The Other Agent
Retail spent this year getting ready for the agent that shops for the customer. The harder question is who works for the shop when it arrives.
Ilyas Kurklu, CEO & Co-founder, Replenit
Friday 27 November, 07:40. A woman, a composite but you know her, asks her assistant to find running shoes for her brother: under a hundred euros, delivered before the holidays. She has bought from you for six years. The assistant reads the catalogues of a dozen retailers before her coffee is poured and comes back with three options. If your feed is in order, you are one of them.
At 07:41 your Black Friday campaign reaches her inbox. Thirty percent off, the same thirty percent everyone in your “Active, 12 months” segment received. She paid full price last week.
Two systems touched her in one minute. One worked for her and knew exactly what she wanted. The other worked for you and knew which list she was on.
That is the Black Friday retail has built for itself. Almost all the preparation this year has gone into the first system. This essay is about the second.
What the numbers actually say
Adobe forecasts that AI traffic to US retail sites will rise 130% this holiday season, and 95% on Black Friday itself. Salesforce predicts that one in three e-commerce sites will have its own shopper agent live by Cyber Week, and that 20% of holiday e-commerce traffic will come from AI chat agents.
Read those carefully before they reach your board. Adobe counts clicks from AI services and publishes no share of total traffic, so the 130% sits on a base nobody outside Adobe can see. Salesforce’s 20% includes bots running background tasks and competitors’ price scrapers.
Now look at how far shoppers will actually let go. In Deloitte’s holiday survey of 4,000 US consumers, 39% call generative AI their primary tool for at least one shopping task, and 16% would permit it to complete a purchase. Accenture, asking a stricter question of 4,048 US shoppers, finds 8% would let an agent buy fully on its own within guardrails they set. BCG puts the same idea at 4%.
These surveys ask different questions and should not be averaged. They point the same way. Discovery is moving to the machine. The purchase, for now, stays with the person. And the relationship, which is what decides the next purchase, stays with whoever bothers to hold it.
Salesforce has a number that should worry anyone calling this a traffic story. Global digital traffic grew 18% in the second quarter. Orders grew 1%.
What the feed buys you
Every holiday playbook this autumn gives the same advice: feed the agents. Clean product data, structured prices, live stock, a delivery promise a machine can verify. Salesforce’s recommendation is exactly this, and it is right. If an agent cannot read you, you do not exist in its answer. Do the work. Do it this week.
Then be honest about what the work buys.
A feed gets you into the comparison. It also makes the comparison easier. Everything in it, the price, the specification, the delivery date, is something your rival publishes in the same fields. A shopper’s agent is very good at this game and it has no sentiment. The better the whole industry gets at being readable, the more the three options on her screen look like one option at three prices.
In “The Third Memory” I argued that a shelf remembers what you stocked and cannot remember who walked past it. The feed is that shelf, opened to a reader who does not work for you.
A feed makes you comparable. Only what happens between purchases makes you preferred.
Preference is decided before the agent is asked
The most useful number of the season sits in Accenture’s survey. Three in four US shoppers want their agent to buy only from preferred or familiar brands, even when alternatives fit better.
Sit with what that means. The agent does not replace loyalty. It enforces it. Whatever she feels about you on the morning she asks is written into the instruction, and from there the machine carries it out with a discipline no customer has ever had. A brand on that short list is protected from the comparison. A brand off it is one more row in a table.
The same survey says 28% want to complete the purchase on the retailer’s own site or app after AI-assisted discovery, while 15% will go wherever the agent recommends. Accenture calls that 15% the swing segment. I would add that nobody is born into it. People arrive there when no brand has given them a reason to name one.
So the question moves from what her agent will find on 27 November to what happened on the other 364 days. Whether you noticed she was running out before she did. Whether you stayed quiet the week she ordered. Whether the offer she received last spring made sense for her, or told her again that she was a row in your table as well.
Why the shopper agent on your site is not the answer
The obvious response is the one Salesforce predicts for a third of the industry: launch your own agent. It is a good thing to do. Salesforce reports that retailers with shopper agents grew 2025 holiday sales 6.2% against 3.9% for those without, and shoppers prefer a brand’s own agent for returns, orders and loyalty.
Get Ilyas Kurklu’s stories in your inbox
Join Medium for free to get updates from this writer.
Look at where it stands, though. It is at the door. It waits for her to arrive, answers what she asks, and stops when she leaves. That is a concierge, and a concierge has nothing to do on the days she does not come in, which is almost all of them.
The work that decides preference happens when no one is asking anything. Someone has to look at her history every day and decide whether today is a day to speak, what would be worth saying, and whether the honest answer is nothing. That job has always existed. No team could do it, because no team can hold a few million people in its head, so the job was given to the segment.
The segment was a reasonable answer to a human limit. It was built for a world where the customer read your message, in your channel, on your schedule, and had no help comparing it to anything. She has help now. A blunt message used to cost you an unsubscribe. In a world where she can hand her next decision to a tireless comparer, it costs you the place on her list.
The agent that works for you
This is the job we built Maestro to do. Maestro is an AI CRM Manager: it remembers each customer, reasons about what her behavior most likely means, decides one thing for her today, or decides nothing, and owns the result. In “The Golden Decision” I walked through one morning inside it, so I will not repeat that here.
Nothing about this asks you to move your stack. Braze, Salesforce, Klaviyo, Bloomreach and Synerise stay where they are and carry the decisions, and your feed work stays useful. What changes is who sits in the middle.
At L’Occitane, post-purchase revenue rose 235%. I mention that number because of where it sits: after the sale, in exactly the stretch of the year this essay is about, when no agent is shopping and no customer is asking.
What I do not know
We have no data on how agent-referred shoppers behave across our clients at any useful scale, and I will not pretend otherwise. Maestro cannot see inside her assistant, cannot influence what it ranks, and does nothing for a retailer whose feed is broken. Memory will not rescue a product that is worse and more expensive. These surveys mostly record what American shoppers say they will do, and what people say about delegating is a poor guide to what they will do once it is easy.
The timing could also go against me in either direction. If delegation stalls at today’s levels, this essay was early, and the argument for deciding one customer at a time stands on its own. If it moves fast, and her agent starts reading your messages for her, a good deal of what works today will need rethinking, including by us.
What I am confident of is narrower. Whatever the share turns out to be, an agent acting for her will carry out a preference she already holds. Someone has to be earning that preference, and a calendar cannot.
On 10 December in Istanbul, at the AI in Retail Summit, we announce Maestro V2. If you run retail and want to argue with any of this in person, ask me for an invitation.
27 November, again
Back to her, and to the three options on her screen. She looks at them for four seconds and says the name of the shop she has used for six years. The assistant adjusts. It was never going to argue.
Whether she says your name was settled months ago, on days neither of you remembers.
The operator asks: can her agent read us?
The owner asks: when her agent asks her, will she say our name?
Ilyas Kurklu is the CEO and co-founder of Replenit, which builds the AI workforce for customer engagement. Its first role is Maestro, an AI CRM Manager. This essay follows “The Third Memory”.
Warsaw, 2026.
Sources:
Adobe, 2026 holiday forecast (28 September 2026);
Salesforce, 2026 Holiday Predictions (20 July 2026);
Deloitte, 2026 Holiday Retail Survey (October 2026);
Accenture, Holiday Shopping 2026,
US respondents; BCG, How Retailers Can Reinvent the Digital Flywheel for Agentic Commerce (October 2026).
Images: James Tissot, The Shop Girl (public domain); Marjory Collins for the Office of War Information (Library of Congress, no known restrictions).
