the approval queue holds only the decisions that deserve a person, and the record shows both. That is what lets a founder raise faster, and what lets a regulated desk tell an examiner exactly what happened and why.
Set the lanes in Kavanah before the agent runs a campaign
- 1
Review parked actions in the approvals queue
Outbound email the agent prepares in governed runs parks for approval rather than sending; the queue lives in Settings → Agent, where each item can be approved or denied. Unattended runs in safe mode have send and delete tools stripped entirely.
- 2
Cadences stop on reply and when a meeting with the contact is booked; opt-out replies and hard bounces suppress the address; suppression is checked when a step is queued and again at send; new cadences include an unsubscribe line and network cadences always do. Investor campaigns release later tranches automatically only if you turn auto-release on.
- 3
Use verified research, and know what it costs
research_contact stores only facts whose quote is verbatim on the source, re-read independently, about the right person, and judged to support the statement — typically 25 to 40 credits per contact. Automatic research for personalised cadence steps is a separate opt-in, off by default.
- 4
Write the rules where the agent reads them
Put the raise's exemption, who may be contacted and how, and what the agent must never say in the workspace charter at /workspace-kvn and in a /knowledge page the agent can read. And if you are regulated, turn on compliance archiving in Settings → Audit logs.
Governance you can show
- Approval outcomes
- Share of agent-prepared outbound actions approved as drafted, approved after edit, or denied.
- Healthy signal: A Kavanah suggestion: rising approve-as-drafted over time. Persistent denials in one category mean the drafts or the instructions need work.
- Messages to suppressed addresses
- Count of outbound messages delivered to an address on a suppression list after it was added.
- Healthy signal: Zero, always. Any non-zero count is a control failure to investigate the same day.
- Personalisation errors reported
- Recipients or reviewers flagging a factual error in a personalised message.
- Healthy signal: Zero. One is a reason to pause personalised steps and review the facts behind them.
Key takeaways
- ·Let the agent read, rank, draft and record alone; require approval for anything that leaves the building; keep legal judgement with people.
- ·CAN-SPAM has no B2B exception, requires opt-outs within 10 business days, and cannot be delegated away.
- ·FINRA treats generative AI like any other technology: the same rules apply to what it produces.
- ·Enforce controls at the wire — approval, suppression, stop on reply, verified facts, a complete record — and measure them weekly.
That completes the course. A raise is a pipeline: the right money, the right instrument, the right investors in the right order, a story backed by evidence, a process with stages and a deadline, and the rules kept throughout. Kavanah is built to hold that pipeline so you can spend your attention on the parts that need judgement — which, in fundraising, is most of what matters.
Sources
- 1.CAN-SPAM Act: A Compliance Guide for Business
Federal Trade Commission · FTC Business Guidance (penalty figure updated January 2024) · 2024
No exception for business-to-business email; opt-outs honoured within 10 business days; responsibility cannot be contracted away; penalties of up to $53,088 per violating email.
- 2.Regulatory Notice 24-09: FINRA Reminds Members of Regulatory Obligations When Using Generative Artificial Intelligence and Large Language Models
FINRA · FINRA Regulatory Notice (June 27, 2024) · 2024
FINRA's rules are technology neutral and continue to apply to generative AI; supervisory use should address technology governance including model risk.
- 3.17 CFR 230.502 — General conditions to be met
U.S. Securities and Exchange Commission (eCFR) · Electronic Code of Federal Regulations, up to date as of October 21, 2026 · 2026
Rule 502(c)'s prohibition of general solicitation and general advertising, which automated outreach in a 506(b) raise must stay inside.