An assistant answers the question you ask and leaves the next step with you. An agent takes an objective, works through the steps, acts in your systems and records what it did, inside rules you set.
One platform carries all three: ChatLima answers, the one built agent acts, and one rule store and one audit trail would govern every agent built after it.
L. Tan asks ChatLima which open lines she can still cancel this week and what they are worth: 38 lines, $412,000, in 4.2 seconds, with the query on show. The next step stays with her.
The PO cancellation agent applies the policy to the same 38 lines and drafts 29 cancellations worth $187,300 for D. Okafor to approve; the 9 lines above $25,000 go to D. Okafor as buyer, for his judgement, with the reason for each.
Ten task agent roles are designed on the blueprint, PO cancellation among them; one runs. Built from the same parts, they would share one rule store, one audit trail and one level setting each, so the tenth would be governed the way the first is.
The difference sits in what comes out. An assistant produces something for a person to use: an answer, a summary, a draft. An agent produces an action in a system of record, or a draft of one for a named person to approve, and it keeps the facts it acted on. The third shape is a design: as roles are built, it would hold their objectives, rules and limits in one place, so a manager would read what ran and who approved it in one report. Today that report counts one agent.
Showcase estate for the Monday’s values; the mechanics are the product’s.
This is the agent the whole story follows: the next chapters set its level, show its settings and build it in a day.
“Cancel open purchase order lines where projected excess is above eight weeks of demand, the line is inside its cancellation window, the line value is under $25,000 and the supplier class is standard. For lines of $25,000 and above, draft the cancellation and ask the buyer.”
L. Tan wrote it on the Friday; Juno compiled it into a planning table, and version 3 ran on the Monday. Showcase estate.
Objective. One sentence says what the agent works towards, and every run’s outcome is checked against it. Here: cancel purchase order lines that stand above the need, inside their cancellation window. The planner, L. Tan, owns it.
Data scope. What the agent may read, under the access rights it carries: purchase order lines, schedule lines, demand, stock and supplier class at plants 1210 and 1220, read in place from SAP S/4HANA and Oracle Fusion. The data lead sets it.
Rules. The policy in plain language and the planning table it compiles to, held as a version: version 1 the first draft, version 2 after L. Tan’s edit, version 3 the one that ran on the Monday. A change means a new version, simulated first.
Trigger. A schedule, an interval, an event in a source, or a person asking for a run. This agent runs at 06:00 each working day, after the planning run, so its drafts are waiting when L. Tan opens her list at 07:40.
Actions. A short list of what the agent may do, drawn from a fixed vocabulary: draft a cancellation, send the change to the ERP, open a query to the buyer. Everything outside the list stays outside its reach.
Limits. Checks that run before every action: a value threshold of $25,000, a daily action budget of 60 and a confidence floor of 0.80. On the Monday the threshold split the 38 lines into 29 under it and 9 above.
Escalation. The named person or desk that receives what the agent hands over, with the reason. Above the threshold, a line goes to the buyer; a failed write-back goes to the owner. The 9 lines needing judgement reached D. Okafor this way.
Tests. Recorded cases with the outcome people chose, used to release the agent and later to raise its level. Before release the policy ran over last quarter: 184 lines matched, worth $1.9M, 31 routed to the buyer, read before J. Ridley released version 3.
The same four steps serve a measure you only watch and a decision you let an agent make.
The rule says forecast error stays at or under 10% for A-class materials. On controller board 7741-X the error reaches 14%, so the planner is alerted with the last six periods beside the figure.
A forecasting agent configured from the same eight parts would test the other forecast methods on the same history and apply the best one inside its limits; above them the planner chooses, and the choice is kept for the next cycle. Illustrative values; the method is the product’s.
The four steps are the eight parts at work: the rule to watch is the rules part, the right person is the escalation route, the limits decide whether the agent resolves or engages, and the execution is an action from its list.
Each test has a screen behind it in the reference pages, so an analyst can verify the answer in the product.
Gartner counts about 130 vendors with genuine agentic capability among the thousands describing their products that way, and expects over 40% of agentic AI projects to be cancelled by the end of 2027 (Gartner, 25 June 2025). Five plain tests let a buyer check any vendor’s claim, this one included.
Showcase estate for the figures in the five rows; the mechanics are the product’s.
Limits and escalation are two of the eight parts, and both depend on how much the agent may do alone.
The $25,000 threshold says where the agent’s own reach ends; the escalation route says who takes over beyond it. Both answer one prior question: how much this agent may do on its own. Set that differently and both parts change with it.
Chapter 1 named unclear limits as one of the four reasons earlier AI stalled. The eight parts give the limits a place to live. They leave open who sets the reach, on what evidence, and whether it can differ between a cancellation agent and a forecasting agent. That is the next chapter.
A workflow follows a fixed path that someone coded. An agent takes an objective and versioned rules, works out which steps fit the case in front of it, and reports what it did against that objective.
The rules engine decides. The language model reads the planner’s intent at design time, explains answers and proposes rules it finds in past decisions. Every action resolves to a rule version and a named approver.
The agent takes the search, the valuation and the draft. The planner keeps the exceptions, the supplier relationships and the judgement calls, which is where the 9 lines of chapter 1 sit.
Each one builds on the one before it.
Reads your rows in place.
Connect and move ↗One record per real thing.
Master and govern ↗Apps for every decision.
Operate and decide ↗Asks and acts, at your level.
Ask and act ↗Bring two systems and one decision your planners make every week. In the first week we register the data, confirm the mapping with your team, answer your questions in ChatLima and run the first agent at the level you choose.