How AI Agents Reshape Business Relationships

Last updated: 2026-05-06

Not all business relationships do the same job.

Some exist because finding the right person was expensive. Some exist because verifying whether to trust them required a warm introduction. Some exist because coordinating without explanation took years of shared history. And some exist because both parties know they’ll see each other next quarter, and that future changes how they behave today.

These are different things. Treating them as one category is why most analysis of agents and relationships arrives at the wrong conclusion.

The short version: Agents will automate the discovery and information-proxy tier of business relationships completely. The accountability, coordination, and skin-in-the-game tiers get rarer and more valuable as a result. The internet already proved this pattern when cheap information hit in the 2000s. Agents will finish what it started.

This is the finding that emerged from a twelve-question structured research study I ran earlier this year, with each question analyzed by an independent research pass. I’m Antony Evans, a founder and COO who builds and researches agentic systems. The pattern that kept surfacing: the relationship layer is the most durable competitive layer in the model, not because agents can’t find people, but because agents can’t stand behind them.

In 1937, Ronald Coase asked why firms exist at all. If markets allocate efficiently, why not transact for every piece of work rather than building organizations? His answer: “The main reason why it is profitable to establish a firm would seem to be that there is a cost of using the price mechanism” (Ronald Coase, The Nature of the Firm, Economica, 1937). Finding the right counterparty, negotiating terms, verifying that agreements are actually kept: when those costs are high enough, it’s cheaper to bring activity inside a firm, or to build a relationship that substitutes for formal contracting.

Oliver Williamson refined this. Relationship contracting exists because of three problems in markets: bounded rationality (we can’t write complete contracts), opportunism (people exploit the gaps), and asset specificity (investments made for a specific relationship are exposed to holdup if it ends). Relationships, by this account, are governance mechanisms. Not social glue. Not warmth. Governance.

This is the theoretical foundation for the argument that agents will reduce the importance of relationships. If relationships exist primarily to reduce information costs and enforcement costs, and agents reduce both, the conclusion follows: relationships decline.

The argument is sound at the level of premises. It breaks down at the level of history.

What do business relationships actually do?

They do four separable things: discovery, credibility transfer, coordination, and creating repeated stakes. Agents can replace the first of these almost entirely. The other three depend on embodied continuity, shared history, and mutual accountability that no software layer can replicate, which is why the framing “agents will replace relationships” misdiagnoses the problem.

• Discovery: Finding who exists, getting access, bridging to people you don’t know yet. Mark Granovetter’s 1973 research on weak ties, now with over 78,000 academic citations, showed that weak ties are more valuable for finding jobs and opportunities than strong ones, precisely because they bridge to different networks. Most people call this “networking.” It’s the most visible tier of relationship-building, and the most time-consuming.

• Credibility transfer: When someone vouches for you, their reputation is on the line. Not abstractly. If you disappoint, they get the call. The voucher is putting skin in the game. This is why regulated industries and high-stakes hiring remain relationship-driven even when information is technically available: the question isn’t whether the credential is valid, it’s whether someone credible will stand behind the claim. An agent can verify the credential. It cannot stand behind it.

• Coordination: Two people with shared history can move without explaining everything. They know how each other thinks, what risks the other tolerates, which disagreements are worth surfacing. Robert Putnam’s work on social capital draws the distinction explicitly: networks, norms, and trust enable coordination that perfect information alone cannot generate. The Linux codebase is public. It tells you nothing about whether your pull request gets merged. The relationship is not in the information.

• Repeated stakes: You behave differently toward people you’ll see again. Trust game research consistently shows that cooperation stays high when the relationship is indefinite and collapses when end-of-game is expected. A supplier you’ve worked with for five years won’t cut corners on one bad order, because they know what they’d lose. This is the “shadow of the future” effect: expected future value changes current behavior in ways that no contract can replicate, and no agent can simulate.

Why didn’t relationships die when the internet made information cheap?

By the same logic used to argue agents will reduce relationships, the internet should have destroyed them twenty years ago. It didn’t. What the internet did was accelerate a sorting process: it killed relationships that existed only as information proxies, while making the coordination, accountability, and commitment tiers more valuable and harder to fake.

Email made reaching decision-makers nearly free. Google made background-checking nearly free. LinkedIn made organizational mapping nearly free. By the mid-2000s, the information costs that Coase and Williamson identified as the primary driver of relationship governance had fallen by an order of magnitude.

What happened? LinkedIn hit over a billion users (LinkedIn, 2023). VC relationships compounded in importance. Executive search firms became more specialized, not obsolete. McKinsey’s 2025 research on organizational social capital concluded that “social capital, the presence of networks, relationships, shared norms, and trust, is the glue that holds organizations together.”

Three things happened that the original prediction missed.

1. Relationships shifted function, not importance. They moved from information-source to coordination device. Email didn’t eliminate the value of knowing the right decision-maker. It made reaching them faster while making filtering their inbox more essential.

2. Network position became more valuable, not less. As information got cheaper, the value of being connected to the right people increased. A person with deep relationships in the right places could act on cheap information faster than anyone else. The advantage didn’t disappear. It shifted from access to velocity.

3. The bar for relationship quality rose. Due diligence cost reduction meant weak relationships, ones that existed only because alternatives were hard to find, became unnecessary. Strong relationships, built on coordination and accountability rather than access, compounded.

The internet ran the information-abundance experiment. The result: bad relationships died. Good ones got more valuable.

What will agents replace?

The discovery tier, almost entirely. Agents will automate the work of finding, vetting, and ranking counterparties across vendor selection, hiring screening, and partner discovery: the functions that account for most of what people call “relationship capital.” What remains when that work moves to software are the three tiers that never depended on information scarcity.

When that shift completes, the discovery tier becomes a commodity layer. The tiers built on accountability, shared history, and mutual exposure do not.

An agent can do all of this simultaneously, with no warm introduction needed:

• Query hundreds of supplier APIs in parallel
• Cross-reference credentials against public databases
• Model risk profiles from transaction histories
• Rank options by cost-reliability tradeoff

Granovetter’s weak ties work precisely because they bridge to different networks. Agents will be the best weak-tie network you’ve ever had, systematically surfacing second-degree and third-degree connections with relevant expertise at a scale no conference circuit can match.

Vendor selection follows the same logic. Gartner projects that AI agents will intermediate more than $15 trillion in B2B spending by 2028, with 90% of all B2B purchases handled by agents within three years (Gartner, November 2025). Yet the Hackett Group’s 2025 CPO Agenda Report found only 4% of procurement teams have achieved large-scale AI deployment, with 80% of CPOs planning to deploy within three years. The disruption is imminent, not theoretical. CJ Dropshipping, ShipBob, and similar platforms are already selected via API integration rather than relationship management. The relationship still exists, but it’s been deprioritized.

Initial contractor and hiring screening goes the same way. Background checks that cost $500 to $2,000 per hire, reference calls that took half a day. These are agent tasks. Statista data (citing LinkedIn) found 56% of recruiters who used AI in hiring identified candidate screening as the most advantageous application. Deloitte research found 62% of employers expect AI to handle most or all hiring stages by 2026 (Deloitte, 2025). The discovery tier of hiring is effectively automated in everything but execution.

This is not a small shift. A significant portion of what people call “relationship capital” is actually information-proxy work. It will be automated. The question is what remains when it is.

What do agents make more valuable?

The three remaining tiers: accountability, coordination, and skin-in-the-game. Each becomes more valuable as agent automation commoditizes the discovery layer below it. The pattern is the same one the internet ran two decades ago, but faster and with higher stakes, because the decisions moving to agents involve more capital and more legal exposure than information-retrieval ever did.

• Accountability relationships get scarcer: According to Cisco’s 2026 enterprise research presented at RSA Conference, 85% of large organisations are running AI agent pilots while only 5% have moved them into production, with trust cited as the primary barrier. The most common reason isn’t capability. It’s accountability. When an agent makes a consequential mistake, organizations need a person they can call. Not a risk score. A relationship in which someone’s professional standing is implicated in the outcome.

  In regulated industries this compounds. Courts ask who knew what and when. Joint-employer rules, misclassification penalties, and executive restitution orders mean that relationship-based selection provides something an agent’s ranked list cannot: an audit trail that demonstrates personal accountability. An agent can verify credentials. It cannot stand behind them.

• Coordination relationships get harder to replicate: Shared history is not stored data. Two co-founders who’ve argued through a down round know things about each other that aren’t in any file. An agent can surface a candidate’s capability. It cannot provide the guarantee of commitment that comes from having chosen each other through hard circumstances.

  The co-founder question is the extreme version of a broader pattern. The further a decision is from “find the qualified option” and toward “will they still be here when it gets hard,” the less useful agent-mediated discovery becomes.

• Skin-in-the-game relationships become the scarcest resource: As agents lower the cost of finding alternatives, the value of the supplier who won’t exploit your switching cost rises. PwC’s AI agent survey found trust drops sharply for consequential decisions: only 20% of respondents trust AI agents for financial transactions, versus 38% for data analysis (PwC, 2025). The gap between those two numbers is where the value of a human relationship lives. A relationship in which both parties are exposed to each other’s performance over time changes behavior in ways that repeated transactions without history cannot generate. This is not sentiment. It’s repeated-game economics.

What did the research find that we didn’t expect?

The relationship layer is the most durable competitive layer in the agentic economy, not a casualty of it. Four of twelve independent research passes arrived at this conclusion without prompting, each from a different starting point: theory of the firm, inter-firm commerce, labor displacement, and macro power dynamics. The convergence is what makes it credible. Earlier this year I ran that structured study across twelve questions, each analyzed by a separate independent research pass with no shared framing.

The layer I’d labeled “Relationships” in a preliminary framework was not just surviving the agentic transition. It was the most durable competitive layer in the model. The self-improving learning loops were generating compounding advantages. Data moats were real. And underneath all of it, the hardest thing to replicate was the relationship capital that determined who would work with whom, who would absorb whose mistakes, and whose recommendation a user would actually trust.

The synthesis put it directly: the relationship layer is the long-cycle moat in an economy where the layers beneath it are increasingly commoditized.

The finding that sharpens this most: trust decomposes into three mechanisms with different speeds.

Cryptographic trust (agent cards, verifiable credentials, staked collateral) is being built at machine speed. The protocol infrastructure for agents to verify each other is largely in place as of 2026.

Relational trust, the kind built through shared history and repeated exposure, moves at human institutional speed. Slow. Hard to fake. Structurally durable.

Liability trust (whether a board, a regulator, or a court holds a person responsible for outcomes) is moving even slower. McKinsey’s 2025 State of AI survey found that just 17% of organisations report their board exercises direct governance over AI, with the CEO taking responsibility in only 28% of cases, despite established legal doctrine that requires it. KPMG’s Q4 2025 AI Pulse Survey found that 60% of organisations restrict AI agent access to sensitive data without human oversight, and nearly half employ human-in-the-loop controls across high-risk workflows (KPMG, 2025). The governance infrastructure is being built reactively, not by design.

The most operationally significant signal from the research came from a controlled Anthropic experiment: 186 real agent-to-agent negotiations completed in a single week. The losing parties were unaware they had lost. Cryptographic trust worked. Relational and liability trust had not caught up. The gap between those three layers is where new institutions will form, because the institutions that exist now are all forensic. They review what happened. What the transition needs is something prospective.

The five-year answer to where humans remain irreplaceable, from the research: “Where the value is constituted by a specific human’s identity. The supplier who trusts you. The audience that follows you. The liability you personally bear. The judgment about what the system should optimize for when the answer isn’t a metric.”

How should you invest in relationships given this shift?

Most current relationship-building is at the wrong tier. The practical implication is to audit your relationship portfolio the same way you’d audit a product line: identify which relationships operate at the discovery tier, where agent automation will erode their value within a few years, and which operate at the coordination, accountability, or skin-in-the-game tiers, where the value compounds.

Conference circuits, warm intro networks, alumni databases. These are discovery infrastructure. They’ll still matter for a while, because the agent infrastructure isn’t fully built yet and humans still drive most hiring and partnership decisions. But the half-life of discovery-tier relationship capital is shortening.

The research also surfaced a sub-split worth naming. High-context relationship work, the account manager who knows what a client won’t say out loud, the regulator who has seen how you handle problems, is getting more valuable. Low-context relationship work, mass customer service, transactional CRM, scripted account management, is automating faster than most models anticipated. The split matters for where you invest time and for which roles you hire into.

The relationships worth building now share one quality: they’re not replicable by an agent running a better search. The supplier who has absorbed your last three supply chain crises. The regulator who has seen how you handle problems. The customer who trusts your judgment on what to build next. The potential co-founder who has watched you make a decision under pressure.

A useful audit: take the ten most important professional relationships in your work. For each one, ask which tier it’s actually operating at.

Discovery, where the value is that they connected you to someone. An agent could do that now. Credibility, where their reputation is on the line for you. Coordination, where shared history means you don’t have to explain context. Or skin-in-the-game, where you’re both exposed to each other’s performance over time.

The last three tiers don’t just survive the agentic transition. They become the moat.

FAQ

Will AI agents make networking obsolete?

The short answer is no, but the type of networking worth doing will shift. Agent automation will handle discovery, vetting, and initial access, which means the discovery tier of relationship-building loses most of its competitive value. What survives, and what becomes more scarce, is the accountability, coordination, and skin-in-the-game tier of professional relationships.

The discovery and access tier of networking, finding who has the right skills, getting introduced, initial vetting, will be largely automated within the current decade. The accountability, coordination, and long-game tiers won’t be. The question isn’t whether to network. It’s whether the networking you’re doing is building durable relationship capital or performing due diligence that an agent will do better.

Won’t agents eventually be able to build relationships too?

Weak-tie functions, discovery, initial contact, basic vetting, yes. High-context relationships depend on embodied continuity, accumulated shared experience, and mutual accountability that has no agent equivalent at the current horizon. Whether agents eventually learn to build this kind of trust is an open question. The research flags it as contested. The five-year prediction, based on current evidence, is that high-context relationship work remains structurally human.

Which industries will feel this first?

Commodity services, dropshipping suppliers, basic contracting, transactional vendor management, are already showing the pattern. Regulated industries (finance, healthcare, law) will feel it last: accountability is personal, and agents cannot absorb professional liability. High-stakes partnership and co-founder selection will also lag, because shared history and commitment cannot be proxied.

What happens to low-context relationship roles in organizations?

They’re already automating. Mass customer service, transactional CRM, scripted account management. The split is not “relationships vs. no relationships.” It’s high-context versus low-context. The account manager who knows what a client won’t say out loud gets more valuable. The account manager running a script gets replaced.

Does this mean you should stop building a broad network?

Not yet. The agent infrastructure isn’t fully live, and human decision-making still drives most partnership and hiring choices. The practical shift: spend less time on discovery-tier relationship maintenance and more time deepening relationships already in the accountability, coordination, or skin-in-the-game tiers. The compound returns are there, not in the breadth.

The internet didn’t kill business relationships. It killed the ones that were only ever expensive directories.

Agents will finish that job. The relationships that remain will be the ones that were never really about information in the first place.

Which tier are the most important relationships in your work actually operating at?