In the spring of 2025, I watched a team pitch an AI phone agent for independent restaurants. One reviewer played a customer and changed her pizza toppings halfway through a sentence, and the agent caught the change. The room liked the demo. The doubts that followed were about distribution and incumbents, and one reviewer pointed out that a large phone-systems vendor had just launched a voice agent of its own. He wanted to know whether an established company would ship the same feature.
Eighteen months later, I sat on the same accelerator's selection committee for a new cohort. A founder presented a collaboration layer for teams of coding agents, and a senior reviewer remarked that he could probably have his own agent build the product later that afternoon.
That reviewer was also the competitor, and none of the startups in the category could match his speed or his cost.
Startup screening has changed a great deal in the last year or two, mostly because of AI, and the committee's conversation shows where. The earlier sample is a pair of final-round sessions, which can show what was missing then but cannot measure it. This fall covers more than a hundred first-round pitches and their finals.
A year or two ago, AI was a category the committee evaluated. By this fall it had become the lens the committee used on every company.
The threat moved from the incumbent to the model
Back then, the defensibility questions were about incumbents. A grocery chain could add a label-scanning feature to its app. A cloud provider could match a startup's price cut. A phone-systems vendor had already launched a voice agent. Reviewers cited a venture firm's market map showing how quickly the voice-agent category was filling up. A crowded category was the risk, and a feature launched by a company with distribution was the threat.
One remark from that session hinted at what came next. A reviewer asked whether a large database of scanned products still counted as a moat when AI tools let any newcomer build the same database through usage and effort. At the time it was a passing doubt. By this fall it had become the committee's working assumption.
The committee now names the threat directly. One partner admitted he found it hard to stay optimistic about any company once someone asked whether a frontier lab could do the same thing, since in his view the labs can do nearly everything. Another reviewer predicted that frontier models moving into healthcare would make a founder's custom AI obsolete quickly. A third recalled the wave of generic agent startups that disappeared once the major labs released agents of their own.
The partner also made an argument about timing. He said the field had shifted so far in six months that a product which made sense half a year earlier might not make sense now. He added that the model the company was competing against was the weakest version that would ever exist.
Incumbents still come up, but the ones reviewers name now are AI platforms: a code host's agent feature, a cloud provider's model foundry, and a weather model from a frontier research lab. One reviewer expected the large office-software suites to build a collaboration tool's core feature into their products before long.
In this fall's first round, the question of why a frontier model couldn't just do the job came up in about one pitch in five. Founders handled it badly six times in ten. The answers that worked named a constraint a lab does not face and cannot cheaply remove. Examples included data that cannot leave the customer's premises, a validated toolchain, and a deterministic verification step a general model does not perform. Answers that argued for speed or focus lost the room.
The reviewer became the competitor
The test in the opening scene never came up in the earlier sessions. This fall it appears across the software pitches. One reviewer estimated he could rebuild a product in half an hour with an AI app builder. Another said developer tools had become a hard sell because strong engineering teams can now build the same thing in-house with AI. A third argued that open agent frameworks already handle the step-by-step integrations a startup was presenting as its differentiator.
Until recently, calling a product easy to implement meant another startup could copy it within a year. Now the same doubt means the person on the call could copy it that afternoon. The comparison point has shifted from a rival company's roadmap to one engineer working with an agent. For a large class of software products, defensibility now has to hold up against someone already sitting in the room.
Traction buys the second question
Earlier, the committee still rewarded customer discovery and early revenue on their own merits. A reviewer praised one team for the quality of its primary market research, and the praise counted.
This fall a reviewer acknowledged that a team's traction was clearly real and then observed that everyone seemed to have traction now. Another noted that nearly every startup was claiming a data flywheel as its moat. Two companies with real revenue were passed over in the same week because nobody could say what they owned once the revenue was set aside. Traction still opens the conversation, and founders still have to answer the question that follows it.
The room listens for a machine's voice
In the earlier sessions, nobody suggested a pitch had been written by a model. The closest anyone came was a reviewer's mild relief that one security pitch hadn't mentioned AI at all.
This fall, reviewers had an ear for AI-written decks. They identified at least seven pitches as generated, and in one session they flagged three out of seven. They also named the tells:
- absolute claims that nobody has ever done something, or that every customer needs it
- top-down market sizing
- strings of statements that deny one framing in order to offer another
- a long problem section before the product appears
- an opening that walks through the history of the field
One reviewer said it was a relief to see a deck a person had actually made. Another advised founders who let a model draft their deck to change enough of it that the reviewers can tell a human looked at it.
The suspicion has spread to live delivery. Reviewers wondered whether some founders were reading from a generated script, and at least once a founder flagged that way was defended by a colleague and went on to a strong reception. The committee also removed a founder's AI note-taker from the call.
"AI" in the name now needs proof
A year or two ago, adding AI to a product was neutral. This fall, one partner compared AI to the old dot-com suffix after noticing that a company's AI appeared only in its name. Others made the same point in different ways. One reviewer called a parking app's AI unnecessary. Another asked whether a compliance tool had any AI in it at all or was ordinary software. A third, hearing a sensor startup describe its model as foundational, asked why every company now claims a foundation model and called for some truth in advertising.
Founders now pay for an AI claim they cannot back up, and they also pay when they build on top of AI without adding anything of their own.
The committee has its own AI
The larger structural change happened on the committee's side of the table, and nothing like it existed in the earlier sessions.
This fall, an AI agent worked alongside the committee. It scored every application before a human read it and suggested which reviewers should look at which companies. During interviews, reviewers asked it for questions to put to the founder and whether a company's claimed customers could be found online. When a pitch lost the room, they asked it to explain what the company did. Staff were enthusiastic about it. The partner who ran it was more careful, reporting that it agreed with human advisors most of the time and that he still did not fully trust it.
Reviewers also brought AI-generated research into the discussion while founders were still presenting. The research included a model's estimate of a market's size, a model's list of competitors, and a model's summary of an acquisition that undercut a founder's position. The earlier committee checked founders against search results and professional profiles. This fall's committee asks a model as well.
The bot has also become a yardstick for founders. After one weak pitch, a reviewer suggested the bot could have presented the company better than its founder had. The bot's verdict on another company, that the deck was sharper than the founder, was repeated around the room.
What has not changed
Several patterns predate the current wave, and both periods show them:
- The team still comes first. In 2025, a reviewer called one company a possible bet on the founders and wished they had brought a sharper insight. This fall, reviewers made the same call on a collaboration-tools team. Another 2025 reviewer described a third kind of bet beyond the team and the product, on a team so plainly fundable that it would win the race for venture money.
- Numbers are checked against what you said before. In 2025, a reviewer caught a customer acquisition cost quoted one way in an email and very differently in the pitch. This fall, a reviewer caught ARR that had dropped between the submitted deck and the interview.
- Live checks on founders are older than the bots. In 2025, a reviewer discovered during the pitch that a co-founder's professional profile still listed a full-time employer. These checks happen more often now and involve more people, but they aren't new.
- Finalists are expected to act on first-round feedback. In 2025 one finalist gave essentially the same pitch as in the first round, and the partner who had given that feedback said so. Another clearly took the feedback on board, came back with a different pitch, and won the room over.
What founders should do differently
Prepare for this year's version of the moat question. Your defensibility answer has to hold up against a frontier lab shipping your product and a reviewer rebuilding it with an agent, as well as an incumbent copying a feature. Name the specific constraint that protects you, such as data access, a validated process, a regulatory position, or a deterministic step a general model does not perform. Then explain why a large lab cannot cheaply remove it.
Assume an AI read your application before any person did, and that a reviewer will ask it about you while you are presenting. Make every claim checkable, keep the deck consistent with the application, and make sure your team's public profiles match what you say.
Write your own pitch. Using a model to draft is fine, but the reviewers have learned to hear its cadence, and a deck that sounds generated costs you credibility before you have said anything specific. Strip out the absolutes and size the market from the bottom up. The product should appear within the first minute.
Treat traction as the start of the defensibility conversation. The revenue earns the question of what you own, and the answer is what moves the room.
The questions have changed faster than the answers founders bring. In the rooms I sat in this fall, the founders who did well had prepared for this year's version of the moat question.