Let agents draft, match and chase. Keep judgment, approvals and anything that moves money with people. Every agent action should be visible, reversible and permissioned.
What an agent is, in ERP terms
An AI agent is software that uses a language model to decide on a sequence of steps and then carries them out using tools. In an ERP, the tools are things like "read this supplier's open purchase orders", "create a draft purchase invoice" or "look up the stock for this item". The model reads the situation, picks a tool, looks at the result and decides what to do next.
That is genuinely different from traditional automation, which follows fixed rules. It is also why agents need more care. A rule does the same thing every time. A model can make a reasonable-looking choice that is wrong, and it can do so confidently.
Where agents help
Agents are most useful on work that is repetitive, text-heavy and easy to check. Good examples:
- Drafting documents from unstructured input. Reading a supplier invoice PDF or a customer's emailed order and preparing a draft purchase invoice or sales order for a person to review.
- Matching. Suggesting which bank transactions correspond to which invoices, or which receipt goes with which purchase order, with the reasons shown.
- Chasing. Drafting reminders for overdue invoices or late supplier deliveries, using the actual balances and dates from the ledger.
- Summarizing. Turning a week of support tickets or production downtime records into a short note on what happened and what keeps recurring.
In each case, the agent prepares and a person decides. The time saved comes from not having to type, search and cross-reference, not from removing the review.
Where they do not
Agents are a poor fit for decisions that need judgment about people, relationships or risk: whether to extend credit to a struggling customer, whether to switch a long-standing supplier, how to handle a quality failure. They lack context that lives in people's heads and in conversations the system never saw.
They are also unreliable at arithmetic done in the model itself. Totals, taxes, valuations and balances should always come from the ERP's own calculations, never from a model adding numbers in text. And they should not act on records they cannot see: if a document is missing, an agent may fill the gap with something plausible rather than say it does not know.
Finally, agents are weak at genuinely new situations. They do well on patterns seen many times and badly on the exception that has never happened before, which is often the one that matters most.
Cost is a practical limit too. Every agent step calls a language model, and long chains of reasoning over large documents add up. Agents that re-read the same records repeatedly, or run on every transaction whether or not anything needs doing, can cost more than the time they save. Good designs trigger agents on specific events, give them only the records they need, and stop when the task is done.
Guardrails that should be non-negotiable
- Draft by default. An agent creates drafts; a person submits. Nothing posts to the ledgers or leaves the building without approval.
- Same permissions as the user. An agent working for someone can see and do only what that person could.
- Everything logged. Each agent action records what it read, what it did and why, in the same audit trail as human actions.
- Easy to undo. Because agents work through normal documents, a wrong draft is simply deleted and a wrong submission is cancelled like any other.
- Numbers from the system. The agent quotes figures the ERP calculated, with links to the records.
How Sundial is designing for this
Sundial is built to be AI-native, which for us means the platform is designed for agents to work alongside the team under exactly these guardrails. Agent features are on our roadmap and arriving in stages; we will describe each one plainly when it is ready, including what it does not do. A typical example of the direction: an agent drafts the purchase order from a reorder signal and waits for your approval.
We think the honest pitch is a smaller one than much of the market is making. Agents will not run your operations. Used carefully, they can take a share of the typing, searching and chasing off your team's plate.
How to get ready
The best preparation is unglamorous: clean master data, clear roles and permissions, consistent document flows and agreed definitions. An agent working on a tidy ERP is useful. An agent working on a messy one produces confident drafts built on bad data.





