Sam Dundas

The through-line

Dundas's consistent argument is that guest communication is the wedge, not the destination. Besty AI began by trying to do one thing extremely well, because "all of the products that are out there um start by doing one thing extremely well and then expanding" [2], and he is explicit that the temptation to sprawl is constant: "there's a million things we could go with the product and a million things we could do but for for one let's just focus on one thing specifically" [2]. What has shifted over time is his sense of how shallow the messaging layer is on its own. Once a conversational assistant became table stakes, available natively inside every major property management system, he began describing chat as "really just like the scratching surface of what these things should really be capable of doing over the long term" [1], and pushed the product toward task detection, review analysis, data writes back into the PMS, and revenue generation [1][2][8][10].

The second constant is a refusal to oversell autonomy. He keeps a human in the loop by design, insists the honest position is that nobody is replacing people yet [2], and treats reliability as the metric that matters more than breadth of coverage [9]. The ambition is scale and a much bigger market, but the sales pitch stays deliberately modest.

On doing one thing well before expanding

The founding product promise was narrow on purpose: connect to the data already in a customer's PMS and help them answer guests, "whether that's an auto pilot that responds when it's confident or just generates a draft for your team to approve," with particular value in "blind spots around inquiries when you're not at the not at your desk 247" [2]. He frames adoption as a design constraint rather than a marketing problem, wanting it "effortless to turn on effortless to start using" [2], and has said Besty focuses on plugging practical AI directly into the PMS rather than shipping a separate tool, precisely so that adoption is easy [9]. Distribution followed the same logic: six integrations with major PMSs, with more planned, has been a key focus [1].

Expansion, when it came, was built out of complaints rather than roadmap theory. He describes a scale of "how much work is it and how much fun is it," and review responses sit at the bad end of both: a lot of work, not fun, and "the differential between like the best possible review response and like the worst one it doesn't make a huge impact but at the end of the day you're still hiring someone to do it" [2]. So Besty automates it, and he is unromantic about the result: "that's nothing to write hum about but that's one self-contained one of the 25 things you do on a daily basis you just don't have to do anymore" [2]. Beyond guest messaging and reviews, the product has moved into upselling, property improvement recommendations and retargeting past guests [10], including gap night upselling [8], with revenue-generating features and deeper integrations flagged as the direction of travel [9].

On autopilot with someone still able to grab the wheel

His clearest metaphor for where the industry actually is comes from autonomous vehicles: even after the cars worked, "up until that point they still had the driver in the seat even though he wasn't touching the wheel they had to have him there," and so "it's autopilot but there's still somebody that needs to be able to grab the wheel" [2]. He puts a horizon on it, arguing that for "at least sort of I would say 18 to 24 months" there will need to be a person, and that trust is the binding constraint rather than capability: "who to sit here and promise that it can do everything like would you even trust that" [2]. He is pointed about vendors who claim otherwise, saying that among the many people working on this problem, including PMSs themselves, "I don't think anybody's out there saying you're going to completely replace people yet I don't think that's honest nor is it really the expectation" [2]. The reframing he prefers is efficiency: "we don't need to completely replace people nor would we really necessarily more to but I think you know giving them better workflow tools and making them more efficient like that's something we can all agree on" [2]. That extends to what AI is allowed to decide. Surfacing a maintenance issue from reviews is fair game, but dispatching people is not: "we don't want the AI assigning your human capital things to do because that still needs sort of a judgment layer" [2]. He has argued elsewhere that AI can cut operational costs and give operators time back while preserving the essence of hospitality [11], that personal touch remains crucial despite the technology [9], and has engaged directly with the harder question of what lower operational costs mean for hospitality jobs [12].

On reliability, error rates and the ways an AI gets gamed

Dundas is unusually specific about failure. Autopilot shipped because "that's what the market wanted," but he was wary of it: "we were scared it was going to make a one or two mistakes and then they would just lose faith in the whole thing," so he first ran it only on his own properties before accepting that "at certain point you have to just let it go of the wild and learn on the go" [2]. Early on the error rate was "like probably 3 or 4% of messages," and the method was brutally manual: find an issue, "we hardcut it to never make that mistake again," repeated over months and thousands of conversations until "you pretty much found it all" [2]. He now puts hallucinations, unauthorised refunds and unapproved early check-ins at "less than 0.1%," with his largest customer running 1,500 properties on 24/7 autopilot for three months without turning it off [2], and he stays candid that "it is still probably not perfect" [2].

The failure mode he finds most instructive is social rather than technical: a guest who says the assistant already promised a refund, and the AI agrees. "Then you're like now you have to backtrack and say sorry that was somebody from our team or do you tell them that was like they're talking to an AI like how do you explain that to the guest you're just gonna end up giving them what they asked for but then you pissed at us which is like understandable" [2]. This is why he ranks reliability above coverage in AI responses [9], and why he describes consistent, dependable answers at enterprise scale as "much easier said than done" against the ease of spinning up a chatbot [1]. He has also argued for transparency in how vendors communicate with operators, and for the value of actually trying software to see whether it fits a particular operation [7].

On the operator scale where AI actually pays

His ideal customer is defined by unit economics, not enthusiasm. The sweet spot is 50 to 250 properties [2]. Below that, the math traps you: at ten or twenty units you are still doing everything yourself, and "if I were to hire someone then that just erases all the margin that I've been able to generate and I'm back to square one," so "the amount of profit you generate at 10 properties and 20 ends up being the same" [2]. Above roughly forty or fifty units, a dedicated onshore or offshore support resource becomes justifiable as a salary [2]. In that band, operators commonly run two or three agents on rotating shifts at once, and the value is either collapsing that to one or keeping all three and redirecting them to other projects [2]. He is clear the product still serves smaller hosts, noting about a hundred customers in the five to fifty range [2], and has argued that AI can democratize efficiency gains across large and small operators alike [6].

At the top end, his conviction is that solving for the hardest customers solves for everyone: "we can make it work for them, we can make it work for for anybody" [1]. That has meant serving 45 users with over 100 listings and seven with over 1,000 [1], on the way to 17,000 paying properties and approaching a million in revenue inside a year [9].

On task management as the next layer

The complaint that drove the current roadmap is that autopilot can quietly recreate the work it removes. "A lot of people switching into autopilot but what we notice is they spend more time just uh reading through the conversations on autopilot then would have taken to just respond manually so kind of defeats the purpose" [2]. He hears the same thing from larger customers: now that an AI handles conversations, how do I stop spending more time reviewing them than paying someone to do the job would have cost [1]. The answer is real-time task detection, reading a whole conversation as a message arrives to spot anything needing physical intervention, towels, a lockout, a clogged toilet, then producing a one-tap summarised task that gets assigned to the right person and pushed into whatever task system the operator uses, native or third party [1][2].

He separates that time-sensitive stream from slower, retrospective analysis, where the review tasks feature reads recent reviews for a listing and drafts actionable items for the team, less urgent but still worth an AI having a perspective on [2]. Further out, he sees the real prize in writing reliable data back into the PMS rather than just reading from it: updating check-in and checkout times semantically when a guest says "we're out" or "my flight arrives at 7 p.m.," pushing that into the PMS as a dependable field [1]. That precursor unlocks the harder problems, the "whole cleaner Tetris" of housekeeping scheduling, and making the most of every turn through dynamic upselling [1]. He puts a realistic timeline on it, roughly four to six months of work even with fresh capital [1].

On hotels, OTAs and the blurring of the two industries

Dundas reads the convergence of hotels and short-term rentals as his next market, and he is candid that he is entering it as a learner: "we don't know anything about hotels right now" [1]. He sizes the opportunity plainly, contrasting roughly $1.1 trillion a year in hotels against short-term rentals at around $170 billion [1]. The entry point is not the full Opera and Oracle stack, which he concedes is the sexier target, but the specific mess created when hotel brands push inventory onto Airbnb and Vrbo. A hotel's customer journey is one-sided, email confirmations and a front desk, with "no inbox" and no concept of a host you converse with [1]. Add seven thousand units to an OTA and "you've now basically opened up this entire Pandora's box of guest communications, which, you know, your team isn't necessarily equipped to do the same way they were handling, say, phone based reservations" [1]. His plan is to be the unified inbox for hotel groups moving into the OTAs, targeting brands on hotel PMSs that offer OTA distribution, with partnerships in the works and a route to direct OTA integration as a byproduct [1].

The underlying bet is about where consumer attention sits. Even if a city like New York cuts Airbnb inventory by 95%, the mind share stays, so "the hotel groups are going to go to where the guests are," pulled by Airbnb's marketing spend and sponsorships [1]. He also thinks short-term rentals have manufactured an expectation that hotels are about to inherit: sub-five-minute response times, an SLA that no newly expanded team can staff manually [1].

On raising money and what investors actually test

He is an open book about venture mechanics and describes his own expectations being wrong. Coming from business school, he assumed fundraising would be rigorous and empirical, like getting a mortgage, "very sort of empirical based like evidence focused," and found it was closer to the opposite [2]. His explanation is structural: venture is a sliver of an institutional portfolio, mostly treasuries, fixed income and stocks, with a few percentage points in high-risk assets, and when interest rates fall the cost of borrowing drops and the appetite for that sliver rises [2]. He and his co-founder were beneficiaries of that in November 2021, raising on a pitch deck and a story with "no product we had no no customers" [2].

The reckoning came when that first startup went nowhere. Rather than dressing it up, they had "a pretty honest conversation like what are we doing here we're sort of phoning it in," threw the idea out, kept "same cap table same investors," and restarted on AI for short-term rentals, a domain where he had genuine operating experience [2]. The second raise, closed in July, looked nothing like the first: about 120 calls with funds over two and a half months, all on Zoom, a five or six minute pitch and then a grilling [2]. The recurring objection was market size, "how big is this Market really um and how does this become like a multi-billion Dollar business," which he accepts is a fair test, because a seed investor's risk-adjusted return does not work on a two or three times outcome [2]. He is equally clear about what the money changes internally: a moment to "breathe and like reset and think like take a longer term view" and build a stronger product than two people could alone [1][2]. On team building, he has stressed that scaling comes down to smart delegation and finding the right people, and that an open mindset is essential for startups [12].

On competition and being everywhere

He does not pretend the space is defensible on features. Between his first and second appearances the number of competitors multiplied: "when we were starting I thought there was a lot but there's probably five times more and you look at every single PMS partnership announcement there's like three or four aai startups" [1]. Front-line communication delivered by a vendor, or natively by the PMS, has become "that sort of expectation table stakes" and is now ubiquitous across every major PMS [1]. His response is to go a layer deeper into workflow and to lean on presence in the market. He is self-aware about the founder circuit, noting that "everybody has a sort of flash independ moment" and that he is deliberately chasing exposure, while insisting the in-person part is genuinely the best of it: "getting to meet like real customers and person and it's it's the funnest part of the job" [2].

Takeaways

• Guest chat is the entry point, not the product: he calls the chat layer "really just like the scratching surface of what these things should really be capable of doing over the long term" and is pushing into task detection, review analysis and PMS data writes [1][2]
• The economic sweet spot is 50 to 250 properties, because below roughly 20 units hiring "just erases all the margin," while above 40 or 50 a dedicated support resource is already justified as a salary [2]
• Keep a human able to intervene: for the next 18 to 24 months "it's autopilot but there's still somebody that needs to be able to grab the wheel," and claiming full replacement is not honest [2]
• Reliability beats coverage: early autopilot error rates of 3 to 4% were driven below 0.1% by hard-coding a fix after every discovered failure, over thousands of conversations [2][9]
• Guard against gamed refunds and unauthorised early check-ins, where a guest claims a promise the AI never made and the AI agrees, leaving the operator to backtrack or explain the bot [2]
• Autopilot fails commercially if operators end up reading every conversation, which is why real-time task detection and one-tap assignment matter more than better prose [1][2]
• Hotels pushing inventory onto Airbnb and Vrbo open "this entire Pandora's box of guest communications" with no inbox culture and no team sized for five-minute response expectations [1]
• Venture appetite is driven by interest rates, not diligence rigour, and the standing objection to short-term rental startups is market size and the path to a multi-billion dollar outcome [2]