The future of RM — AI agents and automation

Tuesday, 6:12 a.m. Daniel’s phone buzzes on the nightstand. Not the alarm — a notification: the overnight rate update has run, prices changed on 14 dates according to the configured rules; 2 dates are waiting for human approval. Over coffee, Daniel scans the list: all 14 moves are inside his limits — small increases where pickup is accelerating, two cautious corrections downward. Of the two waiting dates, one is the edge of an event weekend, the other an unusual combination of signals — the system held back precisely because that is what it was taught to do. Five minutes, two decisions, done.

Five years ago, those same 14 decisions ate the first hour and a half of the morning — calendar scrolling, manual re-typing, channel updates. Today they happened while he slept. And over the coffee, the thought takes hold that every RM will sooner or later have to say out loud: if the machine moves more and more on its own — what remains as my job?

This lesson — the second to last of the course — walks around that question. In lesson 51 we drew the dividing line: the machine sees in the data, the human sees in the market. That picture was a still image — a snapshot of the division of labour. This lesson sets the same frame in motion: what happens as the machine progresses from signalling through recommending to executing? Where does that motion stand today, where is it heading, and — the point — where does the RM’s role shift in the meantime? The answer is no secret: not that the role “disappears”, but that it becomes different — and more valuable. But it is not enough to declare that; it has to be calculated through. So we will.

The five levels of automation

The “how much does the machine do?” question is not yes-or-no; it is a scale. It helps to think in five levels — not an industry standard, but a working tool, and it is exactly this ladder that structures the daily decisions (what do I let go, what do I keep in hand):

A few words on the levels — and on where each stands today. L0 is the world of this course’s beginner level: the machine counts, everything else is human. L1 is the insight layer — the Insight Engine’s terrain (lesson 52): the coverage advantage of a machine that watches everything, every morning — but every move is still yours. L2 is the backbone of today’s advanced practice: rate recommendation (lesson 56), where the machine already names a specific number and you approve or override it with the verification routine. L3 is likewise an existing mode of operation today: automated pricing that applies the recommended rates on its own — while respecting the manually closed exception periods and the hard limits. Lesson 56’s closing image was exactly this: the automation in the quiet zone, RM attention on the outlier days. L4 is partly future: elements of it exist (models retrain on new data), but a system that autonomously tunes its own rule framework is the direction of the coming years, not today’s practice.

Three terms of art will stay with us from here on. A guardrail is the hard boundary the automation may never cross: rate floor and ceiling (lesson 35), maximum daily rate step, exception periods, escalation conditions. Human-in-the-loop is the arrangement in which the machine process requires human approval at defined points — L2 is this in its entirety, L3 at the exceptions. And exception handling is the human half of L3: the machine runs normal operations and hands the decision to the human in precisely defined situations — an unusual combination of signals, a framework boundary, low confidence.

What already works today — and what is likely coming

On a future topic, discipline is mandatory: let us separate sharply what is existing capability and what is industry expectation. The former is a statement, the latter an estimate — and mixing the two is exactly the kind of haze this course has spent 65 lessons teaching you to avoid.

Already today. The agent is the key concept of today’s AI world: a software component built on a large language model (LLM) that does not merely answer but works on a task autonomously — it uses tools, queries data, calculates, reasons in multiple steps. And this is not future tense: on conversational analysis surfaces, specialised agents are at work today. There is a forecast-advisor agent that explains the coming weeks’ forecast — why that number, which events move it, how accurate it has been so far — and there is a pricing-analyst agent that assembles the full signal package for a date (demand, competitor rates, events, own position, plan) and proposes a rate band, with a visible derivation. In Peaqplus these are part of the daily work in Pulse Chat (lesson 54). It also works today that a proven checking question runs as a scheduled routine (Pulse Routines) and its result arrives as a notification; that automated pricing moves on its own inside guardrails — that was the dawn notification Daniel woke to —; and that a decision you state in the conversation goes into the decision log (lesson 64), but only after your explicit confirmation: the machine records, it does not decide for you.

Likely, in the coming years. Three directions are taking shape — emphatically at the industry level, not as any specific product’s promise. The first is agent–agent coordination: today the pricing logic and the channel logic live separately; the next step is that they consult — the channel and restriction consequences of a rate increase (lesson 42) get thought through by the same process. The second is longer-horizon, autonomous campaign management: the agent does not move one rate but runs a programme for weeks — watching the effect, correcting, and reporting back, the way you would expect from a junior colleague. The third — perhaps the most interesting — is natural-language goal programming: the guardrail will not be a settings screen but a sentence. “Keep the RevPAR index — your RevPAR relative to the compset (lesson 44) — above 105, but do not raise loyal guests’ rates by more than 10%.” Two KPIs, one priority, one ethical constraint — in a single sentence you would today hand to a person.

So the frontier today runs in the L2–L3 band, and it slides upward year by year. But watch the direction: none of these developments is “RM without the human”. Every one of them does the same thing — the machine takes over execution, and the human steps up into goal-setting and framework-setting. What that step up means, precisely, is what we calculate next.

The worked example — what a level step is worth, and what it risks

Daniel made the decision half a year before that dawn notification: he stepped from the L2 mode (he approves every rate recommendation) to L3 in the quiet zone. Let’s look at both sides of the decision in numbers — because an automation-level decision is exactly like a rate: without a balance sheet, it is only an opinion.

The yield side: the time balance. Daniel’s weekly pricing work on the two levels:

The difference is 8 hours a week — counting 50 working weeks, ~400 hours a year. A full working day every week; ten working weeks every year.

That time is worth nothing by itself — what matters is what it goes into. In lesson 49 Daniel launched the direct programme — member rate, list building, a three-year target path — but the path had been building slower than planned: the list campaigns, the newsletter segmentation and the refinements to the hotel’s own booking surface kept losing their hours, because the routine ate the week. Now there is time. Let’s count the programme’s yield with lesson 49’s net logic — and here comes that lesson’s most important discipline: separating the gross number from the net one. The starting point is familiar: 65% OTA, 20% direct, 15% corporate, ~21,600 sold room nights a year. The first step is shifting +5 percentage points from the OTA to the closed member rate: ~1,080 nights. In gross terms that looks temptingly large — ~20 EUR of commission per night stays in the house, suggesting ~21,800 EUR a year. But the shifting tool is the member discount, and it has a price: a Booking night nets 90.5 EUR, a member-rate direct night 93.5 — the genuine, net surplus is 3 EUR per night, i.e. 1,080 × 3 ≈ 3,200 EUR a year. At the end of the three-year target path (~15 pp) that is ~10,000 EUR a year — and crucially, it is a durable yield that renews every year, not a one-off campaign result; all of it counted without the repeat layer (the returning member guest’s value — lesson 49’s second layer). Spread over the 400 hours, the first step looks modest — until you notice that what you bought was not one year’s result but infrastructure. The freed time is not rest: it gets repriced — from a routine hourly rate to a strategic one.

The risk side: the price of a bad guardrail. And now the other pan of the scale, because this too is part of automation — and Daniel paid the tuition in the first half year. In November, a mid-sized concert was announced for a Saturday. The date never made it onto the exception list — a human omission, one missing row. The automation saw what its rules said it should see: slow early pickup (concert demand typically explodes late, in the final 2-3 weeks), and stepped down — entirely lawfully, above the guardrail floor. The compression-justified rate for the event would have been 165 EUR; the automation let the date drift to 128. By the time the weekly guardrail review surfaced it, 30 rooms had sold at an average of 128. The bill: 30 × (165 − 128) = 30 × 37 = 1,110 EUR — for a single missing row on the exception list.

Two lessons, and both belong to the spine of this chapter. One: the machine did not fail — the framework was incomplete. On L3, most errors are not “the model was wrong” but “the human drew the boundary wrong” — automation is exactly as good as its guardrail set. Two: the weekly 2 hours of guardrail review is not administration; it is the work that catches the 1,110 EUR errors — ~100 hours of supervision a year guards the whole of the hotel’s rate integrity, and with it the entire yield side of the L3 balance. Automation is not the abolition of attention — it is the relocation of attention.

What the RM becomes — three new roles

The opening question — “what remains as my job?” — can now be unpacked with numbers behind it. The role transformation has three layers, and each one adds; none subtracts.

1. From administrator to strategist

The 8 freed hours a week go where the real value of RM work always was: positioning and compset strategy (lesson 44), channel and guest-relationship building (lesson 49), the total revenue mindset (lesson 58) — beyond the room, the revenue space of events, the restaurant, the spa. Notice: routine rate-fiddling was never the value of RM work — market knowledge and the decision were; administration merely took the time away from them. In this sense automation does not take from the profession — it gives the profession back to itself: every hour moves up the value ladder.

2. Quality control — guardrail design as a new skill

The most important new skill of L3 operation is framework design, and it is genuine professional work, not settings-fiddling. The question list Daniel walks through in the quarterly framework review: What is the maximum daily rate step — and does it differ by season? Which periods are exceptions — event days, priority weekends, group-heavy weeks — and who updates the list, when, from what source? (The November 1,110 EUR was lost exactly here.) When should the machine escalate to a human — below what confidence threshold, at what combination of signals? A guardrail is not a set-once configuration but a living system that ages together with the event calendar and the season — without maintenance it quietly goes stale, and a stale framework produces lawful errors.

The other half of quality control is measurement, and here every earlier tool goes back to work: override discipline (lesson 56 — override only on information from outside the model, logged), experimental evaluation (lesson 61 — who won, you or the machine, by situation type), and decision tracking (lesson 64 — every framework change is a decision, with an outcome). The more the machine executes, the more valuable the RM who can hold it to account.

3. Exception handler — the hardest cases stay with you

The third layer is the least comfortable, so let’s say it upfront. Lesson 51’s three failure points — the unprecedented event, the regime change, the context outside the objective function — do not disappear at any level of automation. The machine runs normal operations; switching to crisis mode (lesson 62), pricing in the new 200-room competitor, calling the structural market shift — “the world works differently from now on” — remain human calls, because they need knowledge from outside the database. More than that: at the first suspicion of a regime change, the correct first move is to rein the automation in — temporarily stepping back from L3 to L2 until the new pattern has built up in the data.

And here is a paradox worth knowing in advance: the better the automation, the more it is only the hard cases that reach you. Your average day gets easier — your hardest day stays just as hard, except now it makes up the larger share of your work. That is not a demotion; it is a promotion in level: the decision density of your working hours rises. This is exactly why the RM who owns the toolkit of lessons 33–65 is worth more — exception handling takes the whole profession, because an exception is, by definition, the thing there is no recipe for.

Adam’s question

At the quarterly leadership meeting, Adam asks the question every GM will ask in the coming years: “If the machine prices most of the days — do we now need less RM?”

Daniel’s answer has two parts. The first is his own time balance: the 12-hour routine has indeed disappeared — but in its place there is no idle time; there is the direct programme (a net ~3,200 EUR a year at the first step, ~10,000 at the end of the target path), the framework supervision (which got sharpened in November, after 1,110 EUR of tuition), and the exception handling that takes his whole profession. Not less RM — a different RM.

The second part is the competitive-edge argument, and it is the deeper one. Automation by itself does not differentiate: within a few years everyone will have it — every member of the compset. What differentiates is who tunes the human–machine pair better: whose guardrails are more precise, who maintains the exception list, who measures their overrides with discipline, who notices the regime change two weeks earlier, and who feeds their market knowledge back into the machine’s context (lesson 51). Two hotels with the same tool will produce results that differ on the order of a hundred thousand EUR a year — the difference will not be in the software, but in the chair next to it.

The human RM’s lasting edge

To close, let’s name what cannot be mechanised at any foreseeable level of automation — the future-proofing test of lesson 51’s frame. Five items:

1. Context. Living in the market — the half-sentence dropped at a hoteliers’ breakfast, the news of a cancelled flight route, the rumour of a competitor’s renovation. What never becomes data does not exist for the machine; in the agent era too, the decisive input remains what you carry into the system.
2. Responsibility. The machine bears no consequences. Towards the owner, the team and the guest, a human answers for the decision — including for the framework of the decisions entrusted to the machine. “The automation set it” is exactly as much of a non-justification as “the machine said so” was in lesson 51.
3. Relationships. The key account, the group organiser, the owner’s trust (lesson 60’s storytelling), the team’s shared decision workshop (lesson 65) — trust capital is built between people, and a surprisingly large share of RM work’s long-term yield rests on it.
4. Ethical and brand judgement. Do we sell at five times the rate in a force majeure situation? Do we raise a loyal guest’s rate because the model says they can bear it? The guardian of the values outside the objective function is the human — on L4 even more than today, because there the formulation of the objective function is itself your work.
5. The ability to ask. The machine answers better and better — but what is worth asking is the concentrate of domain knowledge. Behind a good question sit the understanding of the booking curve (lesson 37), pickup diagnosis (lesson 50), data-quality suspicion (lesson 57), elastic thinking (lesson 36). The agent era does not devalue this knowledge — it becomes the difference between the one who puts the machine to work and the one who merely watches it.

So the RM profession does not disappear: it shifts upward. The administrator layer drains away from beneath it — the strategist, supervisor and exception-handler layers appreciate. Whoever owns this course’s toolkit AND learns to put the machine to work — to frame it, hold it to account, feed it — is worth more on the market than their predecessors ever were. The next and final lesson (67) puts all of this to the test in a single story: a tough year in which forecasting, pricing, crisis, automation and human judgement are all on the field at once.

Key takeaways

• Automation is not yes-or-no but a five-level ladder (L0 measures → L1 signals → L2 recommends → L3 executes → L4 tunes). Today’s practice runs in the L2–L3 band: approving rate recommendations and guardrail-bounded automated pricing are existing modes of operation; L4 — the machine tuning its own rules — is the direction of the coming years.
• Keep “already today” and “likely coming” apart. Already today: specialised agents (forecast advisor, pricing analyst) in Pulse Chat, scheduled routines, guardrail-bounded automated pricing. Likely: agent–agent coordination, autonomous campaign management, natural-language goal programming — an industry direction, not a product calendar.
• The automation-level decision is a computable decision. Daniel’s L2→L3 step: weekly pricing work from 12 to 4 hours, ~400 hours a year freed and invested in lesson 49’s direct target path — a net ~3,200 EUR a year at the first step, ~10,000 at the end of the path (the gross commission figure looks like a multiple of that, but the price of the shift is the member discount). The other pan: one missing row on the exception list cost 1,110 EUR — on L3, most errors are not machine mistakes but bad human frameworks. Automation is the relocation of attention, not its abolition.
• The RM’s three new roles: strategist, quality control, exception handler. The freed time moves up the value ladder (positioning, direct, total revenue); guardrail design and calibration measurement become core skills; regime changes and unprecedented cases — the hardest work — stay with the human, and the decision density of the working day rises.
• Not less RM — a different RM. Within a few years everyone will have the automation; the edge lies in who tunes the human–machine pair better. What stays durably human: context, responsibility, relationships, ethical and brand judgement — and the ability to ask.