KIBO AI: 5 Agentic Functions
In our last post we argued that commerce is not one problem, and that no single agent or model should try to solve all of it. That was the case for composability. This post picks up where it left off and gets specific: what are the actual jobs an agent does in commerce, and why does each one need its own kind of logic?
Not all commerce tasks are the same
Start with two requests that look superficially similar because both go to “an AI agent.”
The first: a shopper types “find me waterproof trail-running shoes under $120 in size 10 that are in stock.” The second: a merchandiser types “write a punchier short description for this backpack in our brand voice.”
Underneath, these have almost nothing in common. The first is a live conversation with a customer that has to feel human, and every claim in it (price, size, availability) has to be exactly right. The second is a focused authoring task, judged on tone and clarity, that ends with a draft dropped into a form for a person to approve. Different inputs, different logic, different definition of “done,” and honestly, different models suited to run them.
Multiply that across everyone who touches a commerce operation (shoppers, B2B buyers, customer service reps, merchandisers, pricing managers, fulfilment ops, developers) and the pattern becomes obvious. These are not variations of one task. They are distinct groups of tasks, each with its own logic.
The five functions
So we grouped them. At KIBO we defined five agentic functions, and every request an agent handles falls into one of them: Engage, Configure, Explain, Analyze, and Optimize. They run through one experience, but each one asks for something different underneath.
1. Engage
Engage is the agent in conversation: with a shopper, a B2B buyer, or a customer service rep. It is the function people actually see, so it needs the most capable model, the one that can hold a natural back-and-forth. But unlike a generic chatbot, every answer is anchored to live commerce data, and every fact has to be correct.
A shopper asks, “find me waterproof trail-running shoes under $120 in size 10 that are in stock.” The agent searches the catalogue, filters by attributes and real-time availability, ranks the results, and comes back with in-stock matches and a top pick, ready to add to the cart. A customer asks, “where is my order 100234, it still hasn’t arrived.” The agent looks up the order, reads the tracking, and gives a plain-language status with an ETA. A B2B buyer says, “reorder my standard monthly list and put it on PO #4471,” and the agent rebuilds the list at contract pricing, attached to the right purchase order, for review before submitting.
Engage even works behind the counter. A service rep can ask why an order hasn’t shipped, and the agent explains that it is sitting on a payment-review hold and offers to flag it. Same function, different persona.
2. Configure
Configure is the agent making changes to the system: promotions, product content, SEO fields, categories, search behaviour, pricing. Here the space of valid outcomes is narrow. There is a correct way to set up a promotion, and the agent’s job is to interpret the request and land on it.
The defining trait of Configure is restraint. The agent stages the change and leaves it unsaved for a human to review. Ask it to “create a 20% off sitewide ‘Summer Kickoff’ running July 15 to 31,” and it populates the discount form with those exact terms and stops, so the change is checked before it goes live. Ask it to “set SKU-1234 to $79.99 across all US sites,” and it prepares the update across the relevant price lists and waits for a save. Ask for a punchier product description in your brand voice, and it drafts the copy and drops it into the product form for approval.
That review step is not a limitation, it is the design. Configuration touches money and customer-facing content, so a human stays in the loop for anything that requires authorisation.
3. Explain
Explain answers the question “why did this happen?” It is the function that turns an opaque system into an accountable one. The logic here is investigative: interpret the question, pull the relevant data from several different places, trace the workflow, and explain it in plain terms. Explain is read-only. It never changes anything.
A fulfilment manager asks, “why was order 100503 routed to DC-EAST, and what else was considered?” The agent reads the routing decision log, ranks the alternatives, and explains that DC-EAST won on a combination of inventory availability and distance, while another location scored lower and a nearby store was missing one of the three items. A customer service rep asks why an order arrived in two boxes, and the agent walks through the split shipment item by item. A developer asks why a webhook stopped firing, and the agent checks the subscription, reads the delivery logs, and points to dozens of failed deliveries all returning errors from the customer’s own endpoint.
This is also where compliance meets daily operations. When a decision needs to be explained, evidenced, or contested, Explain is the function that produces the trace.
4. Analyze
Analyze is conventional data analysis, done conversationally. The agent runs a query over your commerce data and renders the answer as a chart or table. The logic is closer to a data analyst than a chat assistant: pick the right dataset, run the recipe, visualise the result.
“Show me revenue over the last six months as a trend,” and the agent renders a monthly line chart and notes the direction of travel. “What were our top 10 products by revenue in the last 30 days,” and it returns a ranked chart with the leaders called out. “Show me orders by status for the last 30 days,” and it produces a clean breakdown of fulfilled, in-process, backordered, and on-hold. The value is speed: questions that used to mean a ticket to the analytics team become a sentence in a chat.
5. Optimize
Optimize is the most advanced function, and the frontier of where agentic commerce is heading. The other four functions answer questions or carry out instructions. Optimize goes a step further: it analyses the current state of the system, models the impact of possible changes, and recommends how to move to a better one. It is the highest-effort task the agent takes on, and the one that leans hardest on reasoning.
Ask, “which promotions are underperforming and should I pause or restructure,” and instead of a chart, the agent returns a diagnosis: which promotions are failing, why (a segment mismatch here, thin margin there), and what to change. Ask, “where are we losing the most revenue in the checkout funnel,” and it identifies the biggest drop-off point, quantifies the revenue at stake, and proposes a specific test. Ask where to rebalance safety stock, and it models which adjustments free up sellable inventory without risking stockouts.
Optimize is deliberately the most cautious to ship, because a recommendation that moves money or inventory has to be trustworthy. Several of these capabilities are on our roadmap rather than live on day one, and that is by design. This is the tier where getting it right matters most.
Five functions, one experience
Line the five up and the argument from our last post makes itself. Engage wants a conversational model that sounds human. Configure wants a precise interpreter that knows when to stop and ask. Explain wants a patient investigator. Analyze wants a data engine. Optimize wants a reasoner. Five functions, five kinds of logic, and often five different best-fit models.
That is exactly why we built for composability, and why Bring Your Own Model sits underneath all of it. One experience on top, the right function and the right model underneath. This is KIBO AI.
