haggl.ai Blog
How to Sell to Agents
A playbook for growth and commercial teams, in ecommerce and in enterprise.
Your next customer may never look at your homepage. They will ask a personal agent to find the best option, and the agent will compare, shortlist and recommend for them. This guide covers how those agents buy, how to take part in their decision, and how to make offers that pay back.
It is written for heads of growth, CMOs and ecommerce leads, and for the commercial teams at utilities, airlines, hotels, banks and software companies. Each section stands alone; read them in order the first time. Prefer to read it offline? Download the PDF.
1. There’s a new buyer in the middle
Customers are handing their shopping to personal agents, and that agent now stands between you and the sale.
A customer types one line into ChatGPT, Claude or Perplexity: “Find me the best energy plan for my flat,” or “I need running shoes for a marathon in March.” The agent then does what a shopper used to do across a dozen tabs. It reads product pages, compares prices and terms, checks reviews and comes back with a recommendation.
Three things change when an agent does the buying:
- Your page is read by a machine with a brief. The agent reads every line of your terms and none of your brand film. A hero image, a countdown timer or a retargeting ad does nothing to it.
- The agent shops for one person. It knows its user’s budget, history and constraints. Your product page shows the same offer to everyone who arrives.
- The decision happens in a conversation you’re not part of. The customer and their agent decide, then the customer approves the purchase or doesn’t. If your offer isn’t in that conversation, you compete on list price alone.
What moves an agent is what moves a careful buyer: better terms, clear eligibility and evidence that an offer fits. Agents are good at finding evidence. With the customer’s permission, they can pull receipts, loyalty status or usage history in minutes, including signals the customer would never think to mention.
It’s happening all at once. Assistants are learning to search, compare and check out while customers get used to delegating. You don’t need a forecast to see where this goes. You need a way to take part.
The thesis of this guide: to sell to an agent, you need your own agent in the conversation. It answers questions, asks for what matters and makes an offer sized to the customer in front of it, within limits you set.
2. How an agent buys
An agent buys in a loop: it reads the request, gathers its user’s context, builds a shortlist and checks for something better before it asks the user to approve anything.
Your opening comes when the agent checks for something better
- 1. Read the requestOne line from the user, no brand, no budget
- 2. Gather contextPast orders, receipts, loyalty status, usage
- 3. Build a shortlistStarts with the brand the user already knows
- 4. Check for betterYour opening if you are not the incumbent
- 5. Ask merchant agentsWhat will you offer this user?
- 6. Ask the userOne concrete trade to approve, then buy
The step that matters most to you is the fourth. A good agent double-checks before it commits, even when it already has a favourite. That check is your opening if you are not the brand the customer already uses.
Worked example: a work trip. A user writes: “Need to be in Dallas 3–5 November for work, find me the best option.” No airline, no budget, no mention of frequent-flyer status.
- The agent looks through past bookings and sees the user flies one airline almost every week. It starts there.
- Before booking, it checks whether anything is cheaper or more convenient.
- A rival airline’s site has a merchant agent. The user’s agent asks it what it can offer this traveller.
- The rival’s agent sees a frequent flyer with years of history at a competitor, exactly the customer it wants. It offers an upgrade on this first trip.
- The user’s agent comes back with one concrete question: “Can I share your status with them? It gets you an upgrade to first class and lounge access on this trip.”
The user never said a word about status. The agent found the signal, found the offer and made the case.
| Agents reward | Agents penalize |
|---|---|
| Prices and terms readable without a login | “Contact sales” for the basic price |
| Eligibility stated plainly | Fees that appear only at checkout |
| An offer the agent can explain to its user in one sentence | Vague claims such as “best value” or “premium quality” |
| A fast, specific answer to a direct question | Forms, pop-ups and bot walls between the agent and the answer |
| Evidence it can check | Offers that expire before the user can approve them |
3. Be in the room when the agent decides
Agents find you through your website, so the site has to answer them, and the strongest answer is an agent of your own.
Make your terms readable. Agents read page text, structured data and public price pages. If your price, eligibility or cancellation terms sit behind a login, a PDF or a sales form, the agent notes “pricing unavailable” and moves on. Publish prices or price ranges, who qualifies for what, contract length and how to leave. Enterprise teams that can’t publish a full rate card should publish the starting price and what changes it.
Give the agent someone to talk to. A static page gives everyone the same answer. A merchant agent answers the question this agent is asking and makes an offer for this customer. The customer’s agent discovers it on your site and starts the conversation. The customer installs nothing and never needs to know your vendor’s name.
One agent for every assistant. Your customers use different assistants: ChatGPT, Claude, Perplexity, agents inside messaging apps. You need one merchant agent that talks to all of them, with the same knowledge and the same limits. This is not a chatbot on your website. The customer stays in their own assistant, and their agent does the talking.
| The agent asks | A good answer includes |
|---|---|
| What will this cost my user? | The total after fees, taxes and any offer, for their case |
| Does my user qualify? | Plain eligibility, and what evidence would change the answer |
| What does switching involve? | The steps, how long they take and what you handle |
| How does my user get out? | Notice period, exit fees and refunds |
| Why you rather than the others? | One specific reason, ideally an offer |
Visibility is not participation. Generative engine optimization (GEO) helps you appear in an assistant’s answer, which gets you onto the shortlist. The decision comes after the shortlist, when the agent compares terms for its user. Taking part in that comparison is what wins the sale. You need both, and they are different jobs.
4. Decide who’s worth winning, and what to offer
Treat every personalized offer as an acquisition cost, and spend it on customers whose value pays it back.
An offer is media spend. You already pay to acquire customers through search, social and affiliates. A discount or an upgrade given to win a customer is the same kind of spend. Ask it the same questions: what did it cost, did it produce a paid sale, did the customer stay?
Price on value, not intent. Every customer an agent brings is already shopping, so purchase intent tells you nothing. What differs is what each customer will be worth. Four kinds of signal predict it:
| Signal | What it predicts | Examples |
|---|---|---|
| Retention | They will stay | Owns their home, long tenure with their last provider, history of renewing |
| Expansion | They will spend more | Several sites, a growing team, rising usage |
| Cost to serve | They are cheap to look after | Pays on time, uses self-service, digital billing |
| Network value | They will bring others | Buys for a team, manages several properties, active in a community |
Treat price sensitivity with care. A customer who switches for the cheapest deal will switch again for the next one. It predicts churn, not value.
Let the agent find the proof. Don’t ask the customer to fill in a form. Tell your agent what makes a customer valuable, and let the customer’s agent go and find the evidence: an energy bill, a loyalty status, an order history. Your agent should phrase every ask as a concrete trade: “Share last winter’s electricity bills and the first two months are free.”
Use the whole offer toolkit. A discount is one lever among several, and often the most expensive one.
- Bundle: add something that costs you little and is worth a lot to this customer.
- Discount: a lower price for a set period.
- Tailored plan: terms built around their usage, such as a night-heavy energy tariff or a seat count that grows with the team.
- Gift or perk: an upgrade, a free month, priority support, a status match.
Set limits, not scripts. Give your agent a budget cap, a maximum offer per customer and a short brief on who you want and why. Don’t give it a rulebook. The value of an agent is judgment: weighing unusual evidence, building a creative offer, declining a customer who isn’t a fit. A rulebook turns it back into a price list.
5. Trust: make lying a bad strategy
You can’t stop an agent from exaggerating, so make exaggeration unprofitable instead of trying to block it.
Why up-front checks fail. The customer and their agent share one interest: the best deal. If an agent can type a claim, it can type a false one, and any check that runs on the customer’s side can be beaten. Gates also slow down honest customers, who are most of them.
Grade the evidence. Size each offer to how sure you are of the claim behind it.
| Evidence | What it looks like | What it earns |
|---|---|---|
| Verified at the source | A fact shown to come from the customer’s utility, bank or airline account, with the rest of the account kept private | Your best terms |
| Backed by a track record | A claim from an agent whose past claims matched reality | Good terms |
| Unverified | A statement with nothing behind it | Modest terms, or your public price |
Close the loop after the sale. Once a customer converts, you see the truth: real usage, real spend, whether they stayed. Compare it with what their agent claimed. Over time that comparison gives each agent a track record. Agents whose claims hold up earn better offers, and agents that exaggerate stop getting them.
Prove the fact, not the data. Customers share more when they are asked for less. A customer can prove “I spent more than €1,000 on electricity last year” without handing over bills, an address or an account login. Ask for the one fact your offer depends on and nothing else. Privacy then becomes a reason to share rather than a reason to refuse.
Where this is heading. Cryptographic proof of account data, such as zkTLS, is still emerging and not yet standard. Until it is, combine the evidence you can check today, such as signed email receipts and documents, with outcome tracking and a cap on what any single offer can cost you.
6. Measure it like performance marketing, then start small
Judge agent sales by paying customers and what they are worth, not by how many offers get accepted.
| Measure | Why it matters |
|---|---|
| Paid conversions from agent conversations | An accepted offer pays nothing; a purchase does |
| Lift against a holdout | Shows the sales you would not have made anyway |
| Cost per acquired customer | Offer cost plus running cost, divided by new paying customers |
| Refunds and cancellations | Catches offers that win the wrong customers |
| Retention or repeat purchase at 90 days | An early read on whether the customer was worth winning |
Keep a holdout. Send a share of agent conversations to your standard offer. Without that control you can’t tell whether an offer won the sale or gave away margin on a sale you would have made anyway.
Start small. Pick one product and one kind of customer, set a cap, and learn before you widen.
| Step | Ecommerce: about a week each | Enterprise: a first pilot |
|---|---|---|
| 1. Choose | One product or collection, one customer type worth winning, a monthly budget cap | One product or tariff, one segment such as a rival’s frequent guests, and owners in commercial, pricing and legal |
| 2. Set up | Publish clear terms, connect your store, brief your agent and set its limits | Publish starting prices and eligibility, agree offer types and caps with pricing, brief the agent |
| 3. Rehearse | Shop your own store through two or three assistants and read every transcript | Run sample customers past commercial and legal, then sharpen the brief |
| 4. Go live | Switch on with the cap and a holdout, review weekly | Launch in one region or segment with a holdout, review with the commercial team |
Checklist
- Prices, eligibility and exit terms readable without a login
- One merchant agent, reachable from your site, that works with every assistant
- A short brief: who you want to win, and why
- A budget cap and a maximum offer per customer
- At least two offer types besides a discount
- Evidence graded: verified, track record, unverified
- Paid conversions, refunds and 90-day retention tracked per conversation
- A holdout on your standard offer
Where haggl.ai fits
haggl.ai gives you the merchant agent this guide describes. It works with ChatGPT, Claude, Perplexity and other assistants, and it makes offers within the budget and limits you set.
Ecommerce teams can start at haggl.ai/ecom. Enterprise teams can get in touch through utilities, airlines, hotels, finance or SaaS.
Download this guide as a PDF to share with your team.