Business to Agents (B2A): How Business Needs to Sell When Software Starts to Buy
When your best customer never blinks
The most valuable customer you will acquire in the next 24 months will not read your homepage, watch your launch video, or attend your webinar. It will not laugh at your brand voice or complain about your pricing page. It will never take a lunch with your sales representative. It will glance at your API spec, test a few endpoints, benchmark latency, check a cryptographic attestation, and decide.
That “customer” is an AI agent.
For years, we’ve explored how models move from passive autocomplete to agentic action, how software shifts from tools to teammates, how conversational interfaces evolve into operating systems for work. I have written about the discipline required to structure those conversations, and the messy, jagged shape of machine capability that demands new workflows and governance. Those threads converge here in the most practical way yet: commerce built for buyers that are software first and human second.
Business to Agents (B2A) is a go-to-market posture, a re-architecture of products and channels, and a discipline for serving autonomous demand. Kantar calls B2A “an imminent distribution channel” and argues that brands that optimize for agents now will “own the invisible shelf.” We must heed this warning label on the next S-curve.
Why now? Because the rails and the retailers are quietly flipping their switches. Visa, Mastercard, and PayPal have all announced agent-ready capabilities that let AI systems discover, authorize, and complete payments on people’s behalf. That makes “software as shopper” real, not theoretical.
And the top of the funnel is changing shape. Amazon continues to expand Rufus, its shopping assistant that routes attention and spend inside the Amazon app, with early signs of ad monetization flowing into the assistant’s answers. When your product card is evaluated by an algorithmic buying agent living inside a marketplace’s AI, your “brand moment” becomes a schema, a latency number, and a policy on refunds expressed in code.
This is a strategy piece. I’m going to name the shift, challenge some defaults, lay out an enterprise grade architecture and playbook, translate it for regulated industries, and then look up. Along the way I will connect back to prior work on conversation design, compound systems, cognitive browsing, and workforce strategy because B2A is where those threads become revenue engine.
1) Trend and context: The buyer is changing species
Agents are becoming channel, not feature. Kantar’s retail analysis places B2A on a 18–24 month clock for share reshuffling. “Build rails for software agents first, humans second” is the core instruction. That means perfect APIs, structured product data, and instant machine-to-machine decisions. It also means the funnel compresses from weeks to milliseconds.
Payments are going agent-native. Visa’s Intelligent Commerce, Mastercard’s Agent Pay, and PayPal’s Agent Toolkit with an MCP server are all designed so agents can shop, authorize, and reconcile with consent and controls. This is much more than “new checkout buttons.” These offerings standardize how agents prove identity, express preference ceilings, and resolve disputes. That is the compliance backbone of B2A.
Platforms are field-testing agentic retail. Amazon’s Rufus now routes product discovery inside its app and has begun testing ad placements in assistant answers. Performance improvements around speculative decoding show the underlying systems are scaling to peak traffic. When the marketplace’s agent ranks SKUs, brands are competing on schemas and service guarantees as much as copy and creative.
Forecasts point to a steep climb, even if the yardsticks disagree. Depending on the analyst, the global AI agents market climbs from ~$5.4B in 2024 to between ~$50B and ~$236B by 2030–2034. In the United States, Grand View Research estimates ~$1.6B in 2024 rising to ~$13.5B by 2030. The spread is large, but the slope is not in question.
Enterprise adoption is tilting from “copilot” to “agent.” Gartner-cited figures circulating in industry analysis suggest that by 2028, a third of enterprise software will include agentic capabilities and agents will make a meaningful share of day-to-day decisions autonomously. If even half of that lands, your systems will be selling to, and being sold by, non-human buyers.
We have been here conceptually. In Conversation Engineering, I argued that every durable agent is a loop of intention, verification, and escalation. In Jagged Intelligence, I warned that uneven capability requires workflow design that routes to the agent’s strengths. In Cognitive Browsing, I outlined how agents learn the open web like analysts, not chatbots. B2A is where all three become operational KPIs.
2) Challenge the defaults: Your brand is a dataset, your funnel is a loop
Leaders often ask me how to “market to agents.” The honest answer is uncomfortable. You do not persuade an agent. You service it. You make it effortless to evaluate you, reason about you, and supervise you at machine speed. That flips several instincts.
Default 1: Story first
Counter: Schema first. Agents care about product graphs, not prose. If your feature facts, inventory positions, regulatory labels, and service guarantees are not structured, they do not exist for the buyer that matters.
Default 2: Linear funnel
Counter: Closed loops. Agents run sense–orient–reason–act cycles continuously. They price-check your SKU hourly and switch if your SLA slips. Your “retention” is the area under that loop where you remain optimal.
Default 3: Loyalty is emotion
Counter: Loyalty is the cheapest total cost to serve for a given outcome. That includes return friction, traceability, and dispute automation. If your policy is friendly to a human but ambiguous to a policy engine, the agent will not prefer you.
Default 4: SEO
Counter: AEO. Think Agent Experience Optimization. It is not meta descriptions. It is endpoint clarity, semantic types, evidence of compliance, and proof of identity. Agents penalize ambiguity because ambiguity wastes tokens and time.
Default 5: Sales plays
Counter: Service contracts in code. Your top-of-funnel narrative becomes a signed JSON bundle that states capabilities, safety constraints, delivery guarantees, and remedies, all wrapped in a verifiable credential the agent can check without calling your lawyer.
This is not a rejection of brand. Humans still care. Humans will still approve baskets and set preference boundaries. But in B2A, brand is primarily expressed as clean data and predictable conduct. That is the currency agents trade.
3) Deep dive: The B2A reference architecture
This section gets specific. If you are a CIO, CDO, CPO, or the VP who will be blamed if this breaks, here is a blueprint you can run.
3.1 Agent-first surface area
Agent Storefront Manifest. Publish a machine readable manifest that declares what you sell, how you price, how you ship, and how you prove it. Use JSON-LD with schema.org types, GS1 attributes, and your sector’s regulatory fields. Include:
Make it fetchable at a well-known path and sign it. Agents should be able to verify provenance without any “call me for a demo” nonsense.
Action Schemas and Playbooks. Define the actions an agent can take and the proofs you will produce. Example action types:
Payment rails. Support the agentic primitives that payments providers are standardizing tokenized identity, consent windows, programmable spending thresholds, dispute hooks. Visa’s Intelligent Commerce and Mastercard’s Agent Pay are good templates; PayPal’s Agent Toolkit and MCP server show what plug and play feels like for developers. Your job is to adopt and instrument.
3.2 Trust, verification, and safety
Verifiable identities. Give agents a way to verify you and require them to verify themselves. Use decentralized identifiers where appropriate and issue attestations for provenance, quality, and compliance. The agentic economy will be gated by trust primitives; several credible analyses argue that verifiable AI and credentials are the only way to scale agent-to-agent trade.
Audit by default. Log every machine decision with enough detail that a human can reconstruct “who decided what, under which policy, using which evidence.” Make logs queryable. Regulators and customers will ask.
Safety envelopes. Constrain autonomous actions to reversible, low-risk domains before granting higher-impact permissions. Treat safety as a set of composable policies you can turn on per product line, region, or buyer class, not a single switch for the whole enterprise.
3.3 Data plane and latency budgets
Agents are ruthless about time. Build for their clocks.
Data needs to be semantic and synchronized. If your warehouse system says “Size: L” and your PDP says, “42 Long” and your manifest says “Large,” the agent will either normalize or walk away. One of those outcomes costs you money. Invest in canonical vocabularies, unit harmonization, and referential integrity.
3.4 Reasoning, routing, and recovery
Agents succeed not by being universally smart but by orchestrating strengths. In Jagged Intelligence, I argued that knowing where models are brilliant and where they stumble is the new craft. In B2A, route tasks to the right specialists: retrieval agents for policy facts, solvers for combinatorial pricing and routing, planners for multi-step negotiations, and critics for safety. The compound pattern wins.
Recovery matters. Define escalation paths from agent to human and back, with a tight handoff. In Conversation Engineering, I laid out a loop that keeps humans in the circuit without becoming bottlenecks. Your agents should degrade gracefully, announce uncertainty, and invite oversight. That keeps trust high and chargebacks low.
3.5 Evaluation and governance
Treat agents like employees with badges.
4) Proof points and where value shows up first
Procurement and supply chain. Autonomous procurement agents parse supplier docs, reconcile invoices, and manage purchase orders. Vendors like Zycus and Ivalua are publishing detailed guides as enterprises automate routine steps end-to-end. Some template agents claim purchase-order cycle time reductions up to 90 percent in constrained domains, where rules and outcomes are crisp.
Financial services and payments. We covered the rails. The meta story is that payments networks, PSPs, and wallets are converging on agent-friendly primitives. Visa and Mastercard hardened consent and traceability. PayPal exposed a toolkit and MCP endpoint and even announced a partnership with Perplexity to embed checkout inside an AI interface. That is all B2A muscle.
Retail and marketplaces. Amazon’s Rufus is a living test bed. It channels discovery inside the walled garden, appears to be moving toward assistant-native ad inventory, and is being optimized under extreme traffic. Reviews are mixed. That is not the point. The point is incentives. If the marketplace’s agent can generate profit by steering to SKUs that satisfy both customer intent and ad yield, your schema and your service claims are how you defend margin.
Enterprise software spend. Analysts at ARK argue that agentic software will reallocate spend from seats to outcomes, with early examples in customer service and code modernization. Microsoft, Salesforce, and UiPath are actively repositioning around agents, not just copilots. The revenue models that failed to materialize for chatbots are likely to land for agents because agents perform work, not just answer questions.
Public sector and regulated industries. My own work in the U.S. federal civil sector shows strong pull for agentic workflows that can be audited, attributed, and reversed. Think eligibility determinations explained with computation graphs, or case triage with explicit guardrails. Your next hire has a GPU, yes, but your next clearance might be for a non-human teammate that needs an identity, a role, and a memory boundary. We need doctrine for that.
5) Business model design: How B2A makes money
Usage-metered service contracts. Agents do not buy seats. They consume capability. Price on units that correlate with value: verified deliveries, settled disputes, optimized routes, recovered churn. Expose those meters in the manifest so agents can simulate TCO before they act.
Commission and referral at machine scale. Affiliate mechanics will migrate into agent ecosystems. Be careful. Humans accept “nudges.” Agents will optimize to those incentives ruthlessly. Make sure you can prove that your referrals meet a standard of fairness and disclose the terms to the supervising human.
Outcome-backed subscriptions. Where outcomes are measurable (defects avoided, minutes saved), bundle a base subscription with performance-based credits. Agents will choose vendors that let them minimize variance and smooth cash outlay for their humans.
Agent marketplaces and exchanges. Marketplaces that let buyers rent agents with narrow superpowers will emerge. The winning platforms will standardize discovery, evaluation, safety, and payments. Several investors and strategists are already arguing that agent market value could rival or exceed SaaS, especially in verticals with clear success criteria. Treat those claims cautiously, but do not ignore the signal.
6) The B2A enterprise playbook
Here is a practical, sequenced plan. Most organizations can complete a first pass in 90 days.
Phase 1: Make yourself legible
Phase 2: Prove value in a narrow lane
Phase 3: Industrialize
7) Sector notes: Where B2A meets real constraints
Federal and public health. Document provenance and explainability are non-negotiable. Require attestations for data sources, restrict cross-boundary memory, and log every policy decision. Build agent actions that map to CFR requirements. This is where Cognitive Browsing meets compliance and where Conversation Engineering meets due process.
Healthcare. Start with provider operations: benefits verification, prior auth triage, scheduling. Keep PHI out of long term memory unless you have encryption and rotation nailed. Demand verifiable credentials for device and clinician identity.
Mobility and field services. Agents coordinate parts, technicians, routes, and SLAs. Outcome metrics are clear, so value shows up fast. Publish manifests for inventory and service windows by depot. Tie compensations to first-time-fix rates.
Defense and critical infrastructure. Airgap where necessary. Use local inference for critical decisions. Build audit logs that survive in court and in Congress. Your next mission partner might be an agent cluster that needs a badge, a billet, and a budget line.
8) Risks and realities
Hype and definitional fog. CIO coverage has noted that “agent” is being used to market everything from scripted chatbots to autonomous planners. Buyers need definitions and benchmarks. Do not buy language. Buy capability, proofs, and outcomes.
Market concentration. Payment rails and app stores are converging. If a handful of companies control the default agent identities and traffic, they control discovery. Counter by publishing open manifests, supporting multiple rails, and keeping your data portable.
Safety and misuse. If your manifest becomes an attack surface, you just published a menu for fraud. Sign everything, rate limit, and require attestations for risky actions. Stage your rollout like you would a security control, not a campaign.
Data readiness. Agentic value crashes if your data is brittle. Leaders underestimate this. Recent industry coverage hammers the point: most firms are not ready. Fix your data plane before you ship clever agents.
Trust. World Economic Forum’s recent analysis frames trust as the currency of the agent economy. Build it with verifiable claims, transparent intents, and reversible actions. Lose it once and agents will not come back.
9) The horizon: B2A today, A2A tomorrow
If B2A is selling to software, A2A is software selling to software. Payments and marketplaces are moving. Kibo and others are publishing credible playbooks for agent-driven commerce. Analysts and strategists are arguing that vertical agents could dwarf SaaS in certain domains. Skepticism is healthy. Still, the vector is obvious.
We will see:
Call this the commercial layer of digital superintelligence. Not the mythic, limitless kind. The practical, verifiable kind that ships goods, closes books, and resolves issues while you sleep. The kind that acts, explains, and accepts consequences. We have work to do before we trust it everywhere. But the scaffolding is going up.
10) Frequently asked questions I get
How do we “do SEO for agents”? Stop thinking about ranking hacks. Think like a platform. Publish schemas, proofs, and policies. Improve the read and write paths. Prove identity. Make ambiguity expensive for the agent, and you will lose every contested basket.
Will brand die? No. Humans still decide budgets and boundaries. Brand becomes the tiebreaker and the forgiveness budget. It also becomes the promise your manifests must keep.
What should we measure? Agent-qualified traffic. Agent satisfaction and abandonment. Time to decision. Exceptions per thousand actions. Autonomous to supervised ratio. Dispute rates. Cost to serve by agent class.
Will agents kill sales and marketing? They will force new work. You will sell to algorithms with data and to humans with meaning. Both matter. The team that can talk to both will win.
Where do we start if we sell to government? Publish your manifest with attestations. Map actions to policy. Require agent identity. Keep humans in the loop for eligibility or adjudication. Document everything. Ship small, audited loops first.
11) A short field guide (print this)
The Quiet Handover
The internet turned storefronts into APIs. B2A turns persuasion into proofs. If your product can be evaluated by a machine in under a second, it will be. If your promises cannot, they will be ignored.
This is not a small optimization to your funnel. It is a handover of the buyer’s first look, first judgment, and first purchase to a class of systems that reward clarity, integrity, and speed. The leaders who internalize that will compound advantage fast. The rest will still be campaigning for the human’s attention while agents close out the cart.
I will leave you with a question to debate in the comments:
If an agent can prove it bought the right thing, at the right price, with recourse baked in, will your board still insist that a human must click “Buy”? If yes, for how long?
References
Excellent post Bassel. I found the examples provided relative to payment rails to be particularly thought provoking (Drew Leety Jason Rich Delie Minaie) 💡