A board update goes out. Two sentences near the bottom mention you’re adding AI agents to the product. Six months ago that line would have gotten a thumbs up and nothing else. Today it gets a follow-up question, sometimes several. What does the agent actually do. What happens when it’s wrong. Is this something you built or a wrapper around someone else’s model.
Investors didn’t get more skeptical of AI by accident. They’ve seen enough vague AI claims fall apart under a second look, and in a few well-documented cases watched those claims turn into real legal exposure for the companies that made them. This isn’t about whether to mention AI agents to investors. It’s about what to actually say, what to leave out, and how to frame it so the claim still holds up the next time someone asks a harder question.
Why investors got sharper about AI claims
Some of this is regulatory, and it has set the tone for everyone, even companies nowhere near a public filing. US enforcement agencies, including the DOJ, SEC, and FTC, have pursued a growing number of cases against companies for deceptive AI claims. Notably, they have taken action against claims that were technically accurate but overstated or misrepresented, not just outright fabricated.
As highlighted in Global Investigations Review’s analysis of AI washing enforcement, one company settled with the SEC after claiming its product required no human intervention to take customer orders, when in reality most orders still required a person. The company also failed to disclose that the underlying AI technology was licensed from a third party rather than built in house. Two investment related firms separately settled AI washing charges for promoting AI driven capabilities they did not actually have.
None of that applies directly to a Seed or Series A company with no SEC filings to make. But the standard it sets has already moved into private fundraising and board reporting because the investors writing your cheques read the same enforcement updates your lawyers do.
Regulators are now treating forward looking AI roadmap claims as a flag when they are inconsistent with a company’s actual budget, staffing, vendor contracts, or product readiness. That is exactly the kind of mismatch an investor doing diligence on your deck is trained to look for.
A recent startup due diligence analysis by Evalyze found that roughly two thirds of the decks it reviewed contained at least one issue likely to be flagged during formal diligence, most commonly unclear unit economics or aggressive projections without supporting evidence. A vague AI agent claim with no specifics fits that pattern precisely.
What investors actually want to hear
Strip away the buzzword and most investors are asking four practical questions, whether or not they phrase it that way.
What task does the agent actually do. “AI-powered” is not an answer. “It drafts the first response to every inbound support ticket and a human approves before it sends” is an answer. Specificity is the entire signal here. Vague claims read as a placeholder for something that doesn’t exist yet.
Where does a human still sit in the loop, and where doesn’t one. This is the question that separates a credible claim from an inflated one. State exactly what the agent can do without review and what it can’t. If you can’t draw that line clearly internally, you’re not ready to draw it for an investor either.
What’s actually yours versus licensed. If the agent runs on a third-party model API, say so. Investors increasingly expect that disclosure as a baseline, not a weakness to hide. The case above is a direct lesson in what happens when a company implies proprietary technology that was actually a licensed third-party product. The defensibility question investors care about is usually your data, your workflow integration, and your domain tuning, not whether you trained a foundation model from scratch. Almost nobody at this stage did, and investors know that.
What number proves it’s working. Not “users love it.” A real number: tickets resolved without escalation, time saved per task, error rate, adoption rate among existing customers. If you don’t have a number yet because it’s early, say that directly and give a date for when you will.
What not to say
A few specific claims get founders into trouble later, even outside a regulatory context, because they are the ones investors remember word for word.
Don’t claim full autonomy if a human reviews the output. Say “human reviewed” or “human approved” if that’s accurate. Investors are not penalising you for having a human in the loop. Most expect it at this stage, especially as regulators continue to scrutinise how companies describe the role of humans in AI driven processes. As discussed in The D&O Diary’s overview of the 2026 SEC reporting season, companies are facing increasing expectations to clearly explain the extent of AI autonomy and human oversight.
Don’t let the word “proprietary” do work it hasn’t earned. If your competitive advantage comes from your workflow, proprietary data, or deep product integration, describe those strengths directly instead of implying you built the underlying foundation model.
Don’t promise a roadmap your team and budget cannot support. A roadmap that does not align with your staffing, vendor contracts, or available resources can quickly become a credibility issue. Once an investor spots that inconsistency, every future product update is likely to be viewed more critically.
Don’t bury the limitations. Be clear about what the agent does not yet handle, which edge cases still require human intervention, and which workflows you are intentionally not automating. That level of transparency is often more credible than presenting AI as capable of everything. According to Finrep’s analysis of AI related SEC disclosure expectations, SEC comment letter activity around AI disclosures has continued to increase, with regulators placing greater emphasis on companies acknowledging limitations and substantiating their AI claims rather than making broad, unsupported statements.
Framing it differently in a board update versus a fundraising deck
A board update is a hypothesis log, not a pitch. State what you tried, what worked, what you killed, and what’s next. Founders who run their roadmap as a testing process rather than a fixed commitment have an easier time here, because the update can honestly say “we tested this, it didn’t hold up, here’s what we’re doing instead” without that reading as failure.
A fundraising deck carries more weight per sentence, so it needs more precision, not more enthusiasm. The strongest version of an AI agent claim in a deck ties directly to a metric an investor already cares about: retention, support cost per customer, time to value for a new account. The agent is the mechanism. The metric is the actual claim.
Keep the two consistent. The line in your fundraising deck and the line in your board update six months later need to describe the same reality. Investors keep notes. A claim that quietly downgrades from “automates” to “assists” between two updates, with no explanation, reads as exactly the kind of gap diligence teams are now trained to catch.
What this sounds like in practice
The difference is concrete enough to write side by side.
Weak – “We’re AI-powered and built for scale.”
Stronger – “Our support agent now drafts a response for every inbound ticket and a human approves before it sends. It’s not yet running on billing disputes. We’re tracking time-to-first-response as the metric, and we’ll have a resolution-rate number by next quarter.”
The second version says less in terms of marketing language and more in terms of actual information. That’s the trade an investor is implicitly asking you to make every time they ask a follow-up question about an AI claim.
The standard investors expect from AI claims
The pressure to put “AI agents” somewhere in a pitch deck or board update is real this cycle, and pretending otherwise isn’t useful advice. The more useful move is changing what sits behind the sentence.
If you’re preparing to present your AI roadmap to investors or your board, schedule a call with us to review how your AI agent capabilities are positioned before they become part of your next update.
Investors aren’t asking founders to slow down on AI. They’re asking for the same thing they’ve always asked for, applied to a newer category – a specific claim, backed by a specific number, with the limitations named instead of hidden. That’s a higher bar than a buzzword, but it’s also a bar a real product can actually clear.
Your queries, our answers
Generally yes, in substance if not in exact wording. Investors have seen cases where a company implied proprietary technology that was actually licensed from a third party, and it became the central issue in a regulatory settlement. Saying which parts are licensed and which parts are yours protects you more than it costs you.
No. Most investors expect it at this stage. The problem is claiming full autonomy when that's not what's actually happening.
Specific enough to be falsifiable. "It's helping" isn't a metric. "We're tracking ticket resolution rate and expect a real number within a quarter" is, even without a final figure yet.
Not directly in most cases, since that enforcement has mainly targeted registered investment advisers and public companies. But the standard of specificity and substantiation it has set is shaping what institutional investors expect to hear well before any company gets near a public filing.
Leading with the word "AI" instead of the task. The task is what proves the claim. The word alone proves nothing.
What happens after you fill-up the form?
Request a consultation
By completely filling out the form, you'll be able to book a meeting at a time that suits you. After booking the meeting, you'll receive two emails - a booking confirmation email and an email from the member of our team you'll be meeting that will help you prepare for the call.
Speak with our experts
During the consultation, we will listen to your questions and challenges, and provide personalised guidance and actionable recommendations to address your specific needs.
