Your Next Big Customer Is a Spreadsheet With a Budget
Forrester says procurement agents will negotiate with hundreds of suppliers at once in 2026. How to build an ICP for a buyer that is an algorithm, and earn its selection.
On this page
Your next seven-figure account will never see your homepage.
It will not watch the demo. It will not warm to your founder story. It will pull your price list, your SLA, your return policy, and your delivery windows into a comparison matrix at 3 a.m., score you against 214 competitors, and either open a negotiation or delete your row.
Forrester predicts that in 2026, one in five B2B sellers will be compelled to respond to AI-powered buyer agents with dynamically delivered counteroffers, as procurement agents negotiate prices, terms, and service levels across hundreds of suppliers at once. Not sequentially, the way a human category manager works a shortlist. Simultaneously.
Which means the most important ICP you write this year describes a buyer that is literally an algorithm. You are now selling to a spreadsheet. The spreadsheet has a budget, an objective function, and zero patience for adjectives.
The spreadsheet is already shopping
This is not a 2030 thought experiment. Gartner projects that machine customers will directly influence or participate in $30 trillion worth of purchases by 2030, and its CEO research finds executives expecting 15–20% of revenue to flow through them by the same year.
The consumer side moved first. Adobe measured traffic from AI sources to U.S. retail sites growing 393% year over year in Q1 2026, and by March 2026 that AI-referred traffic converted 42% better than traffic from non-AI channels. A year earlier it converted worse. The flip happened in twelve months.
B2B is where the asymmetry gets dangerous. A Deloitte Digital survey covered by Digital Commerce 360 found 38% of buyers already using agentic AI in purchasing, against just 24% of suppliers using agents in the sales process. The same suppliers estimate they have already lost 13% of sales bids to poor buyer experiences. The buyers automated. Most sellers did not. That gap is where market share changes hands.
What the spreadsheet actually weights
The honest way to model a machine customer is as two layers: hard filters in front, weighted scoring behind.
Hard filters come first, and they are brutal because they are silent. Can the agent parse your catalog? Can it verify stock in real time? Can it transact through a protocol it supports? Can it confirm your compliance certifications without emailing anyone? Fail one of these and you are removed before scoring begins. You will never know you lost. There is no "we came in second on price" in this world. There is only a row that got deleted.
Survive the filters and you reach the weighted columns:
| Signal | What the agent checks | Where it looks |
|---|---|---|
| Total cost | Unit price, volume tiers, shipping, payment terms | Structured price feeds, quote APIs |
| Availability | Live inventory, lead times, fulfillment reliability | Inventory APIs, historical order data |
| Terms and policy | Returns, warranties, SLAs, compliance certs | Machine-readable policy pages and docs |
| Data completeness | Whether every attribute can be parsed and compared | Schema markup, product feeds, llms.txt |
| Response latency | How fast your side answers a quote or counteroffer | Quote automation, seller-side agents |
| Track record | On-time delivery, dispute rate, review corpus | Marketplaces, third-party data, past transactions |
Notice what is missing. Awards. Design. Tone of voice. The spreadsheet has no column for charisma.
Why your adjectives are invisible
"Premium craftsmanship" does not parse. "Ships in two business days, 98.6% on-time over the trailing twelve months" parses beautifully. Machine customers compare what they can extract, and extraction is where most brands are quietly failing.
Adobe's readability analysis found the average U.S. retail product page scores just 66 out of 100 for machine readability, with homepages averaging 75%. Translate that: roughly a third of the content on the average product page, the page doing the actual selling, is invisible to the buyer that Forrester says will negotiate with hundreds of your competitors before lunch. You spent six figures on copy the customer cannot read.
There is a second-order effect worth sitting with. When an LLM-mediated agent does the comparing, your brand does not disappear; it gets converted into a prior. Models carry beliefs about which suppliers are reliable, formed from training data, citations, reviews, and everything ever written about you. Reputation becomes a probability distribution. That is a very different asset from a feeling in a human chest, and it is built in public, over years, in text.
Writing the ICP when the buyer is an algorithm
A traditional ICP asks about industry, headcount, and pain points. An ICP for a machine customer asks five different questions:
- Platform. Which agent stack will evaluate you? A consumer shopping agent riding the Agentic Commerce Protocol that OpenAI launched with Stripe behaves differently from an enterprise procurement suite bolted onto SAP. Each has its own retrieval habits, supported protocols, and trust signals.
- Objective function. Is the agent minimizing landed cost, maximizing reliability under a budget cap, or optimizing replenishment cadence? You cannot position against a goal you have not identified.
- Constraints. Compliance regimes, geographic requirements, payment rails, approved-vendor lists. These are the hard filters. Map them per segment the way you would map firmographics.
- Data access. What can the agent actually read about you today? Run your own catalog through an AI visibility checker before assuming the answer is "everything."
- Escalation triggers. When does the agent call its human? That handoff moment is the one place in the loop where persuasion still works, so know exactly what causes it.
Then segment. Agents cluster by platform, objective, and constraint profile just as humans cluster by intent and value, and the modeling discipline transfers directly from predictive audience work in ecommerce. The difference is that agent segments are cleaner. Algorithms do not have moods.
Earning selection when no human reads the page
Four moves matter more than everything else combined.
Make every claim a field. Prices, tiers, lead times, warranty terms, certifications: all structured, all current, all parseable. Start with proper markup via a schema generator and publish an llms.txt file so agents know where to look. This is unglamorous work. It is also the new homepage.
In September 2026 we audited the homepages of 54 well-known B2B SaaS vendors with our AI Visibility Checker. None blocked an AI answer crawler and 59% already published an llms.txt file, but only half carried two or more high-value schema types, 15% had no structured data at all, and just 24% exposed machine-readable publish or modified dates. Those are the homepages of well-resourced vendors, so they show what the leaders do; they do not show whether llms.txt affects citations, and they say nothing about smaller sites.
Answer at machine speed. Forrester's prediction cuts both ways: buyer agents negotiate, which means seller agents must counter. A bot that answers questions is table stakes; the real distinction is between answering and doing, and quote negotiation is firmly in the doing category. If your counteroffer takes a business day, the negotiation ended yesterday.
Feed the prior. Reviews, third-party benchmarks, published reliability data, consistent entity information across the open web. This is the discipline behind AI search visibility, and it compounds slowly, which is exactly why late movers cannot buy their way in.
Enrich what you know about the buyer's owner. Every agent has a principal: a human or a company whose preferences got encoded. The same data enrichment discipline that sharpens human segmentation tells you which constraints and objectives are likely sitting inside the agents that evaluate you.
When does building for the buying agent hurt the human sale?
The trade-off nobody mentions is that a parseable price list is also a visible floor. Publish structured tiers and the agent anchors on them; in practice the deals still closed by phone compress toward the published number. Keep list price and tiers structured and route anything below list through the quote API, where a counteroffer can be conditional.
A seller-side agent that counters at machine speed backfires without floors. Told to win the negotiation, it races to the bottom of its authority on every request, and the first sign is won deals at margins no human would accept. Set floors and escalation triggers before autonomy, as you would for any agent on the roster.
The mistake a competent team makes first is turning every claim into a field before every field has a live source. A lead time published as structured data and then missed is scored as unreliability, which agents weight more harshly than a blank cell; worse, the penalty outlives the fix in the track-record column. Publish only what a system of record keeps current.
It depends on the category. Unless a meaningful share of your buyers already run procurement agents, protocol integrations cost more than the deals they reach; spec-driven categories see agent traffic first. That said, readability work pays off with human buyers and AI search either way, so we would not recommend deferring it, only the negotiation agent.
The humans are still upstream
Here is the strategic comfort, and it is real. Somebody configures the spreadsheet. A procurement lead decides which suppliers get whitelisted. A consumer tells their shopping agent "stick to brands I already trust." Forrester's own 2026 buying research shows procurement gaining influence and buying groups growing, which means the humans setting the constraints matter more than ever, even as they touch individual transactions less.
So brand did not die. It moved. It now lives in two places: the whitelist decision made by a human before the agent ever runs, and the prior baked into the model that scores you. Both are won years before any single deal and neither can be won during one.
The spreadsheet does not hate you. It cannot see you. Every blank cell in your row is a deal that went to whoever filled theirs in.
Fill in the row.
Sources
- Forrester, 2026 B2B Marketing, Sales, and Product Predictions (linked above)
- Gartner, Machine Customers Will Decide Who Gets Their Trillion-Dollar Business (linked above)
- Adobe, AI Traffic Grows but Retail Sites Lag in Machine Readability (linked above)
- Digital Commerce 360, Deloitte Digital survey on B2B supplier maturity (linked above)