What is AI used for in the CPG industry?

The uses vendors are selling into production inside consumer packaged goods companies today are largely back office rather than marketing. The first is accounting and cash: collecting retailer files from many different portals, applying cash to invoices, and managing deductions and disputes. The second is sales planning, meaning sales forecasting and the trade promotion spending a brand commits with its retailers. The third is operational planning: how much to produce and how much raw material that requires. Confido, a Y Combinator company that announced a $55 million Series B led by Insight Partners in September 2026, sells agents across all three and says more than 250 brands run on its platform, including divisions of Unilever, Mars, and Nestle. The consumer insight and personalization use cases that dominate search results are real, but the contracts vendors in this space describe publicly are concentrated in the back office.

Why do CPG companies buy AI agents when they rarely buy software?

Because an agent is not sold to them as software. Consumer brands are not known as big software buyers but are enormous buyers of services: they hire business process outsourcers, retain consulting firms, and pay third-party brokers a share of revenue, around 5 percent by Confido's founders' account, to do commercial work for them. A vendor offering to do the work rather than supply a tool is therefore a familiar purchase, and the founders say buyers never frame it as software, they ask whether it replaces or augments the labor they would otherwise need. That is also why the category is attractive to founders generally. Work that is already outsourced needs no behavior change from the buyer and no new budget line, because you are displacing a vendor and showing up where the money already goes.

Can AI agents replace an ERP for a consumer brand?

Not really, and the better framing is that they do different jobs. An ERP is a system of record: it holds the general ledger, supports the monthly close, and produces the statutory and audit reporting. Agents of the kind Confido sells sit on top of whatever ledger a brand already has and perform the commercial labor around it, pulling retailer data, applying cash, chasing deductions, and building forecasts. Many mid-sized consumer brands were never on a heavy ERP at all, running instead on spreadsheets and point solutions, which is the gap the agent layer fills. Rebuilding the ledger itself around a language model is a separate product category, an AI ERP, and a brand can end up buying both.

What controls should a consumer brand set before AI agents touch its financial data?

Six, and all of them belong in place before the first deployment rather than after an incident. Give each agent its own scoped, revocable identity instead of a shared service account or a borrowed human login, since these agents hold retailer portal credentials. Keep a write boundary so an agent drafts entries and a person posts anything above a defined threshold, which preserves separation of duties. Write that automation threshold down as policy: reconciling a ten dollar item automatically is sensible, settling a six figure dispute is not. Require an audit trail that ties every agent-produced entry to its source document and the rule behind it. Ask in the contract how your point-of-sale data is isolated from competitors on the same shelf. Finally, keep a human reading forecast variance, because a model that sets purchase orders can be wrong in a way that shows up as excess inventory rather than an error message.

AI in CPG: What Consumer Brands Actually Buy

Search for AI in CPG and page one hands you shopper personalization, concept testing, and shelf imagery. That is the demo reel. The work one vendor is doing for divisions of Mars and Unilever is far less photogenic: cash application, deductions, trade promotion, and planning how much raw material to buy. Confido, a Y Combinator company in New York, built an agent workforce for exactly that back office and announced a $55 million Series B led by Insight Partners in September 2026, with divisions of Mars and Unilever among its customers. Its founders also describe an underserved market, one that answered their cold emails within a couple of days and then talked to them for hours. Here is what AI in CPG looks like inside a real consumer brand, why an industry famous for not buying software buys agents anyway, and what the pattern tells any founder picking a vertical.

What AI in CPG actually looks like today

Ask the people selling it and the answer is narrow and operational, not visionary. Confido describes itself as "building the AI agent workforce for consumer brands," and the test for whether a brand is in scope is deliberately domestic: "think of any brand that you might see in your refrigerator or your medicine cabinet." The scope inside each brand is wide. The founders say they "help with everything from accounting automation to sales planning to operational planning in a single system."

What ties those together is data the brand already owns and mostly cannot use. As one founder puts it, "underlying all of this is the unique retail data that these brands have and we use it to power key workflows across the business that automate labor." Retailers hand brands a large amount of detail about products, promotions, contracts, and shipments, and most of it arrives as files in portals that a person has to go collect.

In practice the work falls into three families:

Accounting and cash. Their first workflow attacked "a very manual process of logging into many different portals and pulling down information," then automated cash application and the accounting behind it. Deductions and disputes live here too, and they are where brands quietly lose margin.
Sales planning. Sales forecasting and trade promotion management, the commercial planning that decides what a brand expects to sell and what it spends with retailers to sell it.
Operational planning. How much to produce and how many raw materials that requires, which is the difference between a stockout and a warehouse full of inventory.

The sequence matters more than the list. They did not design a platform and sell it. They solved one portal problem, and then, as they put it, "from there our customers asked us to keep doing more things for them."

Confido's Series B announcement says more than 250 brands now run on the platform and that it carries more than $30 billion in retail sales planning. Its named customers include divisions of Unilever, Mars, and Nestle alongside faster growing brands like Olipop, Simple Mills, and Daisy. In the interview the founders also name Dude Wipes. That is the honest answer to what AI in CPG means today: not a smarter ad, a smaller back office.

Why a famously non-tech industry buys AI agents

The obvious objection is that consumer packaged goods is not a software market. Brands are not known for buying tools, running pilots, or staffing an innovation team. The founders agree with the premise and then turn it over: CPG brands are enormous buyers of services. They hire business process outsourcers, they retain consulting firms, and they pay third-party brokers a share of revenue, around 5 percent by the founders' account, to do commercial work on their behalf.

So "we'll do the work for you" is not a new kind of purchase for them. It is the only kind they are used to. The founders are precise about how buyers frame it: "the way they talk about it is never as software," it is "is this replacing or augmenting the labor we might otherwise require."

That is not a quirk of one category. Y Combinator has been making the same argument as a general thesis, pointing founders at "AI native companies that don't sell software, they sell the service," because services spending is "many times larger than the spend on software and a lot of these services are already outsourced, which makes it much easier to replace with an AI native product." Its longer playbook for the category lists already-outsourced work as the first of four traits of the best markets for AI services: "the work is already outsourced and the customer cares about the final product, not how they got there." The founder advantage that follows is the part worth copying. "You're displacing a vendor, not asking the customer to do something fundamentally different," which means "you're showing up where the budget already lives and doing the work."

Read that against a normal enterprise software sale. You are usually asking a buyer to find new budget, change how their team works, and wait for value. Selling into an already-outsourced function asks for none of those things. The budget exists, the behavior does not change, and the comparison is a vendor invoice rather than a feature grid. This is the core of a service-as-software business, and it is why unglamorous industries keep turning out to be the better opening.

It also sets your pricing floor and ceiling somewhere unfamiliar. As the same playbook puts it, "you're not competing with other software providers, you're competing directly with the cost of labor, internal or outsourced." That is a far more generous anchor than a per seat comparison, and it is why pricing an AI product against the labor it absorbs beats pricing it against the last SaaS tool the customer bought.

The ERP-shaped hole in consumer brands

The natural question is why an ERP did not already do this. Many of the brands Confido started with were not running NetSuite. By the founders' account they were on spreadsheets or a scatter of point solutions, including systems that amount to a thin wrapper over a database: you enter your information and it reads it back to you. What was missing was "a sort of cohesive platform" where the data arrives on its own and, in their words, "the insights are driven agentically."

That gap is the whole opening. The interviewer's framing is that Confido reached NetSuite-level pricing within about two years of starting on the idea, which the founders do not dispute. The pitch is not better bookkeeping. It is reach across the income statement: the founders claim they "can touch almost every line item on the P&L," which is something "an ERP would just simply never be able to do" while it functions as a basic database.

Worth being careful here. This is a vendor describing its own advantage, and a system of record has real jobs that an agent layer does not replace: the audit trail, the close, the statutory reporting. The useful read is that the two are different products. An AI ERP rebuilds the ledger itself around a model. What Confido sells sits on top of whatever ledger exists and does the commercial labor around it. Both are versions of the same bet that vertical software wins where domain depth and permitted data access matter more than model quality.

Where AI in CPG actually makes money

The most useful thing in the interview is a distinction most AI pitches blur. There are two different sources of value, and only one of them is easy to sell.

The first is doing existing work for less. Automate the manual parts of accounting and planning, and the customer saves on headcount. Real, measurable, and a negotiation, because the buyer knows exactly what that labor costs and will price your product against it.

The second is doing work that was never economically possible. This is the better business, and the example is specific. "An agent can reconcile down to a $10 transaction. A human never could." The founders describe a public company that had just signed and that "didn't review transactions less than $10,000 in this area because it was just so manual." Every dollar recovered below that line was money the company was not even looking for. The founders' phrase for it is blunt: "so it's free money."

Planning has the same shape. According to the founders, "billion-dollar businesses are basically forecasting off of historical averages." Nobody chose that. It is what a small planning team can actually maintain in a spreadsheet across a large catalogue and many retail accounts. The ceiling was never ambition, it was labor.

For founders the lesson generalizes past CPG. When you map a prospect's workflow, mark the places where the work stops rather than the places where it is slow. A threshold of that kind is a confession that the economics ruled out the work, and it is the strongest kind of latent demand you can find: no incumbent to displace, no behavior to change, and a number the buyer can verify themselves. The same logic drives the opening in AI for supply chain, where the questions a brand never asked are worth more than the reports it already runs.

Before an agent touches your P&L

An agent workforce in the back office means software with portal credentials, write access to financial records, and influence over purchase orders. That is a serious trust boundary, and the controls are not exotic. They are the ordinary ones, applied before the deployment rather than after an incident.

Give every agent its own identity. Agents logging into retailer portals should not share a human's credentials or one service account. Scope each to the systems it needs and make the access revocable, which is the whole argument for treating agent identity as its own problem.
Keep a write boundary. An agent that drafts an entry and a human who posts it above a defined threshold preserves separation of duties. The automation is in the preparation, not the authorization.
Set thresholds as policy, not as vibes. Reconciling a $10 item automatically is sensible. Settling a six figure dispute is not. Write the number down, review it, and log every exception.
Demand an audit trail per decision. Every entry an agent produces should trace to the source document and the rule that produced it. Without that, your close gets slower, not faster, because nobody can explain a number.
Isolate the retail data. Point-of-sale detail is competitively sensitive. Ask any vendor how your data is separated from the brand you compete with on the same shelf, and get the answer in the contract.
Review the forecast, not just the ledger. A model that quietly sets purchase orders can be wrong in a way that shows up as a warehouse, not an error message. Keep a human reading the variance. This is the practical half of AI compliance: the control that catches a confident mistake.

The pivot that found the idea

Confido is also a useful case study in not dying first. The founders entered Y Combinator's Summer 2021 batch with a different company, called Outfit, building what one describes as a WeWork for fitness instructors at a moment when fitness was moving outdoors. It did not survive contact with the batch. They "had some very sage group partners who told us it may not be the best idea as soon as we joined YC," and on checking, "we looked at the numbers and they were not lying." By the interviewer's account they did not present at Demo Day, and they went through what they call a "long long pivot hell" for roughly a year.

What they landed on was, in their words, "an idea that was hiding in plain sight." By his own account one of them had worked for consumer companies in college, a football helmet manufacturer and Anheuser-Busch, and had seen these exact problems from the inside. The same founder's mother managed a retail store and father sold software to consumer companies. The problem had been in the room the whole time and still took a year to see, which is the normal experience and worth saying out loud.

The validation step is the part to copy. They cold emailed some of the largest CPG companies in the world and got replies within a couple of days. People at Coca-Cola and Dr Pepper were "just willing to talk to us for hours." One exchange became the tell: a Coca-Cola employee asked "do you have 2 hours?" and the founder initially assumed it meant call me back later. It did not. As one founder tells it, "we actually closed our first customer with a Figma," which is to say "prior to writing a single line of code," and read that correctly as evidence of "a real visceral pain and strong need."

That is a cleaner signal than almost anything a founder can manufacture. A warm intro tells you about your network. Strangers at a Fortune 500 giving two unknown founders two unbroken hours, and a buyer signing off a mockup, tells you the pain is real. If you are still refining who exactly you serve, that is the experiment to run, and the broader lesson about surviving a startup pivot is that the year in the desert is not failure, it is the search.

How the team is built

One detail is worth noting because it tracks what this kind of product requires. Asked about it, the founders confirm the company runs without product managers. Engineers own problems end to end instead, and the founders describe "a group of amazing engineers who are very curious and very technical and able to drive to clarity and a lot of ambiguity." Their longest tenured engineer leads a new supply planning module and, by their account, does the road mapping, talks to outside experts, onboards customers, handles enablement, and sits in the sales process.

That is not a cost-cutting choice. When the product is doing a customer's actual work, the distance between the person building and the person living with the output has to be close to zero. It is the same reasoning behind budgeting for a forward deployed engineer before you promise an enterprise logo. The deeper pattern across all of this, choosing an unglamorous job, selling the outcome, and keeping the builders next to the customer, is the thread running through AI for startups more generally.

What to do this week

Pick one already-outsourced function in your market. Write down who does the work today, what the customer pays for it, and whether the buyer cares how it gets done. If they do not care, that is your beachhead.
Hunt for a threshold, not a complaint. Ask three prospects what work they skip because it is not worth the labor. A stated cutoff, the dollar figure below which nobody looks, is worth more than a list of frustrations.
Price against labor, not software. Build the comparison the buyer will actually make: the vendor invoice or the headcount, not the per seat cost of the tool you are replacing.
Send ten cold emails to the biggest companies in your category. Reply rate and call length are your market signal. Two hours from a stranger beats ten warm coffees.
Sell the mockup before you build it. If a prototype in Figma cannot get a verbal commitment, more code will not fix that.
Write the control list before the first deployment. Agent identities, the write boundary, the automation threshold, the audit trail, and who reads the variance. Agree it with your first customer rather than retrofitting it after a bad month.

If you want the full operating system for running a company this way, that is what I teach in AI Operating System for Startups.

Sources

Automating Billion Dollar Brands on YC Root Access, the interview with Confido's founders that this article distills.
Confido's Series B announcement, for the named customer list and the platform scale figures, and the BevNET coverage of the $55 million round led by Insight Partners.
Confido's Y Combinator company page for the Summer 2021 batch and the founders. Profiles: Justin Hunter, co-founder and CEO, and Kara Holinski, co-founder and CTO.
AI-Native Service Companies from Y Combinator, for the services-spend thesis and why outsourced work is easiest to replace.
How to Build an AI-Native Services Company from Y Combinator, for the low-trust market traits, the displacing-a-vendor framing, and pricing against labor.