23 August 2026

From chatbots to agentic workflows: the evolution every leader needs to understand

Anthropic's telemetry shows 77% of enterprise AI usage is now full task delegation. Gartner expects 40% of agentic projects to be cancelled by 2027. Both are true, and the difference is governance.

Three years ago the enterprise AI question was "what can it answer?" It is now "what can it be trusted to finish?" That shift has a measurable start, a measurable middle, and a well-documented failure mode. Leaders who understand all three make very different decisions from leaders who have only read the headlines.

The shift is real, and it is measurable

Anthropic's Economic Index report (15 September 2025) is the cleanest evidence. Analysing usage patterns, it found 77% of enterprise API traffic showed automation patterns — predominantly full task delegation — against roughly 50% on the consumer product. Businesses had already stopped using AI as a conversation partner while the public discourse was still about chatbots.

OpenAI's From assistance to execution (12 August 2026) confirms it from another vantage point: its agentic coding surface accounted for 64% of combined enterprise output tokens by June 2026, with the fastest adoption outside engineering entirely — legal use grew 108x in four months.

The middle is messier than the marketing

McKinsey's State of AI (November 2025, 1,993 respondents across 105 countries) is the sobering counterweight: 62% of organisations are at least experimenting with agents, but only 23% are scaling an agentic system anywhere — and no more than about 10% have scaled agents in any single business function. 88% use AI somewhere; roughly 6% report meaningful EBIT impact.

Gartner's numbers echo it. Their 2026 CIO survey, cited by Forrester, finds 17% of enterprises have actually deployed AI agents while over 60% expect to within two years. And the widely quoted prediction: over 40% of agentic AI projects will be cancelled by the end of 2027, driven by cost, unclear value and inadequate risk controls (Gartner, 25 June 2025).

"Agent washing" is the leadership trap

In the same release, Gartner estimated that of the thousands of vendors claiming agentic capability, only around 130 are genuinely agentic. The rest are rebranded chatbots, RPA scripts and assistants. For a buyer, the practical test is unglamorous but decisive: can it take an action in my system of record, under a permission scope, and log what it did? If not, it is a chatbot with new packaging.

What the field evidence shows

Commonwealth Bank of Australia. Its agentic contact-centre platform resolved roughly 84.6% of self-service messaging interactions end to end in May 2026, across more than two million conversations a month.

Lloyds Banking Group. An agentic framework spanning 21 million accounts delivered around £50M of value in 2025, with a stated target above £100M for 2026.

Klarna — the instructive one. Its 2024 assistant handled 2.3 million chats in month one, work equivalent to about 700 full-time agents, cutting average resolution from 11 minutes to 2. By 2025–26 the company had partially reversed course and rehired human staff. The lesson is not that agents failed; it is that scope failed. Deflection was optimised; the exception path was not.

What leaders should actually take from this

  • The direction is settled. Delegated execution is already the majority of enterprise AI usage. Debating whether it is coming is a year out of date.
  • The cancellations are a scoping failure, not a technology failure. Projects die on unclear value and weak controls — both decided before the build starts.
  • Test for agency, not adjectives. Action, permission scope, audit trail. Anything else is agent washing.
  • Design the exception path first. Klarna's arc is what happens when the human fallback is an afterthought.
  • Scale one function properly before spanning ten. Almost nobody has cleared 10% penetration in a single function — depth is still an available advantage.

The evolution from chatbot to agentic workflow is not a product upgrade. It is a change in what you are delegating and therefore what you must be able to prove afterwards.

Sources