How to choose an AI agent platform: 7 questions that decide it
Build or delegate settles it before any feature list. Plus the pricing two vendors no longer publish, and the platform with a shutdown date on it.
WorkAgent console
Delegate a task · watch it run · get the result
Pick a task to delegate
It works the task end to end and hands back .
| Company | Contact | Fit | |
|---|---|---|---|
Subject:
Research, lead gen, outreach, inbox and reporting, delegated and done.
Like what it did? Put WorkAgent on your real work.
The short answer: ask one question before you look at a single feature list. Do you want to build the agent, or delegate to it? Builders like Make ($9 a month), n8n (free self-hosted), Zapier Agents ($33.33) and Gumloop ($37) are cheap because you supply the labor. Delegates like Lindy ($49.99 to $199.99) and WorkAgent ($149 flat) cost more because the product does that work. Almost everyone unhappy with an AI agent platform bought one while wanting the other. Everything else on this page is downstream of that.
We checked all the pricing below on each vendor's own page in July 2026, and two things everyone repeats turned out to be wrong. Those are in here too.
1. Build or delegate? Settle this first
A builder gives you a canvas: triggers, steps, branches, tools. You design the flow, you test it, and you fix it when a vendor changes an API response at an inconvenient hour. This is genuinely good software and $9 a month is a bargain for what Make does. But be clear about the trade. The reason it costs $9 and not $149 is that you are the expensive part of the system.
A delegate takes a description of an outcome and works out the steps itself. You write "research the twenty largest competitors in this category and tell me how each one prices" and the brief comes back. No canvas, no triggers, no debugging. The trade is control: you cannot rewire a delegate the way you can rewire n8n, and if you have a genuinely idiosyncratic process only you understand, a builder will take you further.
Neither is better. But buying a builder because it was cheaper, when what you wanted was to stop doing the work, is the single most common mistake in this category. The test that settles it: are you willing to own a flow, forever, including the maintenance? If reading that sentence made you tired, you want a delegate.
2. How does it meter you, and does that unit mean anything?
Six different billing units are in play across the mainstream platforms: tasks, operations, credits, actions, activities and seats. They do not convert into each other, which makes headline prices close to meaningless.
Some real examples, all from vendor pages in July 2026. Relevance AI's published top-up rate is $80 per 1,000 actions. Zapier Agents Pro includes 1,500 activities a month for $400 a year. Manus gives 300 credits daily on the free tier and 4,000 a month on the $20 plan. Lindy does not publish its underlying usage metric at all, only that Pro is roughly three times Plus and Max roughly seven times. Structurely, in a different category, at least defines its unit precisely: one credit is one SMS reply, ten seconds of talk time, or two emails. That clarity is rare enough to be worth noticing.
The practical problem with metering is not overcharging, it is behavior change. Metered users ration. They stop handing the agent complex jobs, because complex jobs burn credits fastest, which means they stop handing it the jobs that were worth automating in the first place. An assistant you are afraid to use is not doing anything for you. This is why flat pricing deserves more weight in your comparison than the headline difference suggests, and it is the thing we are most opinionated about. If the low entry price genuinely matters more to you, that is a legitimate call: buy Make at $9 and accept that the building is now your job.
3. Is the price even published?
This one surprised us. Relevance AI's pricing page carries no dollar figures at all as of July 2026: just Enterprise tiers and five links to sales, following a repricing in September 2025. The $19, $199 and $599 tiers still being quoted across the web no longer exist. CrewAI is the same above its free tier. In real estate, Ylopo, Lofty and kvCORE all publish nothing.
Sales-led pricing is a legitimate way to sell enterprise software and it is not a red flag by itself. But it has two consequences for you as a buyer. First, comparison shopping requires a sales call, which is a real time cost you should count. Second, and more importantly, any number you read for those vendors in a third-party roundup is either stale or invented. This applies more broadly than people realize: we found blog posts quoting a $19.99 Lindy tier that is not on Lindy's pricing page either. Check the vendor. Never the listicle.
4. What is the platform's migration risk?
The live example: OpenAI's Agent Builder is scheduled to shut down on November 30, 2026, with developers pointed to the Agents SDK or Workspace Agents. If you are evaluating platforms this quarter and Agent Builder is on your list, that is not a footnote, it is a known migration cost with a date attached.
The general lesson matters more than the specific case. This category is two years old and consolidating. Ask any vendor what happens to your configuration if they get acquired or sunset a product. With a builder, you have real exposure: months of flows you designed, in a proprietary format. With a delegate, you have much less, because the "configuration" is a plain-English brief you could hand to a different tool or a person in an afternoon. That asymmetry rarely shows up in comparisons and it should.
5. Does it act unattended, or does it just suggest?
Ask this in the demo, in these words: can I give it a task, close my laptop, and come back to a finished result?
A remarkable amount of software marketed as an AI agent surfaces a suggestion and waits for your click. That is a faster you, which is useful, but it is not delegation and it should not be priced like it. The demo will not volunteer the distinction, because the demo is a person driving the tool while narrating. Make them walk away.
6. What does it do when it does not know?
This is the question that separates production-ready from demo-ready, and almost nobody asks it.
An agent that invents a plausible answer when the data is not there is worse than useless, because you cannot tell the invented parts from the real ones without redoing the work. The behavior you want is boring: it says it could not verify something, lists what it could not find, and cites sources you can click. In research work specifically, hallucinated or non-supporting citations are a documented problem across deep-research tools, so structural checkability is the only real defense. Ask a vendor for an example of their agent failing. If they cannot produce one, they have not run it on enough real work.
7. What is it replacing, in dollars?
Finish here, because it reframes everything above. Comparing AI tools to other AI tools is the wrong frame and it is why so many buyers stall.
The alternative to an agent is almost never another subscription. It is a person, a contractor, or the work not happening. A US virtual assistant runs $3,000 to $6,000 a month. A dedicated virtual executive assistant is $1,600 to $3,600 depending on hours. A junior hire is $45,000 to $70,000 a year plus payroll tax, benefits, equipment and management time. A US inside sales agent averages $69,398 a year.
Against those numbers, the difference between a $9 builder and a $149 delegate is noise, and the only question that matters is which one will actually get the work done. Optimizing the $140 while the $70,000 alternative sits unexamined is a strange way to make a decision, and it is extremely common.
Working through it in practice
Write down the three jobs you want handed off, specifically, in the language you would use briefing a new hire. Not "sales automation" but "find 60 dental practices in Ohio that recently posted an office manager job, research each, and draft a first email." Vague briefs are how you end up evaluating features instead of outcomes.
Then take those three jobs to two vendors from opposite ends: one builder, one delegate. Make them do job one live. You will learn more in twenty minutes than in a week of comparison tables, and the build-or-delegate question usually answers itself the moment you watch someone assemble a flow in front of you.
Worth knowing before you start: the number of agent products is growing fast enough that browsing what already exists for your specific job is often quicker than describing it to a builder from scratch. A surprising share of the flows people build by hand have already been built by someone else.
The short version
Answer build-or-delegate before anything else. Weight flat pricing more than the headline gap, because metering changes how you use the thing. Verify every price on the vendor's own page, since the roundups are badly out of date. Check migration risk, and note that Agent Builder has a shutdown date. Make it run unattended in the demo. Ask what it does when it does not know. And price the whole decision against the person you would otherwise hire, not against another tool.
The full nine-platform comparison with every price verified is on our AI agent platform and software page. If you are still unsure which side of the build-or-delegate line you are on, AI agent vs automation is the longer version of that question, and best AI agents for small business is the roundup where we do not rank ourselves first.