From 620k USD to 3.1 EBITDA points: the AI content strategy by consumption moments that fixed the 3 p.m. hole, using Radar Gastronomico and meseros.ai

The 3:00 p.m. window cost 3,100 USD a month and nobody was looking at it

Between 3:00 and 6:00 p.m. this 42-seat casual dining room held 19% occupancy with the full front-of-house payroll clocked in, and that dead window drained roughly 3,100 USD monthly in unabsorbed fixed cost, per the reconstruction of its own P&L. The operation billed 620 thousand USD a year, had been open seven years in a Bogotá office corridor with 19 full-time employees and an average check of 41,500 pesos, about 10.4 USD. None of that was broken. The kitchen had process, the menu carried nine years of fine adjustments, and 58% of sales concentrated in three services of the week, which is exactly the kind of concentration an owner celebrates until somebody counts the paid hours with no guest in the chair. The problem was never creative: it was PRODUCTIVITY per hour of open payroll. The incoming brief carried 22 monthly posts, 31 thousand Instagram followers and zero demand attribution by time slot: no dashboard crossed content with sales, so the team measured reach, which costs money and says nothing whatsoever about the register.

They produced 22 posts a month and measured the priciest metric in the trade

That measurement gap is hardly a local quirk. More than 60% of restaurant orders already come through mobile apps (Restroworks, 2025) and roughly 40% of sector sales run through online ordering (Statista), so the digital trail exists; what was missing was tying it to the guest's arrival hour. When an operation publishes blind, it isn't buying demand: it's buying impressions that fill the tables which were going to fill themselves anyway. That is where the audit started, not in the feed. Treating «the week» as a planning unit is far too coarse for a restaurant, which is why the first move of the Masterestaurant method was splitting the seven days into six slots with distinct customer, check and service-speed profiles. The office-corridor slot between 3:00 and 6:00 p.m. does not buy what the Friday 9:00 p.m. slot buys, nor does it tolerate the same wait.

The planning unit stopped being the week and became the TIME SLOT

With that granularity, the Radar Gastronómico projected demand per slot instead of per day, which is where the decision actually lives: how many servers open, what mise en place gets prepped, which message goes out at which hour. We went from 22 generic posts to 18 dated pieces aimed at the three slots with the worst fixed-cost absorption. Less volume, better aim. Publishing more was never the problem. AI did not write the posts here: it forecast occupancy by slot using POS history and the corporate corridor's calendar, and on top of that projection the team decided what to push. It is worth turning down the technological enthusiasm, because the sector itself shows the ceiling: close to 21% of AI-assisted drive-thru orders still require employee intervention (Intouch Insight, 2025), and that margin of human error does not vanish by buying better software. Latin America accounts for barely ~6.4% of global revenue in the restaurant AI market, with a projected 23.1% CAGR through 2034 (Dataintelo, 2025), so the edge of adopting early in the region is real and still cheap.

AI projects the slot; judgment decides what gets pushed

A badly used slot forecast fills the room on the wrong day. Used well, it moves margin to where fixed cost is already paid. Aggregate weekly check went from 14,900 to 17,550 USD, EBITDA climbed from 6.4% to 9.5% and Prime Cost dropped from 68.1% to 63% across seven months, according to the case's monthly closes. Those extra 2,650 USD a week came almost entirely from low-turnover slots, where fixed cost was already paid and only the plate's food cost got added, which is the arithmetic reason a 17.8% sales lift produced 48% more EBITDA. Occupancy in the dead window rose from 19% to 46%. That also explains Prime Cost: the same 19 full-time employees spread over a larger sales base. I got this wrong for years by recommending a price increase first; the correct order is to fill the paid hour and review the menu afterward.

What if the demand in that slot simply isn't there?

Suppose the forecast says 46% occupancy is available at 4:30 p.m. and the room still sits at 19% after eight weeks of dated content.

The reading then is not «the content failed»: it is that the corridor holds no available population at that hour, and the next step stops being publishing and becomes closing the slot, cutting two front-of-house shifts and recovering through cost what sales will not bring. That is the uncomfortable tension in the method: the same measurement that justifies investing in a window is the one authorizing you to kill it. An operator who only accepts the favorable reading of his own dashboard isn't measuring, he's shopping for permission. In this case the demand did exist —the corridor held offices with 4,200 employees eight blocks away— but that verification happened BEFORE the first piece was produced.

Transferable lessons by annual revenue band

Below 500 thousand USD a year, this week's first step is pulling hourly sales for the last 90 days out of the POS and flagging the three slots with the worst fixed-cost absorption: no dashboard required, a spreadsheet does it. Between 500 thousand and 1 million —this case's band— the move is assigning dated content to those three slots and measuring aggregate weekly check, not reach. Above 1 million, demand slot-level attribution in the reservation system before contracting more production. Over 5 million, with several locations, each slot gets projected per venue because an office corridor and a residential neighborhood never share a curve. The archetype of the media chef running a 300-seat themed restaurant above 10 million faces the inverse problem: surplus demand at peak, shortage in the valley, and there the lever is seating capacity per turn. I would not expect this result in three contexts.

Limits of this case

First, a tourist-destination restaurant, where demand arrives by season and visitor flow rather than by the guest's weekly habit: there the slot moves with the airline calendar and dated content has little leverage. Second, an operation with more than 70% of its sales in delivery through aggregators, which already concentrate 67% of global online orders (Business Research Insights, 2025) and hand the restaurant no control over the hour the algorithm surfaces it. Third, any venue with food cost above 38%, because filling a low-turnover slot on broken margin multiplies the loss instead of absorbing fixed cost. The case profile is an anonymized composite of recurring patterns in Diego F. Parra's practice across more than 8,400 restaurants in 43 countries, not an identifiable client. The first difference is one of object: the old method produced CONTENT, the new one produces DEMAND with an arrival time. That sounds like a semantic nicety until you open the P&L and realize a post that fills a table on Friday at 9 p.m.

The four differences that moved cash

contributes nothing, because that table was filling anyway; the same post pushing Tuesday at 4:30 p.m. is worth its weight in margin, since the fixed cost of that hour is already paid and only food cost is added. Second comes temporal granularity. Treating "the week" as a planning unit is far too coarse for a restaurant: the real unit is the DAYPART, and this operation had six of them with completely different customer profiles, check sizes and service speeds. Radar Gastronomico projects at daypart level because that is where the purchase decision lives, not in the weekly aggregate that soothes the owner and never tells him where to act. Third is who decides the discount. It used to be decided at 3:40 p.m. with an empty room and morale on the floor, the worst possible moment; now it is decided fourteen days ahead, with contribution margin calculated dish by dish and one hard ceiling: no offer may push that dish's food cost above 32%, which is the maximum in the costing contract, not the target.

The four differences that moved cash — in practice

The fourth, and the hardest to accept, is that AI does not write the strategy. It executes at a speed no human team inside an MSME can afford, yet the judgment of which daypart to attack, with which dish and at what margin remains a director's call. I got this wrong for years, recommending automation first and thinking later; the correct order runs the other way, always, and this case proves it with seven months of data.