The great AI vocabulary shuffle.
Prompt engineering, agentic workflows, context engineering. Every year brings a dearer noun while the slide on financial returns stays the same.
By Amit Srivastava · 6 October 2026
Somewhere in a head office this morning, an executive slide deck is being updated.
Prompt engineering has been crossed out. Context engineering sits in its place, linked by a thicker arrow to a box marked agentic workflows.
Only one slide in the pack has not changed since 2023: the one showing net financial returns.
This has been the quiet comedy of workplace AI over the past three years. Each time promised gains stall on their way to the accounts, the industry supplies a grander, more expensive noun. Every term describes a real technical step forward. The trouble starts when a new technical label is sold as the cure for an unchanged management habit.
Three years of moving goalposts
Recall how fast the certainty moved.
In January 2023, Andrej Karpathy remarked that the hottest new programming language was English. Within weeks, job boards carried adverts for Prompt Engineers paying up to $335,000 a year. Companies bought prompt handbooks and taught senior executives four-part formulas: assign a role, state the task, add constraints, specify the format. Clear instructions help, of course, just as they help when briefing a colleague or a builder. Yet better phrasing was quietly allowed to stand in for better judgement. When a project disappointed, the user was told they had asked the machine badly.
By January 2025, the browser tab was given an employee badge. Industry leaders predicted AI agents would join the workforce that year, calling tools and completing multi-step jobs on their own. Organisations that still had no reliable way to verify a single AI-drafted board memo were suddenly invited to deploy fleets of autonomous software workers.
Six months later, the goalpost shifted again. In June 2025, Shopify chief executive Tobi Lütke argued that context engineering beat prompt engineering, a point Karpathy amplified days later. Production systems need the right documents, policies, memory and permissions inside the context window, not merely a clever sentence. By late July 2025, Gartner’s verdict was brisk: context engineering was in, and prompt engineering was out.
Then the invoices arrived.
Three years of dearer nouns. One unmoved leadership question.
Track 01 · What the slides called the answer
- Prompt engineering“English is code” · ads up to $335k
- Agentic workflowsAutonomous bots issued credentials
- Context engineeringGartner: “Context in, prompt out”
- 40%+ cancelledGartner forecast · “agent washing”
Track 02 · What reached the accounts
56%
of CEOs reported neither higher revenue nor lower cost from AI over the prior 12 months (PwC, Jan 2026).
“Analyse this proposal and explain why it is the right choice.”
Add clean context and three agents: the output gets faster and more persuasive, while the conclusion hidden inside the request escapes inspection.
In June 2025, Gartner predicted that more than 40 per cent of agentic AI projects would be cancelled by the end of 2027 due to rising costs, unclear business value or weak risk controls, estimating that only around 130 of the thousands of vendors marketing agentic tools possessed genuine capabilities, and calling the rest agent washing. Six months after that, PwC’s January 2026 survey of 4,454 chief executives found that 56 per cent had seen neither higher revenue nor lower costs from AI over the preceding year.
A perfectly written bad request
Why does each new layer of technical sophistication leave the same gap behind? Look at a request that passes every executive prompting course with flying colours:
Analyse this proposal and explain why it is the right choice.
Give that instruction a crisp role, ten megabytes of clean company context and three cooperating agents. You will receive a fluent, well-sourced, board-ready defence of a decision that was smuggled inside the question before work began.
Nothing failed at the interface. Nothing failed in the context window. The software obeyed.
The failure sits in the human system around the tool, precisely where Assumptions, Biases and Complexity live:
- Assumptions: Nobody separated tested customer facts from convenient defaults before framing the request.
- Biases: The question steered the machine to confirm the sponsor’s preferred answer instead of trying to break it.
- Complexity: Nobody traced what happens across downstream teams, incentives and customers once the polished deck is approved: the Transition link where saved minutes either reach the P&L or vanish.
Teaching executives to write tidier prompts without examining how decisions are challenged is like teaching a driver to speak more clearly into a satnav pointed at the wrong city.
Put the discipline where the consequence sits
Prompt craft, context architecture and agentic tools all belong in a modern enterprise. Teach them to the specialists and operators building the workflows.
Leadership teams need a different habit: one for the five minutes before a request is framed and the ten minutes after a fluent answer lands on the table. What did this answer assume? How did our own preference tilt the evidence? Who absorbs the downstream work, and who owns the result in the accounts?
Next year will almost certainly bring another phrase to replace context engineering on the conference stage. You do not need to wait for it.
Ask one plain question at your next review: besides the software and the nouns on the slide, what have we changed about how a weak decision gets stopped?
Sources: Andrej Karpathy posts on X (24 January 2023; 25 June 2025). TIME, “The AI Job That Pays Up to $335K” (14 April 2023). Sam Altman, “Reflections” (6 January 2025). Tobi Lütke post on X (18 June 2025). Gartner press release on agentic AI cancellations and “agent washing” (25 June 2025) and research note on context engineering (28 July 2025). PwC, 29th Global CEO Survey (19 January 2026; 4,454 CEOs).
