Two products shipped this week that vendors are calling AI coworkers and AI agents. Neither one is a coworker. Both are scoring models, and the difference matters more than the marketing wants it to.
What actually shipped
Dynamiks took its Model Context Protocol server to general availability, letting any assistant, Claude, ChatGPT, or Grok, sign into a Dynamiks account and pull deal data out of Salesforce or HubSpot. The company calls the underlying product the Quarterback, an AI coworker with, in its own words, a two-sided brain: one half language, one half a reinforcement-learning model trained on a sales simulator to score a deal’s Impact and Momentum. Sidetrade shipped an autonomous cash-collection agent, Aimie, that calls customers, qualifies disputed invoices, and adjusts its own strategy as it learns payment behavior, drawing on a data lake the company says holds nearly ten trillion dollars in B2B transactions.
Both are genuinely useful. Neither is a coworker in any sense a revenue leader should let pass unchallenged.
Advertisement
The counter-argument, taken seriously
The strongest defense of the coworker framing is that language is doing real work, not just marketing. A coworker, unlike a dashboard widget, takes a scoped piece of a job and owns the outcome without being re-supervised on every step. Aimie does not just flag an overdue invoice for a human to chase. It decides who to call, drafts the approach, and adjusts course as it learns. Dynamiks does not just surface a report. Nicolas Maquaire, the company’s CEO and co-founder, described the goal plainly: “Business tools can now use the Impact and the Momentum of every deal to decide or to trigger GTM flows.” Deciding and triggering are coworker verbs, not dashboard verbs, and a product that genuinely does both has earned some of the anthropomorphizing the industry keeps reaching for.
Why the framing still oversells the product
A coworker can explain why it made a call the way it did, in terms another person on the team could challenge. Neither Aimie nor the Quarterback can currently do that in a way a rep or a collections analyst could push back on before the action is taken. The Quarterback’s Impact and Momentum scores come out of a proprietary reinforcement-learning model trained on a sales simulator that Dynamiks built and controls. A rep asking an AI assistant which deals need attention today is not getting a second opinion from a colleague. They are getting a single number from a model whose training data, failure modes, and blind spots they cannot inspect, wrapped in language, coworker, Quarterback, that invites exactly the kind of unquestioning trust a human colleague has to earn over months of being right.
Aimie raises a sharper version of the same issue because the stakes are external, not internal. A collections agent that misjudges which invoices are genuinely disputed and calls a customer aggressively over a legitimate dispute does not just create an awkward Slack message. It damages a business relationship on the company’s behalf, autonomously, based on a model score nobody at the customer, and often nobody at Sidetrade’s client, reviewed before the call went out. Calling that a coworker frames a real accountability gap as a personnel decision instead of what it actually is: a software vendor’s model making judgment calls that used to require a person willing to sign their name to them.
What it means for the revenue leader buying either category
None of this is an argument against buying scoring-model automation. Chasing invoices and triaging pipeline attention are exactly the tasks that benefit from a system that never gets tired and never plays favorites. It is an argument against evaluating either purchase using the vocabulary the vendor supplies. Before a deal desk or a finance team signs a contract for a product marketed as a coworker or an agent, the right question is not “what can it do.” It is “when it is wrong, who finds out, and how.” If the answer is that a human only learns of an error after a customer complains or a forecast misses, the product is a scoring model wearing a job title, and it should be priced, governed, and audited like the automation it actually is, not trusted like the teammate its name is asking you to imagine.
Source: Dynamiks
