Last week, one of our clients sent us an email. Two lines. They’d like to start doing their AI more end-to-end, because what they have has stopped being enough. And could we share our latest experience with UiPath Maestro?
Some context. This is a big international company. We’ve been automating with them for years, and they have dozens of processes in production. Last year, they told us they’d started building AI agents. Proofs of concept, on different tools, in their private cloud. We said what we always say: lovely, and at some point you’ll need something that orchestrates all of it. And they told us, quite reasonably, that it felt expensive. Especially next to simply paying for tokens. Maybe you’ve said a version of it yourself: “We’ll just pay for the tokens.”
Here’s the truth: they weren’t wrong. Not then. A handful of agents on the cloud you already pay for really is the cheapest option on the table. “We’ll deal with orchestration when we need it” is a perfectly sensible sentence. The only problem is that “when we need it” never shows up as a date in your calendar. It shows up as an email. A year later. With a dozen or so agents behind it.
Funnily enough, a few weeks before that email, a Gartner report landed in my inbox: the Magic Quadrant for something they call BOAT. Sent, as these reports usually are, by a company that sells one. Now, I have a rule for reports like that: enjoy the fish, but never buy the boat from the guy who caught it. And this time, the report is literally about the boat.
So today we read it anyway. Not as a shopping list, but as a checklist for the day an orchestration layer stops being too expensive and becomes cheaper than the mess. By the end, you’ll know how to tell which side of that line you’re on, and what to read in this report before anyone sends you a quote.
Tokens vs orchestration: the hidden cost of AI agents
Let’s start with that reasoning, “we’ll just pay for the tokens”, because it deserves better than a smug podcast host. It’s a good argument. It’s a CFO’s argument.
Tokens are pay-as-you-go. You pay per thought: an agent reads a document, you get a bill for a few cents. And orchestration? It isn’t a flat fee either. On the platform this client asked about, UiPath Maestro, you pay per execution of an orchestrated flow, on top of the tokens. So next to “just the tokens”, orchestration will always look like exactly what it is: a second line on every single run. A surcharge. If I’d been their CFO in the third quarter of last year, looking at a few PoCs, I’d have said no to the surcharge too.
But look at what the token bill doesn’t include.
It pays for an agent thinking. It doesn’t pay for anyone knowing how many agents you have. Who owns each one. What each one costs a month. Which of your existing automations it touches. Who steps in when it gets something wrong. And who explains all of that to an auditor. At three agents, that’s one person’s memory. At a dozen, across two clouds, sitting next to dozens of existing automations, that one person becomes your orchestration platform: part-time, unversioned, and impossible to upgrade.
That’s the trap inside “when we need it.” The surcharge is visible. It’s on the invoice, with a number on it. The cost of not having it isn’t on any invoice. It’s spread across a dozen people’s working weeks. So the line on the invoice always loses the argument, right up until the cost that’s on no invoice arrives as an email that says “end-to-end.”
Which is where Gartner comes in. The report is the Magic Quadrant for Business Orchestration and Automation Technologies, BOAT for short, published this September, with twenty vendors assessed. And before I quote a single line of it, full disclosure: we’re a UiPath diamond partner, UiPath sits in the Leaders corner of this report, and that client asked about a UiPath product. So when I say something nice about UiPath, even with our nobullshit policy, feel free to take it with a grain of salt.
Anyway, Gartner’s definition, translated from analyst speak: BOAT is one platform that orchestrates processes across all your systems, with every kind of worker in them: robots, APIs, AI agents and people. End-to-end, in other words. The word from that email.
Gartner gives BOAT a list of mandatory features and a list of optional ones.
Two features on those lists: RPA, the classic software robot, and “computer use”, meaning AI that operates a screen the way a human would. One of them is mandatory. One is optional.
Pick one.
They say RPA is optional. Computer use is mandatory.
And here I’ll disagree with Gartner, on air, with my name on it. Today, computer use is RPA, just worse. Slower, more expensive, and a lot less reliable. That may change in time, because these models get better every quarter. But if you’re choosing a platform this year, the robot that clicks the same button the same way every single time, I don’t think it’s an optional part.
Where Gartner is completely right though, is the part that isn’t about the worker at all. To even get into this report, a platform had to treat human-in-the-loop as, in Gartner’s words, a “first-class capability.” Step back from the feature list and that’s the real message: this category isn’t defined by who does the task. Robots, agents, APIs, humans, take your pick. What’s not negotiable is the layer above them. Who sequences the work, who governs it, who watches it, who hands it to a human when it gets hard.
Gartner calls that layer the control plane, and the report lists what buyers now expect from it: central policy enforcement, observability, life cycle management, auditability, resilience. Now read that list again as an email: “we’d like to do this end-to-end, and what we have has stopped being enough.” Same list. Written by someone who’s tired.
And I can hear the objection: “Andrzej, this is hyperautomation with a new haircut. Gartner renames the same thing every few years.” Fair. The problem isn’t new. Around 2019, every department bought itself a bot, and nobody knew how many there were. What’s new is what these workers do. A bot nobody’s watching repeats the same mistake. An agent nobody’s watching invents a new kind of mistake. Same problem, higher stakes.
And if you recognized yourself in our client, relax. Gartner’s own forecast is that by 2030, seventy percent of enterprises will have moved to a consolidated platform like this. Today, by Gartner’s estimate, it’s ten percent. So if your agents live wherever they were born, you’re in the ninety percent. You’re normal. For now.
So here’s the one thing to take from this chapter. Tokens are the price of an agent thinking. Orchestration is the price of dozens of agents being accountable. The question was never which one is cheaper. It’s: when will going without orchestration start costing you more than paying for it.
And when that day comes, and a report like this lands in your inbox with somebody’s logo circled in the top-right corner, the first thing to do is stop looking at the circle. Let’s talk about how to read the dot.
Gartner’s BOAT report: how to choose an orchestration platform
So the email arrives. Somebody in your company finally says the word “orchestration” out loud, and the next sentence is almost always the same: “Fine. Let’s shortlist the Leaders and pick one.” The vendors in the top-right corner. Safe, defensible, done.
I understand the instinct. And there’s always someone who adds: “Andrzej, I don’t have time for this. The report says it’s a fifty-seven-minute read. Just tell me who’s in the top-right.” Fair. So let me tell you what the top-right actually measures, and save you fifty-two of those minutes.
The dot sits on two axes: ability to execute and completeness of vision. Ability to execute is built from sensible things: the product, the vendor’s finances, customer experience. But beware, one of its criteria is called Marketing Execution, and it means exactly what it sounds like: how visible the vendor is, how clear its messaging is, how strong its brand is. Nothing wrong with that. A vendor that can’t sell might not be around in five years. But it means the dot answers the question “is this a strong vendor?” It does not answer “is this the right platform for us?” And Gartner says so itself, in plain words: choose on strategic fit, architecture, governance and ecosystem, not on individual AI features or marketing claims.
So where’s the useful part? In the small print, after the picture. Every vendor gets a list of Strengths and a list of Cautions. Skip the strengths. The vendor will tell you those, at length, for free. Read the cautions.
Quick game, then. There are four Leaders this year: Appian, Pega, ServiceNow and UiPath. One kind of caution shows up for all four.
What is it?
Cost.
In four different wordings: pricing, budgeting complexity, commercial complexity, skills and cost commitment. For three of these vendors, it’s about predicting the bill: multiple licensing models, mixes of subscriptions and usage-based fees, consumption metrics that make an enterprise-wide total hard to forecast. For the fourth, Pega, it’s the price of the specialist skills you’ll need as you scale. And to be fair, Gartner actually praises Pega’s agent pricing for being easier to forecast. And remember where we started: our client walked away from orchestration because next to tokens it looked like a surcharge. It turns out that even in the top-right corner, it’s hard to know in advance what that surcharge will add up to. That’s not a reason to avoid the category. It’s the first thing you make a vendor prove.
Here are UiPath’s cautions, out loud. There are three. One: pricing. A broad portfolio and multiple licensing models make sizing and cost forecasting hard. Two: process orchestration in UiPath is still maturing compared to vendors who’ve been doing process management for much longer. Some advanced constructs aren’t supported yet, and Gartner tells buyers with complex, exception-heavy processes to validate them in a proof of concept. Three: UiPath’s customer base is deeply rooted in RPA, and moving to orchestration means changing governance, operating model and skills, not just the licence. I agree with the first two. On the third, I’m on the fence, and I’ll tell you why in a moment.
Now, the second thing hiding in the small print. The headline everybody quotes is consolidation: seventy percent by 2030. But read further and the same report says that no single vendor today is best-in-class across every part of this. Companies might end up with more than one BOAT platform. And the very first recommendation is to rationalize what you already have and reduce sprawl before you expand. So “consolidate” doesn’t mean “move everything under one logo.” It means one layer that knows what’s running, who owns it, and what it costs, even if the agents underneath were born on different clouds. For our client, that changes the question. It’s not “where do we migrate a dozen agents to?” It’s “what sits above them?”
So here’s this chapter’s takeaway. Don’t buy the dot. Take the cautions for every vendor on your shortlist and turn each one into a test. Cost forecasting? “Here’s our volume. Show me the bill for a year.” Orchestration maturity? “Here’s our nastiest process, with all its exceptions. Build it in the proof of concept.” Yes, a proof of concept, again. But this time you’re testing the layer above the agents, not one more agent. Which brings us back to that email, and the question it’s really asking. How do you know you’ve crossed the line?
3 signs you need an AI agent orchestration platform
Here’s the belief that carried our client from that first conversation to last week’s email, and it’s the most reasonable belief in this whole episode: “We’ll notice when we need it.”
You won’t. Not because you’re not paying attention, but because of where the cost sits. The surcharge is on an invoice, so somebody looks at it every month. The cost of not orchestrating is on no invoice, so nobody looks at it at all. Not until a business owner asks for something end-to-end and the answer is “we’d have to rebuild most of it.” So don’t wait to notice. Decide in advance what noticing looks like. Here are three signals. Hit one, and you’re close. Hit two, and you’ve crossed.
Signal one: the count. Ask someone on your team a simple question. How many agents and automations are running right now, who owns each one, and what does each one cost a month? If the answer comes back in a few minutes, you’re fine. If it comes back as “let me check with a few people and get back to you by the end of the week”, that’s your signal. You don’t have a platform. You have a scavenger hunt.
Signal two: the handoff. The first time a request lands that needs two of your agents to pass work to each other, you’ve stopped building agents and started building a process. One agent reads the request, another checks it against a policy, an existing automation enters it into a system, and a human approves the odd ones. A process like that needs something that remembers where every case is, keeps it moving, and hands it to a person when it gets stuck. That’s what “end-to-end” meant in that email.
Signal three: the spreadsheet. Somebody, every month, logs into several consoles, two cloud billing pages and a ticket queue, and builds a status report by hand. That person is doing, manually, beautifully and at great expense, exactly the job Gartner’s control plane exists to do. Observability, auditability, life cycle management, here built in a spreadsheet.
Now, the caution I promised to come back to: the one about UiPath’s customers being rooted in RPA. Here’s why I’m on the fence. A good RPA Center of Excellence already has half of what agents need: an owner for every automation, release discipline, monitoring, exception queues, and people who know that a process in production is a living thing. That half transfers. The other half really is new. You’re testing something that gives a slightly different answer every time, and you’re deciding where a human has to sign off. So yes, the team has to grow. But if you’ve been running automations in production for years, like our client has, you’re not starting from scratch. You have a head start.
And what does getting this wrong cost? I’ll be honest: for our client, I don’t know yet. Neither do they. That’s the nature of a cost that sits on no invoice. You only find out how big it was when you start paying it off. What I do know is the direction. Plugging three agents into an orchestration layer is a project. Retrofitting a dozen, born on different clouds and each built its own way, is a programme. Every quarter of “we’ll notice” moves you from the first to the second.
So here’s the takeaway. Three signals: the count, the handoff, the spreadsheet. Hit two of them, and orchestration has stopped being too expensive. You’re already paying for it. Just not on the invoice.
The answer: AI agent orchestration in 3 sentences
So here’s the whole episode in three sentences, the version for your steering committee.
One: tokens pay for an agent thinking. Orchestration pays for a dozen agents being accountable, and next to tokens it will always look like a surcharge.
Two: read Gartner’s BOAT report for its cautions, not its dot. Every Leader got a cost warning, so cost and orchestration maturity are the first things any vendor has to prove.
Three: don’t wait to notice. Watch the count, the handoff and the spreadsheet. Hit two of the three, and it may be you do need a bigger boat.
Exercise for tomorrow: list every AI agent, its owner and its cost
Your homework is one page. Tomorrow, write down every agent and every automation running in your company. Next to each one, two things: who owns it, and what it costs a month. Not from memory. Ask.
If the page is done by lunch, you’re still on the cheap side of the line, and you can look at orchestration calmly, on your own schedule. If it takes all week, you already know what you need. And you’ve just written the first page of your business case.
