For thirty years, enterprise software has been a filing cabinet. People do the work; software remembers what was done. SAP captures the purchase. Salesforce captures the call. Workday captures the hire. The interface looks different every decade — green-screen, web, mobile, cloud — but the contract underneath has barely moved. Humans act. Software records.
That contract is about to break.
The thing that just happened to software
Around late 2024, three things became simultaneously true. Foundation models got reliable enough to take production actions, not just write summaries. Tool-use protocols stabilized so an LLM could call a real API and not lose its mind. And the unit economics flipped — running an always-on agent became cheaper per outcome than paying a human contractor for the same outcome.
None of these individually was an earthquake. Together, they redrew the map. For the first time, the cheapest, fastest, most accurate way to do a long list of business workflows is not "buy software, hire a person to use it." It is "buy software that does the workflow."
This is not the chatbot story. It's not "AI-powered features sprinkled into your CRM." Those are dressings on top of the old contract. The new contract is structural: software is the worker, not the workspace.
What software-as-a-worker actually looks like
An agent is just a piece of code that observes signals, decides on actions against a policy, executes those actions on systems it has access to, and asks for help when its confidence falls below a threshold. That's it. There is no magic. But the shape of an enterprise built around agents looks very different from one built around forms.
In the old model, your sales team had a list of accounts and ran a sequence of activities against each. The CRM remembered what they did. In the new model, an SDR Agent watches every signal across LinkedIn, news, hiring boards and your own product telemetry. It opens a sequence on every qualified account. It writes the first email. It books the meeting. Your AE walks in to close.
The same shift plays out across every department. A Replenishment Agent watches inventory and reorders before anyone asks. A Returns Agent reads condition reports and decides refund-vs-replace at the door. A Compliance Agent attaches REACH and CSRD evidence to every BOM at design time, not at audit time. A Routing Agent picks a carrier for every parcel and re-routes 38 of them around Mumbai rain at 7:24 AM.
If the cheapest, fastest, most accurate way to execute a workflow is a piece of always-on software, then "humans doing the workflow with software help" is a transitional state. The end state is software doing the workflow with human review when needed.
Why this can't be retrofitted
The instinct, if you sell enterprise software, is to sprinkle agents on top of what you have. Salesforce adds Agentforce. SAP adds Joule. Every booth at every conference next year will have "Agentic ____" on a sign.
The retrofit doesn't work, and the reason is data shape. The old applications were designed for humans to walk through one record at a time, in one module at a time. Each module has its own data store, its own permission model, its own audit trail. An agent that needs to read a customer, then a product, then a SKU, then a parcel, then a return — five different systems with five different schemas — burns most of its budget on integration.
Agents thrive on a graph, not a stack of filing cabinets. To get agentic productivity you need one operating system where customer, product, SKU, parcel and return are first-class connected entities. That's not a feature you can bolt on. It's a different starting point.
What changes for the people doing the work
Two things, mainly. First, the unit of work gets bigger. Instead of "follow up on these 40 leads today," an account executive owns "close these 8 deals this week." The agent does the 40 follow-ups overnight. Second, judgment gets more concentrated. Humans are pulled in exactly where the agent isn't sure — and only then. The shape of an enterprise day flips from 80% chasing and 20% deciding, to 20% chasing and 80% deciding.
This is good news for the people who like the deciding part. It is harder news for people whose value-add was the chasing. We're going to have a decade-long conversation about that, and I don't think anyone has the answer yet. But pretending the shift isn't happening doesn't help.
What we're building, and what we're not
Polluxa is not building one large agent. We ship a useful starter set of specialized ones — and a workshop so every team can build the agents their own workflows demand, across every Polluxa product. Specialization is what makes agents accountable. A "do everything" agent is a marketing demo. A Pricing Agent with a single KPI is an employee. The platform lets you create as many of those single-KPI employees as your business needs.
We're also not building a chat interface that pretends to be a CRM. The agents run quietly in the background, and the humans interact with a tight, fast, beautiful application that surfaces only what needs their attention. Most days, you don't talk to the agents. You watch them work.
The next ten years
By 2035, "enterprise software" will mean something different than it does today. The form-and-record paradigm will feel as quaint as paper ledgers. The companies that win will be the ones that picked their agentic suite early, kept their data graph clean, and trained their humans to be excellent at the small portion of work that can't be delegated. The companies that lose will be running 11 disconnected SaaS apps and wondering why their margins are flat.
That's the bet behind Polluxa. We think it's the biggest bet in software since the cloud. We'd love your help making it real.