Agentic AI

This narrative documents the second stage of the AI editorial progression developed at Sage — the agentic vision, making the case for where generative AI was heading before it fully arrived. It follows Generative AI and precedes Practical AI in the sequence: Generative AI → Agentic AI → Practical AI → Trusted AI.

The problem

Each previous stage of the AI narrative had a relatively clear editorial task.

Generative AI needed to be demystified and given practical shape. Practical AI needed implementation guidance grounded in domain-specific evidence. Trusted AI needed a governance argument that moved beyond capability claims to demonstrated accountability.

Agentic AI is harder to frame. The capability is real — AI systems that operate continuously, monitor conditions, make decisions and complete sequences of tasks without being prompted at each step. But the editorial challenge is not explaining the technology. It is explaining what it demands of the organisations that want to use it.

The market question is: can AI complete meaningful work on our behalf?

The answer is increasingly yes. But the more important question is: what does that work depend on?

The narrative thesis

Agentic AI shifts the failure mode.

In earlier phases, poor AI outputs were visible. A human reviewed the draft, caught the hallucination, corrected the summary. The error was local and recoverable. In agentic workflows, that review step is removed by design. The agent acts. The output enters a downstream process. By the time an error is visible, it may have propagated across multiple steps.

That changes the quality requirement for the knowledge underneath — the context, frameworks, operating rules and narrative logic the agent works from.

The core argument has three parts:

Autonomous execution requires structured knowledge, not just capability. An agent without a clear knowledge base improvises. It draws on general training data, applies generic logic, and produces outputs that are fluent but not aligned. The organisation’s point of view, the specific market argument, the evidence base — none of that is available to the agent unless it has been explicitly built.

Agentic AI does not replace editorial judgement. It raises the cost of its absence. When humans review every output, editorial judgement is applied at the point of review. When agents act autonomously, editorial judgement must be designed in — through frameworks, through defined decision rules, through structured evidence that the agent can draw on. The absence of editorial investment is not a saving. It is a compounding liability.

The shift is from hindsight to foresight. Earlier AI tools helped explain what happened — faster reporting, better summarisation of existing data. Agentic AI operates in the present and the near future: monitoring conditions, detecting anomalies, initiating responses before a human has noticed the problem. That shift changes what organisations need to provide: not records to be analysed, but operating logic for a system that is making decisions continuously.

The content that came out of it

The agentic AI editorial phase required a different kind of argument — less adoption guidance, more structural thinking about what agentic systems need to work well.

“Agentic AI in Practice: How Sage is Reshaping Accounting with MTD” made the concrete case. Making Tax Digital had already moved accounting from annual to continuous. Agentic AI was the logical extension: systems that monitor compliance continuously, detect anomalies before they become problems, and route decisions to the right person at the right time. The MTD context gave the agentic argument practical grounding rather than speculative framing.

The three-pillar architecture that emerged from this work — continuous accounting, continuous assurance, continuous insight — was not a product description. It was an editorial framework for how autonomous AI systems change the structure of finance work. Instead of period-end reviews, continuous monitoring. Instead of static reports, contextual recommendations. Instead of waiting for insight, insight arriving without being requested.

The Authentic Intelligence concept, developed through the trusted AI phase, carried forward and deepened: AI built for a specific domain, with transparent training, human accountability designed in, and professional standards embedded. In the agentic phase, Authentic Intelligence became a more demanding standard — not just AI that could be trusted, but AI that could be trusted to act.

The proof points accumulated:

• AI agents scanning 3.2 billion transactions flagged and corrected 190 million errors
• Implementations removing up to 90% of manual data entry
• Financial period close moving from weeks to two or three days
• 25,000 UK accountants and bookkeepers already operating with Copilot as a daily tool

Aaron Harris, Sage CTO: “The future of AI in finance won’t be about faster answers. It’ll be about encouraging smarter, more autonomous action.”

Signals from this phase

The knowledge gap is the new implementation gap In the practical AI phase, the implementation challenge was embedding AI into workflows. In the agentic phase, the challenge is providing AI with the structured knowledge it needs to act well. Organisations that invested in clear narratives, documented frameworks and structured evidence are discovering that their AI outputs are more consistent. Those that did not are discovering that autonomous execution amplifies generic thinking at scale.

Authentic Intelligence is a differentiating position As agentic AI becomes more widely available, the differentiator is not which vendor’s agents an organisation uses. It is the quality of the knowledge those agents operate from. Domain-specific training, professional standards, transparent governance — these are not compliance features. They are the basis of competitive advantage in an environment where execution is no longer the bottleneck.

The editorial operating system is the infrastructure question Every organisation building on agentic AI is, whether it recognises it or not, making decisions about its editorial operating system: what context agents will have, what frameworks will guide their decisions, what evidence they can draw on, what they are and are not permitted to do. Organisations that design this deliberately will outperform those that discover it through error.

MTD as a preview of what agentic AI requires The shift from annual to continuous compliance that MTD forced was a preview of the operational model that agentic AI assumes. Continuous data, continuous monitoring, continuous decision-making. Accounting practices that adapted to MTD are better placed for agentic AI not because they have better tools, but because they have already restructured their operating model around continuous rather than episodic work.

The editorial intelligence connection

The agentic AI narrative is where Editorial Intelligence becomes most directly relevant.

The knowledge, judgement and narrative logic that Editorial Intelligence builds — frameworks, evidence libraries, defined positions, structured signals — is exactly what agentic AI agents need in order to produce outputs that reflect the organisation’s perspective rather than generic AI capability.

The Editorial Intelligence and Agentic AI framework develops this argument in detail: what the editorial operating system is, why its absence compounds over time, and what the practice looks like in the accounting and finance context.

The consistent thread across all four stages of this narrative — generative, agentic, practical, trusted — is that the technology does not reduce the need for structured organisational knowledge. It increases it. Each stage has raised the stakes.

Open questions

• How do practitioners establish editorial governance for agentic workflows — not policy-level guardrails, but the specific operating rules that guide agent decision-making in practice?
• What does the capacity gap look like when applied to agentic execution rather than just AI-assisted tasks? Where does autonomous work plateau without structured knowledge to draw on?
• How do organisations measure editorial quality in agent outputs — when there is no human reviewer in the loop to catch what does not align?

Each answer becomes a new signal. Each signal strengthens or refines the narrative.

See Practical AI for how the narrative evolved from here — from vision to implementation.

Why this narrative matters

The agentic AI narrative matters because it makes the stakes of autonomous AI specific rather than abstract.

The question is not whether agentic AI systems will become more capable — they will. The question is what organisations need to have in place for those systems to produce outputs that are worth having. The answer is structured knowledge: clear narratives, documented frameworks, connected evidence, defined positions that the agent can draw on rather than improvise around.

That is a different kind of organisational investment from anything AI previously required. Generative AI required prompting skills and workflow integration. Agentic AI requires something deeper: an editorial operating system — the accumulated intelligence that makes autonomous execution coherent rather than generic.

The organisations that understand this earliest will have a compounding advantage: not because their agents are more capable, but because their agents are working from something. The argument this narrative makes — that the knowledge underneath AI matters more than the AI itself — will become more obviously true as autonomous execution becomes more widespread. Getting ahead of that argument, with evidence rather than speculation, is what this phase of editorial work was for.

Articles from this phase