The capability gap between these products is smaller than the marketing suggests. The deployment gap is not — and it is what usually decides which one a company can actually buy.
The first two decide whether it survives a security review. The next two decide whether it survives contact with your actual process.
Vendor tenancy, your account, or a split. This decides the security review and, in most enterprises, the timeline.
Ask about prompts, documents and logs separately. One answer covering all three is not an answer.
A platform hands your team a builder. A service arrives built. The second costs more and needs nobody assigned to it.
Every process has the tenth of cases that do not fit. Whether it stops, guesses, or escalates with context is the product.
Ask for day one and for six months. A product with no answer for the second has not been run anywhere for that long.
Agents drift when the systems around them change. Somebody owns that, and it is worth knowing who before you sign.
Where the audit trail lands matters as much as what it contains — a vendor dashboard is not your audit trail.
A single-purpose agent can be judged on whether the job got done. A general assistant cannot be judged at all.
The four we are asked for most. Three more — operations, engineering and testing — sit on the roles index.
Slack
Google Chat
Telegram
Teams
Gmail
Outlook
ZoomCCallsCCRMBBrowserEERPTriage, the customer’s whole history in one read, and a drafted resolution waiting for the agent.
See the support coworker →Campaign operations, reporting that assembles itself, and drafts written from your own material.
See the marketing coworker →Account research before the call, the follow-up after it, and a CRM that stays true in between.
See the sales coworker →Matching, chasing, and the close pack — with every exception brought to you already evidenced.
See the finance coworker →This is the fact that decides most enterprise agent purchases, and it is settled by architecture rather than by configuration.
A single job can be judged at the end of a pilot. Five half-built agents cannot be judged at all
The reconciliation, the triage, the copying between systems. Adoption is free when nobody was enjoying it
Clean data separates no products at all. The messy week is the only useful evaluation
Not a product list — the jobs companies describe when they first ask about agents. Every one of them is somebody’s worst afternoon.
The questions that come up before a shortlist exists.
Software that takes a business process and runs it, rather than answering questions about it: reading what arrived, gathering context from the systems involved, taking the action, and reporting what it did. The label covers a wide range of products, from a chat interface with tool access to something that owns an entire back-office job.
Mostly deployment and ownership rather than capability. Agent platforms are hosted by the vendor, so your data travels to them, and they hand your team a builder. An AI coworker is deployed inside your own cloud account and built for you around one job, so it arrives already knowing how your process works. There is a full comparison on the compare page.
The ones where the work is high-volume, rule-shaped but not rule-bound, and spread across several systems — accounts payable, support queues, CRM upkeep, reconciliation, operational intake. Work that is mostly judgement, mostly relationship, or mostly physical is a poor fit, and it is worth saying so before a pilot rather than after.
The software rarely is; the procurement is. Enterprise buyers ask where it runs, under whose IAM, against which logging, and what happens at renewal — which is why deployment architecture tends to decide these purchases more than capability does.
One. A single job can be judged on whether it got done, which means the pilot has an answer at the end of it. Companies that start with a platform and five half-built agents usually cannot tell you six months later whether any of them worked.
It varies enough by job and by system count that a published number would be misleading, and we do not publish pricing while we are working with design partners. The comparison that actually decides it is against what the job costs you today, which is a conversation rather than a page.
“Every one of these can read a queue. The difference is what happens at the case nobody described.”
Same workflow, exceptions included.
clean demos separate nothing✓ Ask where the software runsBring the process you were going to point an agent at. We will tell you honestly whether this is the right shape for it.
Get a demo