Upgrading an old London townhouse to full‑fiber internet is painful: drilling through century‑old walls, snaking cable behind ornate mouldings, negotiating wayleaves. It’s slow and expensive—and when you’re done, it’s still an old house with a modern system bolted on. In the UK, street works can account for roughly 70% of the cost of fiber deployment. That’s the retrofit tax.
Most enterprises are those townhouses. They’re pulling “agentic AI” through the legacy foundations of fragile APIs, manual approvals, siloed data, compliance workarounds etc.
They’ll get something working, but the real advantage shifts to the new builds: companies architected, from day one, around agents that can plan, act, and safely close loops.
We’ve seen this movie before. Clayton Christensen ’s disruption pattern explains why incumbents optimized for yesterday’s model don’t spot a new one coming. Netflix was “barely a blip” when it offered to sell to Blockbuster; a few years later, the market flipped. The difference now is the speed. With agentic AI, the cycle compresses from decades to years.
Agentic AI isn’t a feature bolt‑on. It’s an operating model. Agentic systems don’t just answer; they pursue goals, use tools, coordinate with other agents and people, and take reversible actions with audit trails. When you design the business around that reality - data that is action‑ready, policies expressed as code, and processes that default to straight‑through execution - you create a company that learns and compounds faster than rivals who are still threading cable through plaster.
And the first movers and disruptors will be so far ahead that catching up to them will be a Herculean effort (talk about a moat!)
Why retrofits struggle
• Legacy wayleaves: every approval chain, re‑keyed form, or brittle integration is a street you have to dig up. The work is real and recurring.
• Fragmented identity and data: agents need consistent identities and dependable data contracts; most estates have years of sprawl.
• Cultural drag:__ teams built for handoffs rarely start with goals → tools → checks → evaluations as the basic design loop for autonomous actors.
• Pilot theatre: co‑pilots impress in demos but stall at production because nothing downstream is ready for agents that act.
What agentic‑native looks like
• Agent‑first process design: straight‑through when safe; human‑in‑the‑loop by exception.
• Actionable data fabric: governed data products with timestamps and contracts; retrieval‑augmented actions, not just retrieval‑augmented answers.
• Tooling and policy‑as‑code: explicit permissions, rate limits, reversible transactions, tamper‑evident logs.
• Observability and evaluation:__ runbooks, sandboxes, red‑team tests, and task‑specific evals for each agent.
• Risk plumbing from day one:__ identity for agents, allow‑lists for high‑risk actions, full audit trails, and clear lines of liability.
Early signals to watch
• Klarna’s AI assistant now handles a large majority of support chats in production across many markets, reducing repeat contacts and shrinking resolution times from minutes to under two. Those gains came from redesigning a core workflow, not from adding a chatbot to the side.
• In healthcare, post‑discharge voice agents are moving from pilots to production in select hospitals and are tackling the costly gap between a patient leaving and a patient recovering, with clear escalation paths when a human is needed.
• On the platform side, mainstream enterprise stacks are rolling out agent primitives with action frameworks, policies, and observability so that teams can move from “assist” to “act” with proper guardrails.
• Open frameworks such as multi‑agent orchestration libraries make it easier to coordinate specialized agents, keep long‑running state, and enforce safety checks. The tooling is maturing quickly.
A founder’s playbook
1) Choose one workflow you can own end‑to‑end (claims resolution, onboarding, collections, revenue ops). Make it your wedge.
2) Design the agent fabric first (goal → tools → constraints → evals), then wrap UX and operations around it.
3) Ship in weeks. Template each win as a reusable skill‑pack with tests.
4)Instrument ruthlessly : straight‑through percentage, human‑touch ratio, time‑to‑resolution, unit cost per resolved case.
5) Sell the outcome, not the feature : faster cycle time, fewer defects, lower cost per case.
An incumbent’s survival kit
• Stand up a greenfield, agentic line of business with its own data plane and P&L; judge it on throughput and outcomes, not headcount.
• Carve agent‑ready corridors in the core: identities for agents, a minimal set of dependable APIs, and action approvals where money or safety is at risk.
• Use barbell governance: tight guardrails for critical actions; generous sandboxes for exploration.
• Acquire to accelerate—but build an integration factory so the capability survives contact with the mothership.
• Re‑skill for task design and risk: less generic “prompting,” more policy‑as‑code, runbooks, and evaluation design.
A 90‑day plan
Weeks 1–2:
* Pick a high‑volume process with clear guardrails.
* Write the allowed tools and policies as code.
Weeks 3–6:
* Build the MVP agent. Log every action.
* Run in a sandbox with pass/fail evals.
Weeks 7–10:
* Move to a limited production cohort with human‑in‑the‑loop.
* Track throughput, defects, escalations.
Weeks 11–13:
* Publish a one‑page scorecard.
* Either expand the cohort or stop and learn, then try again.
Bottom line
You can pull fiber through old walls, but the new build will always be faster, cheaper, and easier to extend. The first agentic‑native player in your market will look narrow... until one morning it’s the default. If a startup launched tomorrow and resolved your customers’ top three problems without human handoffs, what would you need to change in the next 90 days to keep it from becoming the new standard?
References (for readers who want to dig deeper)
• Building broadband and mobile infrastructure (UK Parliament Library briefing): https://researchbriefings.files.parliament.uk/documents/CBP-9156/CBP-9156.pdf
• Klarna AI Assistant performance highlights: https://openai.com/index/klarna/ and https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/
• Universal Health Services x Hippocratic AI post‑discharge agents (deployment notes): https://uhs.com/news/universal-health-services-launches-hippocratic-ais-generative-ai-healthcare-agents-to-assist-with-post-discharge-patient-engagement/
• Salesforce Agentforce announcements: https://www.salesforce.com/news/
• HubSpot AI agents update (Investor relations newsroom): https://ir.hubspot.com/news
• Multi‑agent orchestration patterns (Microsoft AutoGen): https://www.microsoft.com/en-us/research/publication/autogen-enabling-next-gen-llm-applications-via-multi-agent-conversation-framework/
• LangGraph project: https://www.langchain.com/langgraph and https://github.com/langchain-ai/langgraph
• Clayton Christensen, The Innovator’s Dilemma (overview): https://hbr.org/2015/12/what-is-disruptive-innovation